Future Ventures: Scaling with Clarity

Stuart Frost argues causal AI and rules engines beat LLMs for enterprise value, citing Geminos' railway success.

1h 0m Listen to the original
Causal AIEnterprise SoftwareStartup EconomicsLLM LimitationsCorporate Venture CapitalGeminosGreat Western RailwaysAI Governance

Summary

Stuart Frost argues that AI's true value lies in causal models and rules engines rather than probabilistic LLMs, which he claims are unreliable for high-stakes decisions. He details Geminos' architecture, which combines neural networks with symbolic logic to ensure traceability and safety in enterprise settings. Frost highlights a case study with Great Western Railways where causal AI reduced delay attribution from days to minutes, generating significant value. He criticizes the industry's focus on AGI and fully autonomous agents as a valuation-driven distraction that ignores practical business needs. Additionally, he warns against corporate venture capital conflicts and emphasizes the importance of trust and rapid proof-of-value in enterprise sales cycles.

Episode sections

  1. 0:04 – 2:39

    AI's Impact on Startup Economics

    Stuart Frost explains that AI is drastically reducing the costs and time required for software development and sales cycles, which are historically the primary cost centers that kill startups.

  2. 2:39 – 7:26

    AI in Sales and Product Development

    Frost discusses using AI to analyze sales transcripts for customer insight and training, noting that product development has become hundreds of times faster, allowing for rapid iteration.

  3. 7:26 – 9:35

    Customer-Centricity Over Technology

    Frost argues that companies must understand the customer's job-to-be-done before selecting a technology stack, rather than focusing on technology first.

  4. 9:35 – 13:14

    Rules Engines vs. AI Guardrails

    Frost distinguishes between AI guardrails and traditional rules engines, stating that rules engines provide hard, reliable boundaries for critical decisions, whereas agents often fail to reliably follow soft guardrails.

  5. 13:14 – 15:25

    Inverting the Agent Model with Causal AI

    Frost proposes using causal models for core decision-making and agents only for data gathering, citing the superior mathematical rigor of causal models for business outcomes.

  6. 15:25 – 19:10

    Data-Driven Reality in Enterprises

    Frost claims most large companies are not truly data-driven, relying on outdated spreadsheets, and that causal AI offers a way to implement genuine, testable, data-driven decision-making.

  7. 19:10 – 23:02

    Overcoming Market Noise and Proving Value

    Frost identifies market noise and hype as hurdles for enterprise adoption, noting that Gartner reports only 13% of large enterprises see measurable AI value, and his company uses rapid demonstrations to convince leaders.

  8. 23:02 – 29:46

    Scaling AI Adoption Through Federated Models

    Frost discusses the challenge of scaling AI from central teams to federated models where line managers can test hypotheses, noting they are exploring this but not yet fully offering a self-service platform.

  9. 29:46 – 31:11

    Enterprise Friction and Trust

    Speaker_2 argues that for large enterprises, trust is key, with a focus on protecting downside risk. Speaker_1 notes that organizational friction often stems from a lack of trust.

  10. 31:11 – 33:09

    Sales Cycle Dynamics and Qualification

    Speaker_1 explains that startups must qualify out of deals with high friction, citing a recent win in the food industry after a competitor failed due to excessive barriers.

  11. 33:09 – 34:22

    Sources of Organizational Resistance

    Speaker_1 identifies union fears, internal IT groups, and standard legal/compliance procurement processes as sources of friction in enterprise AI adoption.

  12. 34:22 – 42:11

    Overcoming Procurement Hurdles

    Speaker_1 states that Geminos typically passes enterprise security and procurement checks within days due to built-in best practices, contrasting this with the typical two-to-three-month timeline.

  13. 42:11 – 43:51

    Corporate Venture Capital Conflicts

    Speaker_1 discusses the risks of corporate VC, noting that corporate investors often become risk-averse once the startup becomes a customer, prioritizing their own needs over the startup's broader market growth.

  14. 43:51 – 1:00:00

    Geminos Technical Architecture

    Speaker_1 explains that relying solely on LLMs is insufficient for enterprise scale due to consistency issues, advocating for a neurosymbolic AI approach that combines symbolic, causal, and rules-based engines with neural LLMs.

Assets discussed

AssetKindStanceContext
MicrosoftMSFTequitybullishsold data led growth to Microsoft for 275 million

“Stance” is how the speakers talked about the asset, not a recommendation.

Claims worth verifying

Stuart Frost took Select Software Tools public on the NASDAQ.

Unidentified speaker (speaker_0) ·0:04

How to verify: SEC EDGAR filings for Select Software Tools (ticker: SSWT) listing history. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Stuart Frost sold DataAllegro to Microsoft for $275 million.

Unidentified speaker (speaker_0) ·0:23

How to verify: Microsoft press releases or SEC filings (Form 8-K) regarding the acquisition of DataAllegro. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Stuart Frost has spent nearly four decades building and scaling technology companies.

Unidentified speaker (speaker_0) ·0:04

How to verify: Stuart Frost's LinkedIn profile or biography on the Geminos website. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Only 13% of large enterprises are reporting any measurable value from AI.

Unidentified speaker (speaker_1) ·21:13

How to verify: Gartner research report on AI adoption in large enterprises, specifically looking for the 13% statistic. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Geminos uses causal AI, knowledge graphs, and rules engines to make decisions.

Unidentified speaker (speaker_1) ·8:18

How to verify: Geminos website technical documentation or whitepapers describing their technology stack. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Great Western Railway is a customer of Geminos.

Unidentified speaker (speaker_1) ·25:32

How to verify: Geminos press releases or case studies mentioning Great Western Railway as a client. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Vibe coding is not scalable or reliable for enterprise use.

Unidentified speaker (UNKNOWN) ·28:11

How to verify: Geminos blog posts or marketing materials contrasting 'vibe coding' with their enterprise-grade solutions. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Stuart Frost was selling tools to technologists until last year.

Unidentified speaker (speaker_1) ·19:52

How to verify: Geminos website 'About' or 'History' section detailing their sales motion shift. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

AI can reduce product development time from six hours to six minutes.

Unidentified speaker (speaker_0) ·3:11

How to verify: This is a specific illustrative claim by the host; verification would require internal Geminos metrics or case studies demonstrating this specific speedup ratio. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Causal AI allows testing interventions without live production changes.

Unidentified speaker (UNKNOWN) ·17:56

How to verify: Geminos technical documentation or case studies demonstrating causal modeling capabilities. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Speaker_1 sold DataAllegro to Microsoft for $275 million.

Unidentified speaker (UNKNOWN) ·31:37

How to verify: Check Microsoft's acquisition history or press releases from the time of the acquisition for the transaction value. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Speaker_1 took Select Software Tools public on NASDAQ.

Unidentified speaker (speaker_1) ·29:37

How to verify: Search NASDAQ historical filings or SEC EDGAR for Select Software Tools' IPO filing. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Geminos signed the second biggest contract of Speaker_1's career in the food industry a few days ago.

Unidentified speaker (speaker_1) ·31:11

How to verify: Check Geminos' official website news section or press releases for announcements of food industry contracts around the release date. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

A competitor had a four-month sales cycle for a large enterprise deal.

Unidentified speaker (UNKNOWN) ·31:53

How to verify: This is an anecdotal claim; verify via industry reports on enterprise software sales cycles or Geminos' internal case studies if publicly shared. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Geminos passes security and procurement checks within days.

Unidentified speaker (speaker_1) ·34:22

How to verify: Review Geminos' security compliance documentation (e.g., SOC2 Type II reports) or customer testimonials regarding onboarding speed. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Internal IT groups cannot match the quality of specialized AI startups because AI is not their day job.

Unidentified speaker (speaker_2) ·33:29

How to verify: This is an opinion/claim; verify by comparing Geminos' technical whitepapers or feature sets against typical internal IT department capabilities. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Salesforce acquired Slack.

Unidentified speaker (speaker_2) ·41:38

How to verify: Check Salesforce's press releases or SEC filings for the acquisition of Slack Technologies. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Delay attribution for Great Western Railway took 8-10 days and was reduced to a few minutes.

Unidentified speaker (UNKNOWN) ·53:46

How to verify: Check Geminos' case studies or press releases regarding the Great Western Railway implementation for specific metrics. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

The Great Western Railway use case generates tens of millions a year.

Unidentified speaker (UNKNOWN) ·53:50

How to verify: Check Geminos' financial disclosures or case studies for revenue impact figures. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

LLMs alone will not get us to AGI in any meaningful way.

Unidentified speaker (UNKNOWN) ·46:24

How to verify: This is a prediction/claim; verify by tracking industry consensus or technical papers on AGI roadmaps. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Delaware now allows you to have agent companies, AI companies, with no people responsible for them.

Unidentified speaker (speaker_1) ·57:29

How to verify: Check Delaware state statutes and code regarding corporate formation requirements for AI entities or 'agent companies' to confirm if human responsibility is legally waived. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Jack Dorsey's company (Block) roadmap is now completely delegated to agents.

Unidentified speaker (speaker_1) ·57:55

How to verify: Review Block (Square) press releases, SEC filings (10-K/10-Q), or official statements from Jack Dorsey regarding product development processes and AI integration. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Primary source

Anthropic and OpenAI are racing to justify their trillions in value through AGI and AI consciousness.

Unidentified speaker (UNKNOWN) ·58:44

How to verify: Analyze public statements, investor presentations, and valuation reports from Anthropic and OpenAI to assess their stated goals regarding AGI and their current market valuations. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Stuart Frost has a champion at Great Western Railways who is a leading thinker on AI.

Unidentified speaker (speaker_1) ·56:49

How to verify: Search for public profiles, LinkedIn posts, or industry publications featuring employees of Great Western Railways discussing AI leadership or partnerships with Geminos. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Geminos helps companies arrive at better decisions faster in a deterministic way rather than what LLMs can produce.

Unidentified speaker (speaker_2) ·59:36

How to verify: Review Geminos.ai website materials, whitepapers, or case studies detailing their methodology for deterministic AI decision-making versus probabilistic LLM outputs. For recent events, verify 2026-08-13 through 2026-11-13. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify Gartner's '13% measurable AI value' statistic by checking their latest AI hype cycle or enterprise adoption reports.
  • Review Geminos' technical documentation for their 'neurosymbolic AI' architecture to understand how causal models integrate with LLMs.
  • Check Great Western Railways' public reports or press releases for the claimed reduction in delay attribution time and associated financial value.
  • Investigate Delaware corporate law changes regarding AI companies with no human responsibility to assess legal risks.
  • Analyze the portfolio performance of corporate VCs (e.g., Salesforce Ventures, Intel Capital) to evaluate the 'market of one' conflict claim.
  • Search for news regarding Block (Jack Dorsey) delegating product roadmaps to AI agents to verify the 'sustaining engineering' critique.
  • Review Geminos' security compliance certifications (SOC2, ISO) to validate their claim of rapid procurement passage.
  • Compare typical enterprise AI sales cycles (2-3 months) against Geminos' claimed 'days' timeline using industry benchmarks.
Full transcriptexpand
0:04
Unidentified speaker (speaker_0): Welcome to Scaling with Clarity. Today's guest is Stuart Frost, a serial entrepreneur who spent more than three decades building the scaling technology companies. He took select software tools public on the Nasdaq, sold data led growth to Microsoft for 275 million and went on to build and
0:23
Unidentified speaker (speaker_0): incubate dozens of startups. Today, as founder and CEO of Geminos, he's tackling one of the biggest challenges in business. How to turn artificial intelligence from impressive demonstration into reliable, measurable results. We're exploring why enterprises struggle
0:43
Unidentified speaker (speaker_0): with AI, what they're getting wrong, and what it takes to make technology genuinely useful. Stuart, welcome to the show.
0:51
Unidentified speaker (speaker_1): Thanks, Maxon. Thanks for having
0:53
Unidentified speaker (UNKNOWN): me.
0:54
Unidentified speaker (speaker_0): I am super excited about the conversation today. I mean, I obviously you bring in uh
1:00
Unidentified speaker (UNKNOWN): a
1:00
Unidentified speaker (speaker_0): pedigree of talent and pedigree of experience that like other founders will find useful, but uh what I wanted to start, uh Stuart is kind of like the the back to the origin story. So um you spend much of your professional life building companies as as a written interview you've you've taken a a company
1:20
Unidentified speaker (speaker_0): public you sold another one to microsoft then you you had a venture studio um what is it about building businesses that just keeps pulling you back why aren't you retired like like I mean we you made a ton of money like obviously there's
1:36
Unidentified speaker (UNKNOWN): some kind
1:36
Unidentified speaker (speaker_0): gonna of be a the gravitas gravity that's pulling you back
1:39
Unidentified speaker (speaker_1): Uh it might be insanity, I'm not sure. Um, it's a good question. Uh it's either insanity or addiction, right? And maybe it's a fine line between the two. Yeah. Um, I'll be honest, I just really enjoy it, right? I I mean, I think especially at a time like this, where we've got the opportunity to reinvent Invent the way we build startups. I
1:59
Unidentified speaker (speaker_1): I've actually been writing a few blog posts recently on not just the software development lifecycle, but the whole product development lifecycle, how that's changing, and also how the sales motion is changing. And you know, think about it, those two things, RD costs and sales costs, and the challenge of scaling both,
2:19
Unidentified speaker (speaker_1): have been the kind of death of many a good startup, right? Um so the opportunity to change that dynamic and and massively. reduce the cost and the time involved in both of those activities, uh it's it's fascinating to me, right? And and bringing that to bear and using that as we uh develop Jabinos,
2:39
Unidentified speaker (UNKNOWN): yeah,
2:39
Unidentified speaker (speaker_1): it's fascinating. So what what else would I do?
2:43
Unidentified speaker (speaker_0): Yeah, no, it's uh it it it it's it's strange. Like um I wake up every morning.
2:51
Unidentified speaker (speaker_0): just can't like I can't wait to start working and and and to a lot of people this seems very very foreign for me it's like I just can't like it's like it's literally I find that the current time the current environment it is so formative that like I feel like if I'm not Oh, I'm missing out and it's like I'm
3:11
Unidentified speaker (speaker_0): a kid in a candy store, I just want to play with tools, I want to build like and whatever I could do yesterday in six hours, now I can do in six minutes. It's just insane.
3:22
Unidentified speaker (speaker_1): Absolutely right. I mean, you know, I as you said I've been doing this for over 30 years. I think it's almost 40 years now, actually. Um, and you know, a lot of the things that we've learned in terms of market penetration and so on, the whole Jeffrey Moore stuff, crossing the chasm, and all that, been I've a huge fan of that. I've applied it successfully in startups.
3:42
Unidentified speaker (speaker_1): Um but the sheer speed of what we can do
3:45
Unidentified speaker (UNKNOWN): now
3:46
Unidentified speaker (speaker_1): is changing that. Um not you don't necessarily throw it all out, right? But you it definitely needs to be reconsidered and uh and see how we can, especially on the sales side, right? I mean, often with such a long sales cycle, it might take you a year to realize that a salesperson is not really cutting it, they're not getting the messaging right,
4:05
Unidentified speaker (UNKNOWN): because
4:05
Unidentified speaker (speaker_1): it's only going to show up if they close deals or not. You can't be on every call, otherwise
4:10
Unidentified speaker (UNKNOWN): do
4:10
Unidentified speaker (speaker_1): itself as the founder, right? Um now we can look at every transcript or AI can look at every transcript. Uh it's a good quality transcript, which is that's helpful, right? So we can get a much better view of what the customer's asking for, what the salesperson is telling them. Uh, you know, are we getting our message out there? Um
4:30
Unidentified speaker (speaker_1): and all the other aspects that come with that. Um and and catch it much sooner and train people appropriately. It's not a matter of firing, it's a matter of can we help this person get the right messaging? Um so I'm I'm looking at all of that and actually using the Geminos Geminis technology to build a knowledge graph of Geminos that underpins all that
4:50
Unidentified speaker (speaker_1): work as well. So the AI deeply understands what we do as a company and what our customers need.
4:56
Unidentified speaker (speaker_2): Yeah.
4:57
Unidentified speaker (speaker_1): And let's review every sales core,
5:00
Unidentified speaker (UNKNOWN): you
5:00
Unidentified speaker (speaker_1): know, maybe in in real time before too long, right? Uh and give people hints as to how to best serve the needs of the customer and get
5:08
Unidentified speaker (UNKNOWN): redo
5:09
Unidentified speaker (speaker_1): that as quickly as possible. Right.
5:11
Unidentified speaker (UNKNOWN): Well,
5:11
Unidentified speaker (speaker_0): there's no reason why you shouldn't be doing this, especially like I mean, everyone's talking about the the company brain, you're talking about knowledge craft, but like there's so much data, there's no reason why you're not pulling this data and turning it turning it into actionable insights or signals that you can act upon.
5:29
Unidentified speaker (speaker_1): Yeah, absolutely right. And so that's the thing. I think so far there's been a lot of effort obviously on the software development lifecycle and how much faster that's become and so on. Um, and and for me, the other side of this is that product development lifecycle. It's you know, ideas from my head and and in conversations with customers, how
5:49
Unidentified speaker (speaker_1): quickly can we now develop those and turn them into real shippable products? That's like hundreds of times faster than before, with far fewer resources, right? So you've got to leverage that. You you've got to figure out how to use that uh in your everyday business life and
6:09
Unidentified speaker (speaker_1): and how to get the power of that as well to
6:13
Unidentified speaker (speaker_3): be
6:13
Unidentified speaker (speaker_1): pervasive across the company. And this is not about having agents running the company and making all the decisions. It's about helping us as human beings make much faster decisions, get stuff out to customers quicker. and and reduce the sales cost dramatically as well, right? Because so
6:32
Unidentified speaker (UNKNOWN): much
6:32
Unidentified speaker (speaker_1): gets wasted.
6:38
Unidentified speaker (speaker_0): Now obviously everyone has moved to or
6:41
Unidentified speaker (UNKNOWN): is
6:41
Unidentified speaker (speaker_0): in the process of moving to AI SDLC. I want to double click on this and get your perspective on just two elements. The first element is
6:53
Unidentified speaker (UNKNOWN): the
6:53
Unidentified speaker (speaker_0): principles around SaaS versus agentic. Kind of like if you're building products, how are you building? Is everything that you're building now to be agentic first? And two, um,
7:07
Unidentified speaker (speaker_0): what's your thinking frame of reference around customer discovery? Because
7:13
Unidentified speaker (UNKNOWN): just because
7:13
Unidentified speaker (speaker_0): you build, you can build five. faster doesn't mean you're building the right things. So can I give me your perspective around is it agenda first? And how do you go about customer discovery?
7:26
Unidentified speaker (speaker_1): The end of the day, I don't think any technology should be first.
7:33
Unidentified speaker (speaker_1): Just fundamentally. Um doesn't matter how powerful it is, That shouldn't be the thing that you talk about first, right? I think for me, what I want to deeply understand is what is the job that my customer needs to
7:47
Unidentified speaker (UNKNOWN): do,
7:48
Unidentified speaker (speaker_1): and how do we improve
7:50
Unidentified speaker (UNKNOWN): that?
7:50
Unidentified speaker (speaker_1): Right. Um, and what value can the customer get from that?
7:54
Unidentified speaker (UNKNOWN): Yeah.
7:55
Unidentified speaker (speaker_1): And then only then do we say, okay, which parts of the technology stack do we need to solve that problem? Yeah.
8:02
Unidentified speaker (UNKNOWN): Now, Of course
8:02
Unidentified speaker (speaker_1): there are lots of common patterns here, right? We tend to deal with customers that have a lot of text, a lot of data, structured and unstructured. How do we gather knowledge from that? And really importantly, how do we make decisions based on that information?
8:18
Unidentified speaker (speaker_1): So there are some aspects then of the technology stack that we know we're going gonna to need. We're we're going gonna to need an LM to understand the text. We're going gonna to need agents to run the workflow.
8:27
Unidentified speaker (speaker_1): Um, we're gonna need causal AI to make the final decision. Um, and we're gonna need a knowledge graph underpinning it all that helps us actually understand all of that and and get get the accuracy that we need. Um and also a rules engine, uh very important. Um,
8:42
Unidentified speaker (UNKNOWN): you know,
8:42
Unidentified speaker (speaker_1): it's no good the L LM just going off and and and agents going off and making decisions based on guardrails, which mean very little. in practice, the hints at best. We need to enforce stuff with a real rules engine. Um
8:57
Unidentified speaker (speaker_1): it's that combination of factors. So is that agentic first? Not really. It's what's the customer's problem? Let's piece everything together appropriately then and use the right form of technology, right form of AI
9:11
Unidentified speaker (UNKNOWN): to
9:11
Unidentified speaker (speaker_1): address each part of that problem.
9:14
Unidentified speaker (UNKNOWN): Um
9:15
Unidentified speaker (speaker_0): For for those of our listeners that uh
9:18
Unidentified speaker (UNKNOWN): uh
9:20
Unidentified speaker (speaker_0): need a little bit more clarity around the rules engine, can
9:23
Unidentified speaker (UNKNOWN): you
9:23
Unidentified speaker (speaker_0): jax depose safety guardrails or guardrails um in a model versus rules engine and kind of like you know, suggestions versus a row? Kind of like what well does it mean?
9:35
Unidentified speaker (speaker_1): So rules engines have been around a long time, right? Um in terms of yeah, there's plenty of an open source and that kind of thing. And you know what they do is given a certain set of data, they'll tell you the boundaries that you can operate in. And that might be for a piece of equipment. There might be hard limits on temperature and pressure, for example. And you don't want
9:55
Unidentified speaker (speaker_1): the thing to blow up if you exceed those. So pretty fundamental. Um, And of course, you know, agents, etc
10:02
Unidentified speaker (speaker_1): right now in the raw should not be used for that kind of thing. Right. Because they don't always reliably follow guardrails and suggestions and prompts. So just it's just a no-go area. So what we've done is built in one of those traditional rules engines, um, so that
10:21
Unidentified speaker (UNKNOWN): you know
10:21
Unidentified speaker (speaker_1): we can check and and the the harness around the agents that we build when we're building a solution. of the code will check the rule and not just let the agent uh you know make that decision. It'll it'll stop it before it does. So that's for me the big distinction, right? These are hard and fast
10:41
Unidentified speaker (speaker_1): actual rules. Yeah. You know, in a traditional rules engine, there's been yeah, that's it's decades of uh of work have gone into those things. Um why not reuse that?
10:51
Unidentified speaker (speaker_0): Yeah.
10:52
Unidentified speaker (UNKNOWN): You know,
10:53
Unidentified speaker (speaker_1): because trying to build that into agents is not working very well because the very nature of the technology underneath them like we get steadily better. But
11:03
Unidentified speaker (UNKNOWN): I
11:03
Unidentified speaker (speaker_1): I just think if technology's there and it's and it's proven, um why not use it? The neat thing is now we can use AI to discover rules. So we can, for example, go through contracts and go through uh manuals and that kind of stuff and extract rules and populate the rules engine with that and then curate those, check them, uh, and so on. So
11:23
Unidentified speaker (speaker_1): we can make that whole process a lot faster, a lot more efficient, and then you know deploy the engine to solve the problem.
11:31
Unidentified speaker (speaker_0): Yeah, it's amazing what it can do. Um
11:35
Unidentified speaker (speaker_0): recently I was trying to build a keepability capability capability map. for for the team okay we are hearing now if the outcome we're seeking is there what do we need to do from a team capability but respect to be able to equip the team and so I just like I I asked the AI that we use is like okay go and inventory all the work
11:55
Unidentified speaker (speaker_0): we do who does it what are the handovers from system to from machine to machine or machine to human human to machine and did a pretty amazing job in identifying this. And then obviously, this earth is uh the input in terms of us kind coming up with a capability map, but it's amazing what it can do. Um
12:15
Unidentified speaker (speaker_0): when you have structured it the right way. Now, on the structure side, um I have uh massively adopted personal agents,
12:26
Unidentified speaker (speaker_1): and
12:26
Unidentified speaker (speaker_0): so
12:27
Unidentified speaker (speaker_0): Instinct is my favorite, but I use other like I use Muse, I use quad. Um
12:35
Unidentified speaker (speaker_0): I was playing around over the last couple of days trying to see can I put a harness? Can I put some guardrails and rules? And so
12:46
Unidentified speaker (speaker_0): Instinct flat out told me that there's no way that it can do Muse is a little bit more even more broke when it comes to like so
12:56
Unidentified speaker (speaker_0): from your perspective, how how is there a way that can you define your rules at a global level? And so any technology you plug in that says these are my rules, stick to my rules, don't break the rules. Is there a way to do so? Or
13:11
Unidentified speaker (UNKNOWN): how
13:11
Unidentified speaker (speaker_0): do you make personal agents work?
13:14
Unidentified speaker (UNKNOWN): Well, the
13:14
Unidentified speaker (speaker_1): problem is. right, you're with a personal agent, you're delegating responsibility for doing most of the work to that agent. And as you rightly say, right, it's
13:24
Unidentified speaker (UNKNOWN): it's
13:24
Unidentified speaker (speaker_1): got a mind of its own well, it doesn't, but you know what I mean, right? Yeah. In in this context at least. Um and and it's it's a fundamental thing, isn't it? You give an agent a goal and then you give it guidelines, really, right? They're not guardrails,
13:38
Unidentified speaker (UNKNOWN): guidelines
13:38
Unidentified speaker (speaker_1): as to what to do and what not to do but the goal is always primary in i to an agent so it will ignore those uh guidelines if if
13:49
Unidentified speaker (UNKNOWN): it
13:50
Unidentified speaker (speaker_1): chooses to to to get if it thinks it's necessary to get to the goal
13:53
Unidentified speaker (UNKNOWN): yeah so
13:54
Unidentified speaker (speaker_1): what we're doing is really inverting that um and we're saying look if there is a goal for the business that's got to be optimized let's build a proper causal model for that because the underpinning math there is much more powerful than an agent will ever be. Let's use agents to feed it and to understand and
14:14
Unidentified speaker (speaker_1): manage the workflow around it. But at the end of the day, let's do a proper data-driven decision that you know drives the business outcome uh that the customer needs and use the agents in the L. LLMs then to, like I say, you know, understand the text, understand the workflow, gather the information,
14:32
Unidentified speaker (speaker_2): yeah,
14:33
Unidentified speaker (UNKNOWN): and
14:33
Unidentified speaker (speaker_1): feed it in. Uh so let's use the an appropriate form of technology that is well proven now. The map works great on cause. We've been doing it for years for all in all kinds of different industries. Why try and use an agent for that I when mean it just doesn't work?
14:50
Unidentified speaker (UNKNOWN): You know,
14:51
Unidentified speaker (speaker_1): probably won't for quite a while, right?
14:53
Unidentified speaker (UNKNOWN): Yeah.
14:54
Unidentified speaker (speaker_0): Stuart, in in our last conversation, you talked about that you primarily work with companies or divisions of companies with a billion plus and dollars in revenue. And so we're talking big companies with with big resources and capabilities. Um, from your perspective, how many of these companies actually make
15:14
Unidentified speaker (UNKNOWN): well
15:15
Unidentified speaker (speaker_0): informed substantiated decision and how quickly can they get to clarity to make those decisions?
15:24
Unidentified speaker (speaker_1): So
15:25
Unidentified speaker (speaker_1): if you're talking about day to day operational decisions, right? Um
15:30
Unidentified speaker (speaker_0): yeah, like I mean, how do I know like that I'm doing the right capital allocation? Like, is this the right company to acquire? Like, like obviously, the bigger the decision, the more vigor it should have. And and some of it is art, some of it is science, but like I feel like
15:46
Unidentified speaker (speaker_0): there's too much reliance on the art and not really not enough reliance
15:50
Unidentified speaker (UNKNOWN): on
15:51
Unidentified speaker (speaker_0): the science, or the science is dated.
15:53
Unidentified speaker (speaker_2): Yeah.
15:53
Unidentified speaker (speaker_1): Well, I mean, today, like if you go into most industrial companies or any kind of company really, and they say, you know, yeah, we you we do data-driven decisions. What they really mean is they've got some spreadsheet that's full of macros and stuff, and they do some BI, right? And maybe we use AI on top of that to help. And a lot of those macros in the spreadsheet
16:13
Unidentified speaker (speaker_1): were written 20, 30 years ago by someone who's no longer at the company that breaks all the time, but no one really knows, right? Um, so they're not really making truly data-driven decisions at all. What they're doing is applying a lot of expertise through that, that's still in humans' heads, right? Somewhat aided by the data. Too much gut
16:33
Unidentified speaker (speaker_1): field, too much, too many emotional decisions.
16:37
Unidentified speaker (UNKNOWN): You
16:37
Unidentified speaker (speaker_1): know, we saw this in in a in a supply chain uh a couple of years ago where um if a shortfall of some particular inventory, some component fell short and stopped the production line, the person responsible for that got yelled at pretty hard, right? Somebody shut a multi-billion dollar production line down for
16:57
Unidentified speaker (speaker_1): a week while that part was sourced and so on. So what they tend to do then is make sure that part is over
17:05
Unidentified speaker (speaker_1): stocked, right? So that never happens again. Um or they, you know, there's a lot of rush shipping and everything to solve that problem. Um
17:13
Unidentified speaker (UNKNOWN): what
17:14
Unidentified speaker (speaker_1): you need to do to to have a truly data-driven decision, to my mind, is have a math that actually supports that, um, and an engine that supports that. That's where the causal AI really comes in. Um, so
17:25
Unidentified speaker (UNKNOWN): you know,
17:26
Unidentified speaker (speaker_1): and it's brilliant. You you can set an outcome, understand cause effect chain that kind of gets you there. Business people can understand that, it's in their terms, it's it's their variables, it's their business. And then we can put the math underneath it with the uh data history and so on that they give us.
17:44
Unidentified speaker (UNKNOWN): And
17:44
Unidentified speaker (speaker_1): the neat thing then is if an outcome's changing in a bad way, good way, whichever,
17:49
Unidentified speaker (UNKNOWN): yeah, you
17:49
Unidentified speaker (speaker_1): can see what the root cause is of that, but also you can test interventions. So I can say if I change the temperature by
17:56
Unidentified speaker (UNKNOWN): two
17:56
Unidentified speaker (speaker_1): degrees on this industrial process, what's gonna going to happen?
17:59
Unidentified speaker (UNKNOWN): Right? Is it
18:00
Unidentified speaker (speaker_1): gonna going to fix the problem or not?
18:02
Unidentified speaker (UNKNOWN): And I can play around.
18:03
Unidentified speaker (speaker_1): with that in the math without actually necessarily doing it in live production.
18:09
Unidentified speaker (speaker_1): So for the first time, you've got this ability to do truly data-driven decisions.
18:15
Unidentified speaker (UNKNOWN): What
18:16
Unidentified speaker (speaker_1): we've then got to do is get there as quickly as possible in terms of gathering the data, gathering the information, might come from a lot of text, might come from uh some structured data as well. That's where I think the power of LLMs and agents really comes in, then to gather that information, make sure it's you know
18:36
Unidentified speaker (speaker_1): fits the rules, et cetera. We can check the rules against it as we go, um, and then feed it into the causal model, and we end up with truly data-driven Driven decisions for really the first time in history, I think, you know, at any kind of scale.
18:51
Unidentified speaker (UNKNOWN): So
18:52
Unidentified speaker (speaker_2): when you're having this conversation with customers, I mean obviously that that uh makes a hundred percent sense. Like, why would you not want to get there? Um, what's the biggest hurdle?
19:02
Unidentified speaker (speaker_0): What is the biggest hurdle? Like what is perhaps stopping companies from saying yes to
19:08
Unidentified speaker (UNKNOWN): this?
19:10
Unidentified speaker (speaker_1): So noise, noise in the
19:12
Unidentified speaker (UNKNOWN): market,
19:13
Unidentified speaker (speaker_1): right? So they're bombarded with oh we can vibe code this, right? We can LMs are gonna get there eventually, they're gonna understand causality, no, they're not. Um and it's so for the typical business exec who doesn't have time to, you know, understand what the latest model can do and what it means and everything, right?
19:33
Unidentified speaker (speaker_1): There's bombarded with stuff. Um and it's very hard to pick out then, you know, a a better way. forwards. Um now, once we get to them and we explain what we do, frankly, it's a really high hit rate and it doesn't take us very long. That's the beautiful thing about this, you know, selling tools to technologists, which we were
19:52
Unidentified speaker (speaker_1): until last year, is a long sales cycle, right? So you've got to persuade them that you're always better. And we were ending up building solutions for the business just to prove that it was better. which is great. But you know, we're still same hurdles and some into the technology side of it. So into the business, we can go in and say, let us show you. We'll
20:12
Unidentified speaker (speaker_1): build something in a few days that demonstrates to you that we can solve your problem. It's not going to be the final app. We're not vibe coding this, we're doing it right, but we'll show you, you know, on a a basic knowledge graph with a basic causal model.
20:27
Unidentified speaker (UNKNOWN): that
20:27
Unidentified speaker (speaker_1): it drives the right outcome for you, and and we'll explain that in your terms. And once we do that, they're solved, right? Because they they don't get that from anything else. If someone's vibe code is something there's too many holes. It's not gonna be a recurring solution to the problem.
20:43
Unidentified speaker (speaker_0): Yeah.
20:44
Unidentified speaker (speaker_1): And we can we can show that. Um so that side of it goes really quickly. You We sort started of to get through contracting and procurement, but you know
20:51
Unidentified speaker (speaker_1): it wasn't with us.
20:53
Unidentified speaker (speaker_0): Yeah, yeah,
20:53
Unidentified speaker (speaker_2): yeah. For sure. Um
20:56
Unidentified speaker (UNKNOWN): out
20:56
Unidentified speaker (speaker_2): of the C-suite, which persona is the easiest to convince? Uh CEO, C O O, kind of like what's your what's it take?
21:06
Unidentified speaker (speaker_1): So it's basically I would say the CEO or the head of a business unit, right? Um
21:13
Unidentified speaker (speaker_1): so it's they've been tasked typically by the board to do something cool with AI, show some value. Yeah, they're struggling with it. Gartner just put out numbers. I think only 13% of large enterprises are reporting any value, measurable value from AI. So very early in this, right? And that value could be really small in one
21:33
Unidentified speaker (speaker_1): division. So there's a lot to go at, and people are struggling with this.
21:37
Unidentified speaker (UNKNOWN): Um,
21:38
Unidentified speaker (speaker_1): and I think. we've got a great answer to that problem. And yeah we can go in and say, look, you know, co-pilots and everything. It's cool. You know, you can maybe write a PowerPoint a bit quicker, whatever. But in give us your key decision that you your business needs to take every day.
21:56
Unidentified speaker (UNKNOWN): Yeah. It's
21:57
Unidentified speaker (speaker_1): done at some scale, right? It involves a few people.
22:00
Unidentified speaker (UNKNOWN): And
22:01
Unidentified speaker (speaker_1): let us show you how we can get that down from taking dates, because it's a lot of paperwork, phone calls, et cetera, multiple data sources. We'll get it down to minutes. And we'll do it with best in class accuracy. Not perfect. Not claiming that,
22:17
Unidentified speaker (UNKNOWN): but
22:17
Unidentified speaker (speaker_1): it's going to be as good as your, in most cases, good as your very best employees, right? Yeah.
22:23
Unidentified speaker (UNKNOWN): So
22:23
Unidentified speaker (speaker_1): we're leveling up your organization to the very best levels. giving you more traceability on it, more transparency on those decisions, and doing it, you know, in a thousandth of the time. All right. But we'll go from eight days to eight eight minutes, you know, it's it's a big difference. Right. Um for
22:41
Unidentified speaker (speaker_0): sure.
22:42
Unidentified speaker (speaker_1): Yeah. So that's that's the kind of fundamental story. And once you get to the business person that has responsibility for that type of decision
22:53
Unidentified speaker (UNKNOWN): it's
22:53
Unidentified speaker (speaker_1): not a hard sell.
22:57
Unidentified speaker (speaker_0): Makes sense.
23:02
Unidentified speaker (speaker_0): One thing that I have been thinking about a lot is
23:09
Unidentified speaker (UNKNOWN): I
23:09
Unidentified speaker (speaker_0): I mean obviously like
23:11
Unidentified speaker (UNKNOWN): uh
23:11
Unidentified speaker (speaker_0): getting in front of getting getting in front of the economic buyer buyers, having a conversation, uh, demonstrating the value, that's clearly it it it makes it makes total sense. Uh you're building a solution. Now,
23:24
Unidentified speaker (UNKNOWN): the
23:25
Unidentified speaker (speaker_0): one challenge I have with this model is it's predicated upon a a single party, or maybe like a lot of those organizations will have uh
23:34
Unidentified speaker (speaker_2): some form of an innovation and function and they kind of centralize
23:37
Unidentified speaker (UNKNOWN): the
23:37
Unidentified speaker (speaker_2): innovate in the innovation. innovational transformation thinking. Um
23:43
Unidentified speaker (speaker_2): what I've seen work really well is
23:48
Unidentified speaker (UNKNOWN): but
23:49
Unidentified speaker (speaker_2): it's predicated upon first building the foundation, having that central function, and then kind of federating
23:54
Unidentified speaker (UNKNOWN): across is
23:55
Unidentified speaker (speaker_0): like
23:55
Unidentified speaker (speaker_2): once you move from a centric model to a federated model where you put in the the know-how, the technology,
24:01
Unidentified speaker (speaker_0): the tools, like the um in in the hands of the operators and then uh at least my in my experience then at that point it started to spread like wildfire because people now can model build things on their own and that's not to say that they do that they're doing it without guardrails or um they just
24:23
Unidentified speaker (speaker_0): the they're testing hypothesis continues continuously. that oh can can I try I
24:25
Unidentified speaker (UNKNOWN): this try this, can can try I
24:26
Unidentified speaker (speaker_0): this try can this, try can this I try this, can try this. And I I find that that's that that at least from my experience, that's the inflection point that they see, like where they move from like, oh, I'm gonna build one model and yes, that's gonna drive value for me. No no question that what you're proposing would drive a lot of value for them. But how do you get this from like I don't know,
24:46
Unidentified speaker (speaker_0): 10x to a thousand X?
24:49
Unidentified speaker (UNKNOWN): Well,
24:49
Unidentified speaker (speaker_1): it's one of the in interesting challenges, right? And I think the state of the technology right now is such that we can get people up to speed on building these solutions using this platform pretty quickly.
25:03
Unidentified speaker (speaker_2): Yeah.
25:03
Unidentified speaker (speaker_1): Um
25:05
Unidentified speaker (UNKNOWN): now
25:05
Unidentified speaker (speaker_1): we ha we're not yet really offering that to customers, but that is the next step.
25:12
Unidentified speaker (UNKNOWN): One
25:12
Unidentified speaker (speaker_1): of the things we're doing is in the next phase of RD on our side is really understanding where the bottlenecks have been over the these first set of projects that we've been doing. And what do we need to grind down on in terms of really automating more of that and so on. So there's a few areas there. So and
25:32
Unidentified speaker (speaker_1): I think once we've done that, it'll make our costs lower. So therefore we can make it cheaper for customers, etc. But also I think it'll be a more recurring offering to a customer. There's another aspect to it that actually we've been talking to one of our existing customers, uh Great Western Railway about, where and and
25:52
Unidentified speaker (speaker_1): it's interesting you mentioned hypothesis, right? Because one of the big things with causal modelling is you it's a more scientific approach to solving these kind of problems. And one of the things that they want to do is have their line managers be able to set a hypothesis and test it using the platform.
26:10
Unidentified speaker (speaker_1): So we're trying to figure out how we
26:15
Unidentified speaker (UNKNOWN): I
26:15
Unidentified speaker (speaker_1): don't industrialize that, I suppose, right? Because you know, cause of modeling at the high level of hey, you know, these variables change, that one changes, okay. A business person, business analyst can certainly do that. Once you get into the detail of the math, it becomes more data science-like, right? You need more training to do that right. Um, so
26:35
Unidentified speaker (speaker_1): we're trying to figure out how to make that more recurring and and more automatable. Um, so that that is one of the key challenges, actually, that we're looking at. You know, how do you put this whole platform in the hands of business analysts as opposed to our forward deployed engineers
26:55
Unidentified speaker (speaker_1): and make it more flexible like that? We're not quite there yet. Uh, gotta be honest with you, but it's something that's definitely something I I'm putting quite a bit of thought into as we go forwards.
27:07
Unidentified speaker (UNKNOWN): Um,
27:08
Unidentified speaker (speaker_1): kind of everything. It's not uh, you know, vibe coding can be put in everyone's hands, but it doesn't really work in any kind of recurring. way. It's gonna be really interesting on our side as we build these knowledge graphs for customers
27:22
Unidentified speaker (UNKNOWN): and
27:22
Unidentified speaker (speaker_1): they start to become wider and more have more knowledge in them about the the company. Does that enable us to
27:30
Unidentified speaker (UNKNOWN): put
27:31
Unidentified speaker (speaker_1): more power in the hands of of those kind of line managers and business analysts to do it themselves more? for. Um interesting. question that I don't have the full answer to yet, but we're definitely exploring it.
27:42
Unidentified speaker (speaker_2): Yeah, for sure.
27:46
Unidentified speaker (UNKNOWN): Did
27:47
Unidentified speaker (speaker_0): the Garten study offer any kind of insights as to why companies fail to realize the the the benefits and value of of AI? And I I'm just
27:58
Unidentified speaker (speaker_2): kind of curious.
27:59
Unidentified speaker (speaker_1): Yeah.
28:00
Unidentified speaker (UNKNOWN): I
28:00
Unidentified speaker (speaker_1): mean a lot of it is the the actual penetration of AI into the organizations is is still at the copilot level in a lot of cases, right? Um
28:11
Unidentified speaker (UNKNOWN): and
28:11
Unidentified speaker (speaker_1): I think it's it's just a fundamental issue of vibe coding. Once you take vibe coding into the enterprise, it's great at our level, right? You can create some awesome stuff, you know, very quickly.
28:21
Unidentified speaker (UNKNOWN): Is
28:21
Unidentified speaker (speaker_1): it scalable? Does it is it reliable? Is it does it have any kind of governance security? etc.? Well, probably not. Um, so it's simply not going to work at any kind of enterprise scale, right? Um, and we're a long way from from making that work. Um,
28:37
Unidentified speaker (UNKNOWN): so
28:38
Unidentified speaker (speaker_1): I just think what's being offered to the enterprise at the moment isn't by other people, not general, um, but it's just isn't fit for purpose, right? It's not good enough. Um, so yeah, it's understandable that they're going to gonna have pockets of success at best,
28:53
Unidentified speaker (UNKNOWN): and
28:53
Unidentified speaker (speaker_1): the penetration is is still very low. Um
28:57
Unidentified speaker (UNKNOWN): you know,
28:58
Unidentified speaker (speaker_1): which obviously creates a great opportunity for us because you know, I I I don't feel like and and you know, in conversations with Gartner analysts, right? They're confirming this. Wow, you guys are putting
29:08
Unidentified speaker (UNKNOWN): all
29:08
Unidentified speaker (speaker_1): the pieces together
29:09
Unidentified speaker (UNKNOWN): uh
29:09
Unidentified speaker (speaker_1): in the right way. This is really unique and compelling, and you know, as you go forwards and get more evidence of this and more referenceability and so on, this is gonna be very valuable and very, you know, very breakthrough. Um,
29:24
Unidentified speaker (speaker_2): for sure,
29:25
Unidentified speaker (speaker_1): which is fantastic for me as a founder to hear that, right? Um, so yeah, I I feel like that landscape is is perfect for where Geminos Jaminos is at right now,
29:37
Unidentified speaker (speaker_1): and if it that also gives gets me out of bed in the morning, that excitement, right, to go and and talk to customers about it and prospects about it and uh solve some problems for
29:46
Unidentified speaker (UNKNOWN): them. I
29:46
Unidentified speaker (speaker_2): mean, when it comes to big enterprise, probably the the key operative worth is trust.
29:53
Unidentified speaker (speaker_2): Um
29:57
Unidentified speaker (speaker_2): for for them it's more about the protecting the the risk of the downside than realizing the risk from the upside. And so um trust becomes really important. I'm going to if I'm going to bring a tool, I need to know that it's going to work and work within my um
30:16
Unidentified speaker (speaker_2): established tolerance limits. Anything outside of that I'm not willing to tolerate.
30:26
Unidentified speaker (speaker_2): And even like co-pilot cowork is not enabled. So just co-pilot chat is enabled to you like, but why? Like it follows your Active Directory settings, like it it it follows the
30:38
Unidentified speaker (UNKNOWN): persona
30:38
Unidentified speaker (speaker_2): you get like why not co work? Um
30:43
Unidentified speaker (speaker_1): but you know, like a a lot of these large organizations have the issue of friction is built up in the organization.
30:51
Unidentified speaker (speaker_1): because of that problem, right? That they don't trust everything that comes through the door. So over time bureaucracy grows and and people get put in positions of gatekeeping and and a position of power like that. So they block a lot of things that could add an awful lot a of value to the uh the business. So
31:11
Unidentified speaker (speaker_1): in situations like that, we qualify out relatively quickly, quite honestly, because there's so much opportunity out there, so many different markets we can go to,
31:20
Unidentified speaker (UNKNOWN): so
31:20
Unidentified speaker (speaker_1): many different customers we can go to. If you're gonna put that much friction in the way, then we can't can't get there. And we actually signed a uh the second biggest contract of my career, which is quite a decent one, um, in the food industry just a few days ago. And we we'd
31:36
Unidentified speaker (speaker_2): nice nice
31:37
Unidentified speaker (UNKNOWN): to
31:37
Unidentified speaker (speaker_1): to one one of of the their competitors, competitors. our I'll I'll remain nameless. But and we talked to one of the competitors six months before, we'd gone quite a long way. We'd done a sample app.
31:47
Unidentified speaker (speaker_1): And they were just putting too many barriers in the way to actually getting the deal done. So we
31:53
Unidentified speaker (UNKNOWN): said,
31:53
Unidentified speaker (speaker_1): okay, you know, it's not for us. We can't work with you. We'll go elsewhere. Thank you very much. And yeah, I saw the company, I think it's a four month sales cycle. you know, for a a pretty large enterprise deal, multiple use cases, a lot of long term potential. Um
32:13
Unidentified speaker (speaker_1): you know it's that's the kind of organization that that is so much easier for us to work with as a startup, right? For
32:20
Unidentified speaker (UNKNOWN): sure.
32:20
Unidentified speaker (speaker_1): People who see the challenge look at it but and then make a decision and get on with it. Uh and don't put a lot of gates in the way that are unnecessary. Do all the diligence and everything, don't get me wrong. It's been
32:34
Unidentified speaker (UNKNOWN): very
32:34
Unidentified speaker (speaker_1): thorough, been great to work with.
32:36
Unidentified speaker (UNKNOWN): And
32:36
Unidentified speaker (speaker_1): then they did it and signed it and got on with it, you know, and it's it's refreshing. Yeah.
32:42
Unidentified speaker (speaker_2): So I I I just I just wanted to get a little more clarity on this. Where do you think the greatest amount of friction comes? from? Is it that bureaucracy, the inertia, the organizational culture, leadership, fear of changing how things are done? Like like
33:03
Unidentified speaker (speaker_2): w like what is what what is your sense as to where the friction, the resistance comes
33:07
Unidentified speaker (UNKNOWN): from?
33:09
Unidentified speaker (UNKNOWN): You
33:09
Unidentified speaker (speaker_1): know, it's different in different places, but uh you know some areas you even see unions getting involved. and scared of what AI is going to do to jobs. So there's some friction there that slows things down. Some of it is, you know, technology groups that think they want to do it themselves, they're excited to do
33:28
Unidentified speaker (speaker_1): AI.
33:29
Unidentified speaker (speaker_2): Yeah.
33:29
Unidentified speaker (speaker_1): Um,
33:31
Unidentified speaker (UNKNOWN): you know,
33:31
Unidentified speaker (speaker_1): the reality is it's not their day job, day in, day out, to do this. We're it is ours. Um, so the quality and and the features we can bring. to bear are let's just say a little beyond, um, what what they're capable of doing. Um it's not just a matter of plugging together vibe coding around some open source here, right? This is
33:48
Unidentified speaker (speaker_2): this
33:48
Unidentified speaker (speaker_1): is visionary stuff, I like to think. And uh you know, that day to day driving of that roadmap forwards is is a big deal. Uh and no internal IT group is ever gonna match that. Um, my humble opinion. So um
34:02
Unidentified speaker (UNKNOWN): with
34:02
Unidentified speaker (speaker_1): that, you know, uh and then there's just the regular legal compliance uh procurement stuff that almost no matter what you do is gonna take two or three months, right? Um and it just adds to uh the sales cycle. So you know on the business side and even you know the technology gatekeeping keeping side so I we usually get through it in a
34:22
Unidentified speaker (speaker_1): few weeks.
34:23
Unidentified speaker (UNKNOWN): Um
34:24
Unidentified speaker (speaker_1): because we've got a pretty compelling technical story. and and it's enterprise grade, you know, security, etc. They pass us, uh you know, they do the the checks, they pass us within days, because we've built this in the right way.
34:36
Unidentified speaker (UNKNOWN): Um
34:37
Unidentified speaker (speaker_1): best practices throughout. Um
34:39
Unidentified speaker (UNKNOWN): so
34:39
Unidentified speaker (speaker_1): those don't usually take too long unless someone's just gonna be difficult.
34:44
Unidentified speaker (UNKNOWN): In
34:44
Unidentified speaker (speaker_1): which case we kind of walk away pretty quickly. Um but yeah procurement and contracting always takes a while. But so
34:51
Unidentified speaker (speaker_1): you know
34:52
Unidentified speaker (UNKNOWN): I
34:52
Unidentified speaker (speaker_2): mean this is um
34:55
Unidentified speaker (speaker_2): this is a really important
34:58
Unidentified speaker (speaker_2): aspect of business that that is especially big companies because the bigger the company the more rigor m is applied to uh procurement decision or third or managing third party risk. So how do big companies get to a point of being able to say yes to a pilot or contract match
35:18
Unidentified speaker (speaker_2): faster and squeeze the the evaluation time from 90 days to maybe like 90 minutes? Is it like them continuously building like this view on a a particular third party, being able to run it through this? What what's your perspective? Because I feel like um you only score.
35:39
Unidentified speaker (speaker_2): if you if you take a shot and go.
35:41
Unidentified speaker (UNKNOWN): If you
35:42
Unidentified speaker (speaker_2): don't, um
35:44
Unidentified speaker (speaker_2): how do you know? I mean, I I I feel like you leave it a lot on the table.
35:48
Unidentified speaker (speaker_1): Yeah. So one of the things it it's generally is easier to get a proof of value, proof of concept, whatever you want to call it, through, right? Because it's a limited time, it's usually yeah pretty low cost, especially given the tools we're using, and and that's another aspect of all this, right? The business model is also changing under our feet because our costs are changing under our feet.
36:08
Unidentified speaker (speaker_1): So where do you pitch this? Um, but it gives me the luxury of having a conversation with a customer saying, Look, what are your levels here? You want to do this, I want to do it for you.
36:20
Unidentified speaker (UNKNOWN): So
36:21
Unidentified speaker (speaker_1): let's work together to figure out a deal that at least we can get some proof points. Because once you've got the proof points, everything else gets a lot easier, right? Um, and especially given that
36:31
Unidentified speaker (UNKNOWN): you know
36:31
Unidentified speaker (speaker_1): the beautiful thing here is people become champions for us really quickly. So, because they're the ones that are proving value from AI when no one else
36:40
Unidentified speaker (UNKNOWN): in our
36:41
Unidentified speaker (speaker_1): organization
36:41
Unidentified speaker (UNKNOWN): or
36:41
Unidentified speaker (speaker_1): the industry is doing
36:42
Unidentified speaker (UNKNOWN): that.
36:43
Unidentified speaker (speaker_1): That's great for their career, right? Great for their They're ego, everything else. So, you know, and and that is that's gold when you you're selling, right? Um, because you've got that internal champion. And it's not that they're you know giving you information that they shouldn't be giving you, they're just saying,
36:58
Unidentified speaker (speaker_2): yeah,
36:59
Unidentified speaker (speaker_1): you know, here's what I can make work on my side to get to get to that proof point.
37:03
Unidentified speaker (UNKNOWN): And
37:03
Unidentified speaker (speaker_1): then once we're at that proof point, it's not only me that's your champion, it's the board because they're gonna love this, right? They're gonna be turning around and going to procurement. What do you mean you do you
37:15
Unidentified speaker (UNKNOWN): need
37:15
Unidentified speaker (speaker_1): to put this out to tender and stuff? These guys are delivering value when no one else has. Get on with it.
37:22
Unidentified speaker (speaker_1): We're actually seeing that in a couple of cases, you know, where they're going, no, stop it. You know, you you're just
37:27
Unidentified speaker (UNKNOWN): putting
37:28
Unidentified speaker (speaker_1): friction in the way. We've we're seeing value from this.
37:33
Unidentified speaker (speaker_2): Mm-hmm. We
37:33
Unidentified speaker (speaker_1): need more of it. Let's get on with it, you know. And
37:36
Unidentified speaker (speaker_2): yeah.
37:36
Unidentified speaker (UNKNOWN): And
37:37
Unidentified speaker (speaker_1): again, that comes down to the kind of organization. organizations that you you want to work with, right? And and from a startup point of view, at least at this stage. It's related to Jeffrey Moore's kind of crossing the chasm, right? And uh and those kind of ideas, but it's more on the business side. It's not just the adoption of a product
37:57
Unidentified speaker (speaker_1): and the whole product solution and all that. It's is the business do they have the urgency? and the vision
38:05
Unidentified speaker (UNKNOWN): to
38:05
Unidentified speaker (speaker_1): actually go for this? Right. And and are they willing to open the door to let you prove it to
38:11
Unidentified speaker (UNKNOWN): them?
38:12
Unidentified speaker (speaker_1): That's constantly in my mind, right? When I'm, you know, when the team is qualifying an account, what questions are we asking and what kind of answers are we seeing that indicate that that's their pattern of behavior as opposed to oh, you know, we're gonna do a three-month security review and then we're gonna and the
38:31
Unidentified speaker (speaker_1): contract has to be on our paper and it has to be this and that, you know.
38:36
Unidentified speaker (UNKNOWN): Yeah,
38:37
Unidentified speaker (speaker_2): yeah,
38:37
Unidentified speaker (speaker_1): you guys have just built friction into your organization to the extent that you
38:42
Unidentified speaker (speaker_2): yeah,
38:42
Unidentified speaker (speaker_1): you're the opposite of of uh an innovative company, and and that's a big reason why I'm going back to why it's the that thirteen percent, right?
38:51
Unidentified speaker (UNKNOWN): Yeah.
38:52
Unidentified speaker (speaker_1): The others just can't even get their act together to even try it.
38:55
Unidentified speaker (UNKNOWN): They're
38:56
Unidentified speaker (speaker_1): too scared or too much friction, right?
38:58
Unidentified speaker (UNKNOWN): And
38:59
Unidentified speaker (speaker_1): that's it is what it is, and they're big companies for a reason, right? Um
39:03
Unidentified speaker (speaker_1): you know, whether they will stay big companies for too long with that attitude is another matter. But you know, we'll see.
39:10
Unidentified speaker (UNKNOWN): But
39:10
Unidentified speaker (speaker_2): it seems that it it seems like
39:12
Unidentified speaker (UNKNOWN): the
39:13
Unidentified speaker (speaker_2): bigger can com it at least I I uh this is my hypothesis. I th some of it is is supported by science, but the bigger the company, the more that growth comes from acquisition rather than organically. And and so um there's this
39:31
Unidentified speaker (speaker_2): inherent barrier uh in taking big big risks because they may not pay off and what the market rewards especially big companies typically they're publicly traded the market rewards consistency they don't want that volatility in in earnings and so that forces companies to behave in certain
39:51
Unidentified speaker (speaker_2): certain way where steady eddy wins the race uh rather than like I'm gonna take big shots and go away. Some of them are gonna going to pay off and some of them are not but the ones that pay off it's kind of like the power of distribution law where whereas they're going to pay off like a hundred X.
40:05
Unidentified speaker (speaker_1): But that's the problem, isn't it? You know, you get stuck on that quarterly. Oh my god, I've got to set the set the expectations, I've got to meet those expectations. And what a lot of these companies are missing then is well, next year or two years out, well, probably less than that now, because it's getting shorter and shorter. Someone else is taking advantage of this technology
40:25
Unidentified speaker (speaker_1): and their growth. rate is going to be dramatically faster than you.
40:29
Unidentified speaker (speaker_2): 100%.
40:31
Unidentified speaker (UNKNOWN): And
40:31
Unidentified speaker (speaker_1): you know, meeting your expectations and so on is no longer going to be good enough. Um, it it's just uh, like we said, but they're gonna record, right? It's a fascinating time to be in business, and yeah, especially on the startup side of it and on the AI deployment side of it. So much is changing. Every aspect of the business model that I'm deploying.
40:51
Unidentified speaker (speaker_1): is changing. Um and how far do we go with that in and where does it lead us? It's it's gonna be amazing. It's it's gonna be impressive
41:01
Unidentified speaker (UNKNOWN): years
41:01
Unidentified speaker (speaker_1): ahead of us in it. Yeah.
41:03
Unidentified speaker (UNKNOWN): Yeah
41:03
Unidentified speaker (speaker_2): I mean I um
41:07
Unidentified speaker (UNKNOWN): I
41:07
Unidentified speaker (speaker_2): love one what one of the big companies company I that really like in terms of what they're doing. I think it's really special is with Salesforce and I think that's that what Mark Benny Benioff is doing on this is is phenomenal.
41:18
Unidentified speaker (speaker_2): I mean obviously an amazing product. It's a genet focus. The acquisition of Slack in in in serving is is the layer in the communication layer. Like I I I love this, but the the one thing that I find really interesting is this kind of triple strategy of like, yeah we're gonna grow organically. Yeah we're gonna acquire so
41:38
Unidentified speaker (speaker_2): acquisitive growth and three. We're gonna invest like like maniacal uh corporate VC, we're gonna invest in companies, and and the ones that fit synergistically with what we're doing, what those become the acquisition targets. I just I just find this interesting, but I rarely see this in practice. Because I I obviously we deal with a lot
41:58
Unidentified speaker (speaker_2): of corporate VCs,
42:00
Unidentified speaker (UNKNOWN): but
42:00
Unidentified speaker (speaker_2): they're not the norm, and it's it is proliferate. proliferate. That's not to say that it's not corporate VC, there's a ton of them. It's
42:06
Unidentified speaker (UNKNOWN): just that
42:06
Unidentified speaker (speaker_2): it's not the norm.
42:08
Unidentified speaker (speaker_1): Well, and you know what I've found is that
42:11
Unidentified speaker (UNKNOWN): I've
42:12
Unidentified speaker (speaker_1): done a lot of corporate VC uh over the last 20 years. Um
42:18
Unidentified speaker (speaker_1): the trouble is, you know, they they tend to become a customer as well, which is great, right? In the early days of that corporate startup VC relationship, it can be awesome because it can get you into customers that you otherwise would really struggle to do.
42:31
Unidentified speaker (speaker_2): I agree.
42:31
Unidentified speaker (speaker_1): But then, you know, the the VC at that corporate, who's not necessarily a traditional VC, right? I was gonna say real VC, that's a bit unfair, but they're different, and they're driven by different incentives on their side. And once you've got your product in there and it's being used at scale.
42:51
Unidentified speaker (speaker_1): they tend to go quite risk averse on you
42:54
Unidentified speaker (UNKNOWN): because
42:54
Unidentified speaker (speaker_1): it's more important for them then to s to have you servicing them as a customer
43:00
Unidentified speaker (speaker_2): than
43:00
Unidentified speaker (speaker_1): it is growing in the market more generally, right?
43:03
Unidentified speaker (UNKNOWN): So
43:03
Unidentified speaker (speaker_1): they end up with a bit of a conflict of interest, in my my view, um, that can be quite difficult to work around, if especially if they're on your board,
43:11
Unidentified speaker (UNKNOWN): if
43:11
Unidentified speaker (speaker_1): they're on your board and dictating a your strategy
43:14
Unidentified speaker (speaker_2): in
43:14
Unidentified speaker (speaker_1): a way. that is really just oriented around them as a customer,
43:19
Unidentified speaker (speaker_1): you know, a market of one isn't going to work for
43:21
Unidentified speaker (UNKNOWN): you.
43:21
Unidentified speaker (speaker_1): Um
43:22
Unidentified speaker (UNKNOWN): and
43:22
Unidentified speaker (speaker_1): I've seen that over and over again. And it's
43:26
Unidentified speaker (UNKNOWN): you know,
43:27
Unidentified speaker (speaker_1): it can be the death of uh far too many startups, unfortunately. Um I'm saying that with the benefit of hindsight, because we did a lot of corporate copper VC at the incubator and fighting us quite a bit, but um
43:38
Unidentified speaker (UNKNOWN): yeah.
43:39
Unidentified speaker (speaker_1): you know, it can work, but it's it's challenging.
43:44
Unidentified speaker (UNKNOWN): Yeah. I
43:45
Unidentified speaker (speaker_2): want to come back to Geminos. Um
43:50
Unidentified speaker (speaker_2): you
43:51
Unidentified speaker (speaker_2): your approach to Geminis really it, I mean, we talked about it, it combines the LLMs, the agents, the cause of AI, the enterprise technology graphs, crafts, the business rule. Like,
44:02
Unidentified speaker (speaker_2): why? walk me through uh and walk the audience through as to why all those pieces need to work together. Why can't you just rely on LLM? I mean, it it if LM produces convincing results, even if they if they're even if the underlying reason it may be flawed.
44:21
Unidentified speaker (speaker_1): So it's about the scale, it's about the accuracy,
44:25
Unidentified speaker (UNKNOWN): about
44:26
Unidentified speaker (speaker_1): the consistency. of it, the traceability behind it, a lot of different factors.
44:30
Unidentified speaker (UNKNOWN): And
44:31
Unidentified speaker (speaker_1): it's not about the first application or two, it's about the tenth or the hundredth in an organization. So if you vibecode and use LLMs, and it's so if LLMs are centric to everything you do with agents and so on,
44:46
Unidentified speaker (speaker_1): you can't really do any of that that I've just listed, right? Uh consistently. You you can pull it off maybe with a pretty comprehensive harness around it, but you know, it it's hard and it's not something every, you know, certainly amateur developer, let's face it, a lot of people doing vibe coding in that way are not trained
45:06
Unidentified speaker (speaker_1): developers. Um
45:10
Unidentified speaker (speaker_1): let's just say it's the kind of thing that, you know, don't do this at home kind of thing. Um certainly for the enterprise at home, it's not mixing my to what's there. Anyway, um, but yeah, it's just not a solid platform for real widespread use. And what we're seeing is because we can show value on the first use case,
45:30
Unidentified speaker (speaker_1): we really quickly and then get demand from the rest of the business for more use cases. Well,
45:36
Unidentified speaker (UNKNOWN): if
45:36
Unidentified speaker (speaker_1): we're not getting the foundations right, that's really hard for us to deliver. Right. So I don't want I don't want to be in that position either of,
45:43
Unidentified speaker (UNKNOWN): you know,
45:44
Unidentified speaker (speaker_1): saying to the organization, yeah, yeah, we can do a hundred use cases for you and then failing miserably, right? It doesn't help anybody. So it doesn't help me. So we've put a huge amount of thought into getting those foundations right with the knowledge draft, with the rules based, base with using causal where relevant, and then using LLNs LLMs and agents around that.
46:04
Unidentified speaker (speaker_1): It's um,
46:06
Unidentified speaker (UNKNOWN): you know,
46:06
Unidentified speaker (speaker_1): it to use some technical terms here, it's a form of neurosymbolic AI.
46:11
Unidentified speaker (UNKNOWN): So
46:11
Unidentified speaker (speaker_1): it's the combining the symbolic AI, the causal and the rules engines and so on, with the neural LLMs,
46:19
Unidentified speaker (UNKNOWN): uh,
46:19
Unidentified speaker (speaker_1): which I think is the ultimately the right way forward. LLMs alone,
46:24
Unidentified speaker (UNKNOWN): you
46:24
Unidentified speaker (speaker_1): know, they're not gonna get us to AGI in any way. meaningful way. Um, amazing technology that can do an awful lot.
46:33
Unidentified speaker (speaker_2): There's
46:33
Unidentified speaker (speaker_1): an awful lot more to AI than LLMs, and we we need to relearn some of these things and bring them to bear to solve problems, you know, and and to meet customers' needs. Uh
46:45
Unidentified speaker (speaker_1): LLMs LMs alone don't do
46:46
Unidentified speaker (UNKNOWN): that.
46:47
Unidentified speaker (speaker_3): Completely agree with
46:49
Unidentified speaker (speaker_2): you. Um so with regards to this, um
46:54
Unidentified speaker (UNKNOWN): I I
46:54
Unidentified speaker (speaker_2): mean ultimately what we're trying to uh increase is the level of trust, the level of certainty, the level of precision. So if if I'm a CPA of a background, so if we're to take this approach from a governance and assurance perspective, in financial reporting, we don't promise absolute certainty. Like um uh uh audit
47:14
Unidentified speaker (speaker_2): opinion provides reasonable assurance, right? It doesn't provide absolute assurance. So we design controls, we collect evidence, and to provide this reasonable assurance that the information is provided, it that the information is reliable.
47:28
Unidentified speaker (speaker_2): Should we be thinking about AI systems in a similar way, or do we have a sense of expectation? Do we have a an expectation that we could get?
47:39
Unidentified speaker (speaker_2): to an absolute certainty because it's reading data. Now, if the data is structured versus unstructured versus neural, like kinda like w what is your thing?
47:48
Unidentified speaker (UNKNOWN): Yeah,
47:48
Unidentified speaker (speaker_1): this is a really interesting area, isn't it? Because when I sit in front of a computer and I use a spreadsheet,
47:54
Unidentified speaker (UNKNOWN): as
47:54
Unidentified speaker (speaker_1): long as there are macros that I actually understand or myself, etcetera, um I'm relying on that to be a pretty deterministic system, right?
48:02
Unidentified speaker (speaker_2): Yeah.
48:02
Unidentified speaker (UNKNOWN): So
48:03
Unidentified speaker (speaker_1): yeah, as long as the code is valid and etc. It's gonna give me the result.
48:09
Unidentified speaker (speaker_2): Yeah, I'm
48:10
Unidentified speaker (speaker_1): gonna get a consistent result. That's simply as you know, not the case with LMs, right? They'll give you a tight different result every time almost. It's like you know, um, depending on on the path through the matrices. So it it's a fundamentally different beast. Um you know, it's probably probabilistic, not deterministic. Now, in
48:30
Unidentified speaker (speaker_1): most cases, that's actually okay because the world is probabilistic and and not deterministic in most things, right? So when we're making decisions,
48:42
Unidentified speaker (speaker_1): expecting perfection isn't gonna happen, right? Let's just not let's be realistic about this,
48:49
Unidentified speaker (UNKNOWN): and
48:49
Unidentified speaker (speaker_1): the best human. decision maker does not get it right every
48:54
Unidentified speaker (UNKNOWN): time.
48:54
Unidentified speaker (speaker_2): Now
48:55
Unidentified speaker (speaker_1): that's not an excuse for saying, oh, okay, then we can, you know, just use LLMs and it's fine. It's not because they they get get it wrong in unpredictable ways and they can get it wrong and be very confident, right? And also one step in a long agent thread that gets it wrong means it's entirely wrong. Uh if you go down and and go down some very
49:15
Unidentified speaker (speaker_1): bad paths. So,
49:16
Unidentified speaker (UNKNOWN): you
49:16
Unidentified speaker (speaker_1): know, so what we're we're doing here is getting the balance right. We're putting the framework of rules and causal AI, etc., around the LLM, not using the LM for the primary decision, to get to a point where we can help human operators make better decisions overall across a team
49:36
Unidentified speaker (UNKNOWN): on
49:36
Unidentified speaker (speaker_1): a day-to-day basis consistently, and we can tell them when we're unsure and give them a pretty good idea of okay, there's a bit of data missing here, or there are two causes here, you know, in the causal model that you know, one it might be 60, 70 percent for one and 30 percent for the other.
49:54
Unidentified speaker (UNKNOWN): You
49:55
Unidentified speaker (speaker_1): gotta think about that, right?
49:57
Unidentified speaker (speaker_0): Yeah,
49:57
Unidentified speaker (speaker_1): um, and training people and getting to understand that and giving them the information and they've the ability to ask questions using natural language, which is great, right? So LMs are plugged in with agents and in their workflow, they've got a
50:11
Unidentified speaker (UNKNOWN): you
50:11
Unidentified speaker (speaker_1): know some information in front of them which says hmm 7030.
50:16
Unidentified speaker (speaker_1): Why? Let me ask about that. Natural language questions about it. I'm not you're not just stuck with whatever information is uh in front of
50:25
Unidentified speaker (UNKNOWN): you.
50:26
Unidentified speaker (speaker_1): Um And users are finding that really powerful. They can query it, they can understand the probabilities better, where there's uncertainty, you know, um, and and where there's missing information that they need to go fill in. Um
50:40
Unidentified speaker (UNKNOWN): so
50:40
Unidentified speaker (speaker_1): that's really you know what this is all about, right? It's about human in the loop, still the human making the decision, but giving them all the information that they need in order to. make that decision much faster and on the whole more accurately than they've been capable of doing before.
50:56
Unidentified speaker (UNKNOWN): And
50:56
Unidentified speaker (speaker_1): if we achieve that, then we've got something of real value. But we're not pretending it's perfect every time.
51:02
Unidentified speaker (speaker_2): So, Stuart, speaking of making the making the decisions, so
51:09
Unidentified speaker (UNKNOWN): and
51:10
Unidentified speaker (speaker_2): you're you talking talk about use uh having a four-deployed forward deployed engineering model that's there to support the client. So I'm assuming that when you go and start working with clients, you have a multitude of opportunities that you're looking at and deciding upon. So, how do you decide or how do you help the company decide
51:30
Unidentified speaker (speaker_2): which opportunity is merely interesting from the one that's going to drive the most value? And how do companies or how do you help companies define what is truly value?
51:41
Unidentified speaker (speaker_1): Hmm.
51:42
Unidentified speaker (UNKNOWN): So
51:42
Unidentified speaker (speaker_1): it's obviously a fascinating conversation that we uh, you know, and love to get into every time. And it is different almost each time. And it's not necessarily about driving the most value in the first use case.
51:56
Unidentified speaker (UNKNOWN): What
51:56
Unidentified speaker (speaker_1): we're looking to do is choose one mutually that gets us to value quickly.
52:02
Unidentified speaker (UNKNOWN): Okay.
52:03
Unidentified speaker (speaker_1): Right? Because that is that's the initial goal.
52:07
Unidentified speaker (UNKNOWN): Keep it
52:08
Unidentified speaker (speaker_0): in,
52:08
Unidentified speaker (speaker_1): show value fast, a sensible cost, right?
52:12
Unidentified speaker (speaker_2): Yeah.
52:12
Unidentified speaker (speaker_1): Everything changes in terms of the sales cycle. And then we can tackle some of the more you know high value uh projects. I I actually just did a blog post on this. I'm not sure if it's on the website yet.
52:23
Unidentified speaker (speaker_2): Okay.
52:23
Unidentified speaker (speaker_1): Jeffrey Moore has a uh he talks about zones. for a company, and it's things like you know the productivity zone, the innovation zone, etc. And we use that as a bit of a framework for thinking about different use cases. Um and in general, I would say what we're looking for initially is things in the productivity zone. Things where
52:43
Unidentified speaker (speaker_1): a big company makes a decision very consistently every day, and moving the needle on that decision is going to drive some value quite quickly, right?
52:53
Unidentified speaker (speaker_2): Yeah.
52:54
Unidentified speaker (speaker_1): And often it's it's a nice sweet spot for us where there's a lot of text involved with some structured data. Because that combination has been traditionally really hard to automate and use a software application to solve, right?
53:09
Unidentified speaker (UNKNOWN): Uh
53:09
Unidentified speaker (speaker_1): at any kind of scale.
53:11
Unidentified speaker (UNKNOWN): Um
53:12
Unidentified speaker (speaker_1): so the fact that we can do that now. So the the Great Western Railway example of delay attribution for trains, right?
53:20
Unidentified speaker (UNKNOWN): A
53:20
Unidentified speaker (speaker_1): lot of text, Slack messages,
53:23
Unidentified speaker (speaker_1): emails, phone call transcripts,
53:26
Unidentified speaker (UNKNOWN): a lot
53:27
Unidentified speaker (speaker_1): of rules about what happens. You have to understand the train network, the railway network, uh understand what's happening with drivers, etc. et cetera. Um,
53:37
Unidentified speaker (UNKNOWN): You know,
53:37
Unidentified speaker (speaker_1): it's 15 different sources of data. Um so an individual trying to figure out then whose fault that delay was and who should pay for it ultimately,
53:46
Unidentified speaker (UNKNOWN): it was
53:47
Unidentified speaker (speaker_1): taking them eight to ten days. Um
53:50
Unidentified speaker (UNKNOWN): we
53:50
Unidentified speaker (speaker_1): got that down to a few minutes and and it was you know a couple of months of uh development work for us to get it to production level. So that that's a dramatic impact.
54:01
Unidentified speaker (UNKNOWN): Is
54:01
Unidentified speaker (speaker_1): it their highest value use? case? Not necessarily, but it's a nice one that generates, you know, tens of millions a year. Um and and makes the job of the person doing delay attribution much nicer, frankly. They're not dealing with a lot of tedium of looking through lots of data and trying to piece it together. They're presented with
54:21
Unidentified speaker (speaker_1): it's 65% this one, 20% this one, 10% this one. Here's something else to, you know, a missing piece of data that will help us move that needle if you want,
54:32
Unidentified speaker (UNKNOWN): uh
54:32
Unidentified speaker (speaker_1): make it more certain. Go call Bill in, you know, the station at Reading or whatever and get this bit of information for us and we'll add that in and and redo it. Or it's ninety percent and you know, you're pretty confident in it. It's what here's why. And here's a complete report to send to the other parties. party that needs to pay for this, you
54:52
Unidentified speaker (speaker_1): know, and give them all the the reasons why. Um
54:56
Unidentified speaker (UNKNOWN): I
54:56
Unidentified speaker (speaker_2): mean what's interesting about the the example that you're providing Stewart is I mean obviously going from eight days to eight minutes it's it's it's a massive gain of like I I don't know I haven't calculated but it's probably Maybe a hundred 100x X. And so
55:12
Unidentified speaker (speaker_2): but what's neat about this is that if it's just the start of things. So you start with okay, well, let's go do some root cause analysis around delay attribution. And so for each one, we have specificity of what caused the delay. But then you can take this information and pipe it into like okay, well, okay, if if if if
55:32
Unidentified speaker (speaker_2): whatever the root With cause cost was for the delay, let's go and fix it, fix it, fix it,
55:36
Unidentified speaker (UNKNOWN): and It just
55:36
Unidentified speaker (speaker_2): kind of start this like virtuous, yeah. exactly.
55:39
Unidentified speaker (UNKNOWN): Yeah.
55:40
Unidentified speaker (speaker_1): Well, and the other one that we're talking to uh another rail company about is on the announcement side.
55:46
Unidentified speaker (UNKNOWN): Okay. So
55:47
Unidentified speaker (speaker_1): a train's delayed, right? And they've got a bunch of stations. And and instead of you know the announcer just saying, sorry, trains delayed, what if
55:56
Unidentified speaker (UNKNOWN): we
55:57
Unidentified speaker (speaker_1): could actually say, if you want to go to London, go to platform three, take the train. to Sheffield or whatever, and you know it'll take you slightly longer, but you'll get there roughly this time.
56:11
Unidentified speaker (speaker_2): Yeah.
56:11
Unidentified speaker (speaker_1): You know, and so we're actually going forwards now and and saying, Okay, great, if there's a delay, what's the best path around that? Not just for the rail company, but for passengers. You know, what should they
56:23
Unidentified speaker (UNKNOWN): do? Uh
56:24
Unidentified speaker (speaker_1): rather than you know, k y 'cause because you know what happens. now, right? They go to the whoever's standing on the station and they go, Yeah,
56:30
Unidentified speaker (speaker_2): how
56:30
Unidentified speaker (UNKNOWN): do
56:30
Unidentified speaker (speaker_1): I get to London from here if that train's delayed three hours? What's the best way?
56:35
Unidentified speaker (UNKNOWN): And
56:35
Unidentified speaker (speaker_1): go, uh, let's look at the timetable, let's do uh yeah.
56:38
Unidentified speaker (speaker_2): Yeah.
56:39
Unidentified speaker (speaker_1): So it absolutely spreads, right? It's um it's a viral effect across the business. But it's also a viral effect across the industry. Um
56:49
Unidentified speaker (speaker_1): because of that champion. side of it, right? So, you know, the guy at Great Western Railways has been a fantastic champion reference for us. He's very well known in the industry now as one of the leading thinkers on AI. Um and he's spreading that word to all the other rail companies. So great for us, right? Yeah, land and expand in multiple
57:09
Unidentified speaker (speaker_1): dimensions. Awesome. also
57:10
Unidentified speaker (speaker_2): A 100%. hundred percent. Um I know we're coming at time. Uh just one last question for me. What is one widely held belief about AI and entrepreneurship that you think is fundamentally wrong?
57:24
Unidentified speaker (speaker_1): Boy, that's a that's a thoughtful one, isn't
57:27
Unidentified speaker (UNKNOWN): it?
57:28
Unidentified speaker (speaker_1): Um
57:29
Unidentified speaker (speaker_1): well, I mean, beyond just the the normal you know, LLMs are everything, which I think an awful lot of people are falling for, um, and you know, we're seeing del Delaware now allow you to have agent companies, AI companies, with no people responsible for them understand, which is crazy, right? So that kind of
57:49
Unidentified speaker (speaker_1): extreme for me, uh and you know, exemplified by things like Block, uh Jack
57:55
Unidentified speaker (speaker_1): Company, where okay, the his company roadmap is now completely delegated to agents, as I understand it. So a customer says they want a feature, the agent goes and builds it and puts into the product. That to me is the antithesis of of what a startup should be about, right? And it it's like yeah, playing Christian uh
58:15
Unidentified speaker (speaker_1): sustaining engineering versus uh innovation, you know, disruptive engineering. Let's just do sustaining engineering. How about no? Right? So when we're talking about that kind of level of thing. things, you know, it's
58:27
Unidentified speaker (UNKNOWN): like,
58:27
Unidentified speaker (speaker_1): no, we don't do that. It
58:29
Unidentified speaker (UNKNOWN): takes just
58:30
Unidentified speaker (speaker_2): because you can doesn't mean it should.
58:32
Unidentified speaker (speaker_1): Yeah, exactly, right? And I think that's that's true of AI right across the board, right? That just because we can do it, uh, no. And I think ultimately that goes to
58:44
Unidentified speaker (UNKNOWN): you
58:44
Unidentified speaker (speaker_1): know AI consciousness and
58:47
Unidentified speaker (speaker_1): AGI and everything else to that anthropic and OpenAI. racing to right to justify their trillions in value.
58:56
Unidentified speaker (speaker_1): Why do we need that?
58:59
Unidentified speaker (speaker_1): Why is that necessary to get value from this technology?
59:03
Unidentified speaker (UNKNOWN): You
59:03
Unidentified speaker (speaker_1): know, and given all the dangers of it, why the hell are we allowing them to do
59:08
Unidentified speaker (UNKNOWN): that?
59:09
Unidentified speaker (speaker_2): Mm-hmm.
59:09
Unidentified speaker (speaker_1): You know?
59:10
Unidentified speaker (UNKNOWN): Okay.
59:10
Unidentified speaker (speaker_1): So maybe that's not quite on point with
59:13
Unidentified speaker (UNKNOWN): the
59:13
Unidentified speaker (speaker_1): question, but that's what what a are the things that are top of mind for me right now.
59:17
Unidentified speaker (UNKNOWN): No, I
59:17
Unidentified speaker (speaker_2): really appreciate you. This has been a fascinating conversation. Um thank you so much for joining me on Scaling with Clarity and sharing your experience and perspective. It clearly demonstrated you have a ton of knowledge and experience and and truly can help companies in in interrupt arrive at better decisions faster in a more
59:36
Unidentified speaker (speaker_2): deterministic way rather than what LLM can produce. So for everyone listening can learn more about Stewart and his work at Geminis.ai we'll drop links to to Stuart and where to find him. Um until next time keep building keep learning and keep scaling with clarity. Thanks Stuart. Thanks
59:55
Unidentified speaker (speaker_1): Mike Sim. It's been a real pleasure.
59:57
Unidentified speaker (speaker_2): Yeah
59:57
Unidentified speaker (speaker_1): I've enjoyed
59:58
Unidentified speaker (UNKNOWN): it.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights