IBM CEO Krishna outlines enterprise AI adoption stages, open-weight model economics, and software consolidation trends.
Summary
IBM CEO Arvind Krishna discusses the state of enterprise AI, noting that only 10-20% of companies have moved from experimentation to scaled deployment. He identifies coding, customer experience, and enterprise operations as key use cases, predicting that tech spend will double as a percentage of revenue over the next decade. Krishna argues that data accessibility matters more than cleanliness for AI agents and highlights IBM's internal success with AI automating HR and procurement processes. He forecasts that open-weight models will capture 80% of enterprise token usage within three years due to cost and control advantages, drawing parallels to the adoption of Linux.
Key takeaways
Episode sections
- 0:18 – 1:56
Introduction and IBM's Software Pivot
Host Anurag Rana introduces IBM CEO Arvind Krishna, noting IBM's transformation since the $34 billion Red Hat acquisition in 2019 and subsequent deals like Apteo, HashiCorp, and Confluent. Krishna is welcomed for his role in leading IBM's shift toward software, which has improved organic sales, margins, and cash flow.
- 1:56 – 5:24
Enterprise AI Adoption and Use Cases
Krishna describes enterprise AI adoption as being in the 'second innings,' with 10-20% of enterprises moving from experimentation to scale. He identifies three 'no-regrets' use cases: coding/programming, customer experience (where 60-80% of clients prefer AI interaction), and enterprise operations like procurement and HR.
- 5:24 – 9:18
Unlocking Data and the Role of Agents
Krishna explains that unlocking data across silos is critical for AI agents to function, citing a complex sales proposal example involving multiple departments. He argues that data does not need to be clean, only accessible, and suggests using streaming data (via Confluent) to unlock value without moving it to a central repository.
- 9:18 – 16:29
System of Record Agents and IP Protection
Krishna discusses the tension between system-of-record vendors (like SAP) exposing APIs versus keeping data locked. He asserts that sophisticated enterprises will avoid letting third-party models access core IP to prevent implicit leakage of methodology, preferring to run their own agents on-premise or in private clouds.
- 16:29 – 18:51
Enterprise Spending Predictions
Krishna predicts that enterprise technology spending, currently 5-6% of revenue, will double as a percentage of revenue within five to ten years. He clarifies that while token usage is a component, the majority of this spend will be on code, applications, data access, and workflow modernization, not just inference tokens.
- 18:51 – 27:50
Software Disruption and OpenShift Tailwinds
Krishna categorizes software by disruption risk: interactive UIs are vulnerable, while deterministic, high-volume systems (payments, core databases) are safe. He notes IBM's software portfolio is largely in the non-disruptive bucket. He highlights OpenShift as a key tailwind, stating the industry is less than 10% into the containerization journey.
- 27:50 – 32:57
Services Business Evolution and Headcount
Krishna addresses why services headcount hasn't dropped significantly, citing tailwinds from data preparation and security needs against headwinds from AI-driven efficiency in coding and documentation. He predicts a 50-60% improvement in elapsed time for tasks like SAP implementation and expects IBM's services headcount to decrease over the next few years.
- 32:57 – 34:41
Database and Services Consolidation
Speaker 1 states that most database vendors have disappeared, leaving only Oracle, Microsoft, and IBM as majors, with others replaced by key-value or JSON stores. In the services sector, consolidation is expected among large players due to high AI investment costs, while a new expansion of domain-specific experts will emerge due to lower barriers to entry via AI tools.
- 34:41 – 38:20
Internal AI Agent Implementation at IBM
Speaker 1 describes IBM's internal use of AI agents to automate HR processes, noting that 94% of HR activities by volume are now handled by a single agent called 'Ask HR'. The agent accesses multiple systems (payroll, directory) with strict guardrails to protect PII, reducing the need for human intervention in routine tasks.
- 38:20 – 40:35
Procurement and IT Support Automation
IBM is applying similar agents to procurement to verify statements of work and pricing, and to IT support to provide expert-level assistance beyond traditional help desks. Approximately 60 enterprise processes are in good shape with these agents, with another 60 making progress, reducing the need for large teams of rote workers.
- 40:35 – 42:19
Prediction on Open Weight Model Adoption
Speaker 1 predicts that within three years, 80% of enterprise token usage will be on open weight models or models chosen by the enterprise to run on-premise or in the cloud. This shift is driven by the closing gap in quality between open weight and frontier models, with open weight models costing one-fifth to one-tenth as much while being 96% as good for many tasks.
- 42:19 – 44:08
Barriers to Adoption and Future Ease of Use
Currently, average enterprises lack the engineering expertise and guardrails to deploy open weight models effectively. However, Speaker 1 expects these barriers to be resolved within a year through easier-to-use systems, similar to the evolution of Kubernetes and Linux, allowing enterprises to leverage open weight models with minimal infrastructure complexity.
- 44:08 – 45:31
Vertical-Specific Model Training
The discussion covers the potential for auto companies to fine-tune open weight models using specific data like warranty claims. A 10% improvement in handling warranty claims could save a company billions annually, providing a strong economic incentive to adopt and customize open weight models rather than relying solely on frontier models.
- 45:31 – 48:27
Linux Analogy for Enterprise AI
Speaker 1 compares the shift to open weight models to the adoption of Linux, which reached escape velocity around 2000 and now runs 70% of enterprise workloads, projected to reach 90% by the end of the decade. This analogy suggests that open weight models will eventually dominate enterprise AI workloads due to cost and control advantages.
Assets discussed
| Asset | Kind | Stance | Context |
|---|---|---|---|
| MicrosoftMSFT | equity | neutral | We are left with three majors Oracle, Microsoft, IBM. |
| NVIDIANVDA | equity | neutral | NVIDIA buying hugging face and more open stuff coming through, that may change is |
“Stance” is how the speakers talked about the asset, not a recommendation.
Claims worth verifying
IBM acquired Red Hat for $34 billion in 2019.
Unidentified speaker (speaker_0) ·1:11
How to verify: IBM 2019 Form 10-K filing with SEC EDGAR; press releases from 2019. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Primary sourceIBM has undertaken software deals including Apteo, HashiCorp, and Confluent.
Unidentified speaker (speaker_0) ·1:11
How to verify: IBM investor relations newsroom; SEC filings for material acquisitions. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Primary source10-20% of enterprises are moving from experimentation to scaled AI.
Unidentified speaker (speaker_1) ·2:16
How to verify: Internal IBM surveys or third-party market research reports (e.g., Gartner, IDC) on AI adoption rates. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
60-80% of clients prefer interaction with AI over humans in customer experience.
Unidentified speaker (speaker_1) ·3:19
How to verify: IBM internal customer satisfaction surveys or case studies published by IBM. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Enterprise technology spending is currently 5-6% of revenue.
Unidentified speaker (speaker_1) ·16:53
How to verify: Analysis of aggregate financial data from top 10,000 enterprises; industry reports on IT spend as % of revenue. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Enterprise tech spend will double as a percentage of revenue in 5-10 years.
Unidentified speaker (speaker_1) ·17:13
How to verify: Longitudinal tracking of enterprise IT spend ratios; verification against future financial reports. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
IBM software unit accounts for 46% of total sales, up from 32% when Krishna took over.
Unidentified speaker (speaker_0) ·18:51
How to verify: IBM quarterly earnings reports (10-Q) detailing segment revenue; historical data from 2026. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Primary sourceIBM software was 22% of IBM in 2018.
Unidentified speaker (speaker_1) ·19:39
How to verify: IBM 2026 Annual Report (10-K) segment reporting. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Primary sourceOpenShift adoption is less than 10% of the journey to becoming the default standard.
Unidentified speaker (speaker_1) ·23:28
How to verify: Market share data for container orchestration platforms (Kubernetes/OpenShift) vs. total enterprise applications. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
AI tools can improve elapsed time for SAP implementation documentation by 50% or more.
Unidentified speaker (speaker_1) ·29:25
How to verify: IBM internal efficiency metrics for services projects; case studies comparing traditional vs. AI-assisted implementation times. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Most database vendors have disappeared, leaving only Oracle, Microsoft, and IBM as majors.
Unidentified speaker (speaker_1) ·33:02
How to verify: Check recent market share reports for enterprise database software from firms like Gartner or IDC; review IBM, Oracle, and Microsoft earnings calls for mentions of database market consolidation. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Primary source94% of all HR activities by volume at IBM are done using a single agent.
Unidentified speaker (speaker_1) ·36:32
How to verify: Review IBM's quarterly earnings reports or investor presentations for internal efficiency metrics; look for press releases detailing the 'Ask HR' agent implementation. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Open weight models cost one-fifth to one-tenth of frontier models and are 96% as good.
Unidentified speaker (speaker_1) ·41:45
How to verify: Compare pricing APIs for open weight models (e.g., via Hugging Face Inference API) against frontier model providers (e.g., OpenAI, Anthropic); review independent benchmarks like LMSYS Chatbot Arena for quality comparisons. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Within three years, 80% of enterprise token usage will be on open weight or enterprise-chosen models.
Unidentified speaker (speaker_1) ·40:59
How to verify: Track enterprise AI adoption surveys from firms like IDC, Gartner, or McKinsey over the next three years; monitor token usage data from major cloud providers and AI model marketplaces. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Linux runs 70% of all enterprise workloads.
Unidentified speaker (speaker_1) ·45:51
How to verify: Check server operating system market share reports from Synergy Research Group or Statista. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Linux will run 90% of enterprise workloads before the decade is out.
Unidentified speaker (speaker_1) ·45:51
How to verify: Monitor server OS market share trends annually via Synergy Research Group or Statista to verify the 90% projection by 2026. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
NVIDIA bought Hugging Face.
Unidentified speaker (speaker_0) ·44:08
How to verify: Search SEC filings (Form 8-K) for NVIDIA Corporation or press releases from NVIDIA and Hugging Face regarding acquisition transactions. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Primary sourceSalesforce released a sector-specific model.
Unidentified speaker (speaker_0) ·44:08
How to verify: Check Salesforce press releases or blog posts for announcements regarding new AI models, specifically those tailored to verticals. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
IBM has 60 enterprise processes in good shape with AI agents and another 60 making progress.
Unidentified speaker (speaker_1) ·38:43
How to verify: Review IBM's internal case studies or earnings call transcripts for specific metrics on AI process automation. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
AI could help keep GDP growth north of 3%.
Unidentified speaker (speaker_1) ·47:30
How to verify: Compare historical GDP growth rates with periods of significant technological adoption; review economic forecasts from the Federal Reserve or IMF regarding AI's macroeconomic impact. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.
Do your own research
- Verify IBM's Red Hat acquisition price ($34 billion) and subsequent deals (Apteo, HashiCorp, Confluent) via SEC filings.
- Check IBM's latest quarterly earnings report for organic sales growth, margin improvements, and cash flow metrics.
- Research the current market share of Oracle, Microsoft, and IBM in the enterprise database sector.
- Review IBM's internal case studies or press releases regarding the 'Ask HR' agent's implementation and PII guardrails.
- Compare the performance and pricing of leading open-weight models (e.g., Llama 3, Mistral) against frontier models (e.g., GPT-4, Claude 3) using independent benchmarks.
- Analyze IBM's services headcount trends in recent financial reports to assess the impact of AI-driven efficiency.
- Investigate the adoption rates of containerization (OpenShift) and streaming data platforms (Confluent) in the enterprise sector.
Full transcriptexpand
- 0:18
- Unidentified speaker (UNKNOWN): Hi
- 0:18
- Unidentified speaker (speaker_0): everyone, welcome to another episode of the Tech Disruptors Podcast, brought to you by Bloomberg Intelligence. Bloomberg Intelligence is the investment. research platform on the Bloomberg Terminal. We have 500 analysts and strategists working across the globe, and we are focused on all major world markets. Our coverage includes over 2,000 equities and
- 0:38
- Unidentified speaker (speaker_0): credits, and we have outlooks on more than 90 industries and 100 market indices, currencies, and commodities. I'm Anuragrana, a technology analyst at Bloomberg Intelligence, and I'll be your host for today's episode.
- 0:51
- Unidentified speaker (speaker_0): It is our absolute delight to welcome the chairman, president, and the CEO of IBM, Irvin Krishna. With over three decades at IBM, Irving has played a pivotal role in shaping the company's journey and is one of the world's leading technology visionaries and thought leaders. IBM has undergone many transformations in its history, and Arwin is leading its next big
- 1:11
- Unidentified speaker (speaker_0): shift, which started with the $34 billion acquisition of Red Hat. in 2019. Since then, IBM has undertaken several other software deals, including Apteo, HashiCorp, and more recently, Confluent. The pivot to software has helped IBM's financial profile in multiple ways, from improving organic sales growth,
- 1:31
- Unidentified speaker (speaker_0): its margin profile, and cash flow. On today's podcast, we will dive into multiple issues from enterprise AI to our Urban's long-term vision for its various businesses. Arvind, welcome to the podcast.
- 1:46
- Unidentified speaker (speaker_0): Thank you. So, Arvind, to start off with, where are enterprises in terms of AI adoption right now, and what are some of the biggest barriers to achieving ROI?
- 1:56
- Unidentified speaker (speaker_1): Yeah. Interesting question, Anurag. Um, I used to say that AI adoption to use a base baseball analogy in the first innings. I'd actually say it's maybe more like the second innings right now. So the players are on the field, people have kind of warmed up. And I think there's about 10 to 20 percent
- 2:16
- Unidentified speaker (speaker_1): of enterprises who are moving from experimentation to scaled AI. There's three use cases that I think are no regrets, meaning there's very little pushback against.
- 2:27
- Unidentified speaker (speaker_1): it. Doesn't quite mean that people have their arms around how to do it, but they agree they will do it. First, which I think the world talks about, is in programming or coding. I think large enterprises have embraced it in pockets, if not across the whole enterprise.
- 2:45
- Unidentified speaker (UNKNOWN): I
- 2:45
- Unidentified speaker (speaker_1): think medium-sized enterprises are getting there. And the risk there is well understood. and well controlled, but the productivity gains so far outweigh the risks that people are jumping all in.
- 2:59
- Unidentified speaker (UNKNOWN): The
- 2:59
- Unidentified speaker (speaker_1): second use case is around customer experience. It's not about cost cutting only, and it's not about productivity only. If you can use AI to augment your customer experience, it's is 24 by 7. AI agents don't get angry and don't get upset. They don't really react to the tone of the
- 3:19
- Unidentified speaker (speaker_1): other side. And people who are embracing this find that at least 60 to 80% of the client set much prefers the interaction with AI as opposed to uh, let's call it with carbon. Uh, I think that that's a no-regrets. no regrets. And I've heard people talk about
- 3:39
- Unidentified speaker (speaker_1): maybe over time we can simplify the customer experience dramatically, improve uh sort of uh r resolution on the first touch, as well as um uh have everybody more delighted at the end of it. And I would say that that is playing out completely.
- 3:58
- Unidentified speaker (UNKNOWN): The
- 3:58
- Unidentified speaker (speaker_1): third one, which is not quite as mature as the first two, is trying to go across all enterprise operations: procurement, compliance, HR. And as you think about that, every enterprise has a hundred to two hundred of these processes, and you can go across it.
- 4:16
- Unidentified speaker (UNKNOWN): So
- 4:17
- Unidentified speaker (speaker_1): if I give you sort of my brush across it, coding is getting very embraced. I think customer experience is getting quite embraced. There are some things there that are tricky that people worry about. Are all our systems got clean data? What about things like HIPAA compliance and uh uh too
- 4:36
- Unidentified speaker (speaker_1): much PII, et cetera? So those concerns have to get resolved. So that is a bit of an inhibitor.
- 4:43
- Unidentified speaker (UNKNOWN): And
- 4:43
- Unidentified speaker (speaker_1): the third one around enterprise ops, the big question is the data and the systems. Nobody's going to be able to put all the data in one place. So how do you unlock the data when you have to cut across 10 or 15 different systems? But as opposed to doing it using human glue, which is people looking at a lot of screens, I think AI agents are going to be the
- 5:03
- Unidentified speaker (speaker_1): de facto answer here. If you note, I didn't touch on the revenue side of the equation. That's where I think there's a lot of opportunity still in front of us. So while I mentioned those three, and they are pretty big, um, it is typically 20% of uh a typical client's total spend. There's the other side on revenue that
- 5:23
- Unidentified speaker (speaker_1): we've yet to get to.
- 5:24
- Unidentified speaker (UNKNOWN): So,
- 5:24
- Unidentified speaker (speaker_0): Ivan, perhaps we can dive into the third aspects, which is the AI diffusion across the enterprise. And I think during the third quarter call, you talked about how 90% of the enterprise data still lies in companies' company's own databases. So, what does a company need to do to unlock that data so that you get true value out of it?
- 5:44
- Unidentified speaker (speaker_1): Yeah. So, Arurak, let me maybe give an example from inside IBM, and that'll sort of uh make this very, very real. Let's take a complex sales proposal that we give to one of our clients. So there is a sales team in front of the client who gives it to them. They also obviously work worked with the solutioning
- 6:04
- Unidentified speaker (speaker_1): team who created that. Then you have to get approval from the pricing team, which is typically living inside finance and needs to make sure that it's a fair price. And how do you
- 6:14
- Unidentified speaker (speaker_1): ascribe revenue to all the component pieces? You're going to have somebody from Legal who's looking at what's the risk because there is always liabilities associated with complex project deliveries.
- 6:25
- Unidentified speaker (UNKNOWN): Then
- 6:25
- Unidentified speaker (speaker_1): there's going to be a delivery team.
- 6:27
- Unidentified speaker (UNKNOWN): If
- 6:27
- Unidentified speaker (speaker_1): there's hardware involved, there's got to be inventory, manufacturing. So you can begin to see that what I began with a simple thing touches lots and lots of different systems. So if I want to change that instead of going across seven different silos, which normally then takes the two weeks of approval, I want to have agents go do this.
- 6:47
- Unidentified speaker (speaker_1): You've got to unlock the data in each of them. So if it's uh the legal compliance where there is a repository of contracts, and then you have to ask the question: is there a repository of contracts? Because your AI can only learn what is the risk from looking at past contracts and what happened to those. Almost nobody has that data today.
- 7:07
- Unidentified speaker (speaker_1): So that is where the unlock has to come by saying, okay, let's look at these. And for if it has enough value, how do I get enough data? to be able to train the AI? And how do I get enough data to be able to make a real-time decision? I think that this is the question that is holding back the full adoption of enterprise
- 7:27
- Unidentified speaker (speaker_1): AI in that third bucket. But the price is so big that I think this is just going to happen. It takes though probably three to five years for us to get to: is it 50% fifty percent done? done? 50% Fifty is percent all is always the magic number I look at. Are half the people there or not? Because that's a tipping point. We can debate at the tipping point 30% or 50%.
- 7:47
- Unidentified speaker (speaker_1): And it will take time and that is why there is a lot of work required on opening up the data fabric. I'm not saying the data has to be clean, but it has to be accessible.
- 7:56
- Unidentified speaker (speaker_0): So how long would it take? So if I go back and say uh you know in a traditional services way, um first you gotta do c technology consulting, then you have to ask them to upgrade their infrastructure, then you have to figure out how to get data on a common platform. In this schema, am I how what is the flow chart that you tell your clients that this is what
- 8:16
- Unidentified speaker (speaker_0): you need to do before your agents can perform that task?
- 8:20
- Unidentified speaker (speaker_1): Um, less than six months from the moment somebody gets serious to when they can begin to get some value with agents. So I am actually not a believer in taking all the data into one place only. I think that is where people then suffer. If your core underlying system of record or system of engagement is not going to change, and you
- 8:40
- Unidentified speaker (speaker_1): take a copy of the data someplace else, then that's a perpetual task. You have to keep doing it. I actually believe put something in front of your systems of record and systems of engagement and use the data where it is. So unlock it. Uh I mean you mentioned us buying Confluent um a few months ago. Now you
- 8:59
- Unidentified speaker (speaker_1): understand why we felt that was such an important asset. Because keep the data where it is so those systems can keep going, but unlock it in real time in a really efficient way using uh some form of streaming data is I believe the much better answer. for unlocking the value inside an enterprise.
- 9:18
- Unidentified speaker (speaker_0): So in that same framework, a number of these system of record companies are creating their own agents. So, you know, SAP has some work, they will have it. So when you go to a client and say, Let me create a central agent for you, and that agent would then talk to those system of record agents, or do you construct it differently?
- 9:36
- Unidentified speaker (speaker_1): I think that this is going to become interesting.
- 9:39
- Unidentified speaker (speaker_1): So without naming companies, uh, let's take the domain areas. You kind of mentioned finance, ERP, and HR.
- 9:49
- Unidentified speaker (speaker_1): Some of these companies will say, I am not really going to expose enough APIs so agents can act on it. Others Brothers will say, look, I understand there is a lot of value in my core business logic and the core repository. because it has all kinds of audit, tax, compliance issues.
- 10:07
- Unidentified speaker (UNKNOWN): Um,
- 10:07
- Unidentified speaker (speaker_1): I'll give you my agent, but I'm not going to stop you from having your own agent or a custom agent also act on the underlying repositories. Certainly, when we work with SAP, that seems to be their modus operandi. They'll give agents, but they're not stopping you from doing it. And so I think that those people are going to
- 10:26
- Unidentified speaker (UNKNOWN): be able to do that.
- 10:27
- Unidentified speaker (speaker_1): find that they have more value because they cannot possibly own CRM and ERP and HR and procurement and and and. So they recognize that they'll give you some agents that are really great at operating on the single domain system, but they got to give you APIs to be able to
- 10:46
- Unidentified speaker (speaker_1): cross. And just in business, we're going to find some. are going to pretend that they don't want to give you those APIs. I actually believe that those are the ones who are going to have a saturation in the market, and they might find that they are the ones who get displaced because the value of having it through agents is so much that the
- 11:06
- Unidentified speaker (speaker_1): enterprises then are going to make a different decision.
- 11:09
- Unidentified speaker (speaker_0): There's another argument right now. So the evangelist of the model companies will say, hey, listen, I will just use my models and the models can go out and figure out the other thing. And while the system of record companies are saying I want to bucket these agents into procurement, into you know TNE and other things so that whatever is the task, because I not only
- 11:29
- Unidentified speaker (speaker_0): have the data but the context as well. How how which ways are your clients leaning towards?
- 11:35
- Unidentified speaker (speaker_1): So let's acknowledge that at a technical level, if I exclude and don't worry about revenue or profit or intellectual property, both methods are possible at a technical level. Okay. What's the danger in the first? If what you're exposing out as
- 11:55
- Unidentified speaker (speaker_1): your core intellectual property is really the way you make uh revenue and profit, if you are going to let the model companies in and play around with all of that and discover things, I'm not saying they're going to steal your data. I don't think that that's accurate. However, they are implicitly going to learn your
- 12:15
- Unidentified speaker (speaker_1): methodology and happily give it to everybody else. We have seen that example play out. So I think that sophisticated enterprises are very quickly going to get to the conclusion that we can't actually. do that. While it's technically possible, it's not good for our long term business. So then they are going to resort to, hey, I need
- 12:35
- Unidentified speaker (speaker_1): to run my own agents and I need to run them in a way that I can make sure that my proprietary IP does not leak out in an implicit way. It's not in an explicit way.
- 12:46
- Unidentified speaker (speaker_0): No, that's a fair point. And have you seen any change in uh enterprise behavior over? the last 12 months since all these incidents of hacking and the, you know, the models becoming so smart that they can go anywhere that they want? Um, have you seen any slowdown in that or are they trying to figure out putting some kind of guardrails around it?
- 13:06
- Unidentified speaker (speaker_1): I think there are three issues happening. So larger enterprises are very reluctant to let third-party agents inside the enterprise.
- 13:15
- Unidentified speaker (speaker_1): They may be okay with a third-party agent at what I call the old word would be DMZ or some system on the perimeter. They're certainly not okay with an unfettered agent from the outside coming inside the enterprise. This has actually got a little bit to do with the hacking and what happened around Mythurst Mythos and
- 13:35
- Unidentified speaker (speaker_1): Glasswing. It generally has to do with I don't want agents to be able to just access all data inside the enterprise because you don't quite know what they're going to do with
- 13:45
- Unidentified speaker (UNKNOWN): it.
- 13:46
- Unidentified speaker (speaker_1): Where consumer agents are now being put up by other companies. I've seen many large financial institutions essentially say, we cannot let the agent act on behalf of the individual. And I think there's both technology, but there's also all the legal apparatus we have built
- 14:05
- Unidentified speaker (UNKNOWN): up.
- 14:05
- Unidentified speaker (speaker_1): If the person is making the decision, we often will reduce to say, okay, if you make an error, you have a very small liability, we'll take care of the error. What happens if an agent makes the error and nobody realizes it for a few weeks? I think people haven't understood all of these issues because if you say the agent is acting on behalf of the individual, well, then
- 14:25
- Unidentified speaker (speaker_1): it can do trades, it can do money transfers, it can do purchases. Can you really unwind those? I don't think that's possible. So this is why they're being a bit cautious in saying, I'm not sure we're going to let other agents act on behalf of an individual. I think, but it's always the
- 14:45
- Unidentified speaker (speaker_1): mixed world.
- 14:46
- Unidentified speaker (UNKNOWN): On
- 14:46
- Unidentified speaker (speaker_1): the other hand, if you're a retailer, why do you care? I mean, if the agent wants to buy something, that's good for your business. So go ahead and buy it. If the agent wants to browse around, hey, plenty of search sites. In any case, uh, browse a retailer's site or an uh online or a physical inventory site and do things. So I think we're going to find
- 15:06
- Unidentified speaker (speaker_1): that for a lot of small things, I can see it for travel, for hospitality, for um uh lodging, uh for retail. it'll go on. And then you get the counter pull. Hey, hang on, but I want loyalty. And if I want loyalty and I want my brand and I want my experience put up, I do not really want
- 15:25
- Unidentified speaker (speaker_1): agents. I want the person uh coming in, or I want you to use my agent only, not somebody else's. So I think this is going to become an explosion of unfortunately many different options before we come back to there's only two or three use cases and use patterns that happen here. But I think for
- 15:45
- Unidentified speaker (speaker_1): medical, for financial, for taxation, probably government, it's hard for me to believe we will let uh agents in the wild uh be around. But there's many other things that are not that uh critical and that are in some sense not reversible, but the consequences. are minor. There, why not have the
- 16:05
- Unidentified speaker (speaker_1): convenience? Is kind of where I think I see it going.
- 16:09
- Unidentified speaker (UNKNOWN): No,
- 16:09
- Unidentified speaker (speaker_0): it's fair. And uh, as far as budgets are concerned, these enterprises are the are they funding these AI initiatives at the expense of other things? Because I I mean I would say that over the last two years, I have not seen discretionary IT spending, you know, improve. And probably in covering the space. for over 25 years, this is the longest slump that
- 16:29
- Unidentified speaker (speaker_0): I have seen. What's your view on the whole uh aspect of?
- 16:33
- Unidentified speaker (speaker_1): I'm going to give you a very bold prediction, Anurag. So your listeners are going to get excited on this. Let's go back to take long aucs. Go back 25 years ago. Average IT tech spend in enterprises, let's take the top 10,000, was about 3% of revenue. There's a few exceptions. Banks were always higher
- 16:53
- Unidentified speaker (speaker_1): in those days, and actually retailers were at the other end. Let's call it 3% in the aggregate. That number is already around 5 to 6% today. So while we don't notice it, it has doubled. While revenues, of course, have exploded. So the total tech spend as a proportion has doubled, and as an absolute
- 17:13
- Unidentified speaker (speaker_1): amount is, of course, five to six times. higher than uh 20 years ago. I'll give you a prediction that within five to ten years, 10 years at the outside, five years at the inside, tech spend is going to double as a proportion of revenue. But it goes back to the things I said. If I'm using
- 17:33
- Unidentified speaker (speaker_1): technology to do customer service, that means I can have a lot more revenue. revenue with the same number of people, but with technology augmenting my people.
- 17:44
- Unidentified speaker (UNKNOWN): If
- 17:44
- Unidentified speaker (speaker_1): I can get a lot more programming done, then I can probably get a lot more things built that help the entire enterprise. So I will assert that over the next five years, we're going to find tech spend going up.
- 17:58
- Unidentified speaker (UNKNOWN): And
- 17:59
- Unidentified speaker (speaker_1): we can see this happening in people's token bills, in people's agent bills. Right now, it's since it's still small, it's on the fringes, 1 to 10% of the total, they can absorb it, they can manage it. But I think everybody is going to start budgeting for how much do I have to spend on all this? And over that period, 5 to 10, I'm going
- 18:19
- Unidentified speaker (speaker_1): to predict total tech spend is doubling. Note, I'm not saying that this is all token. An agent is largely code and CPU. and data access with a little bit of token. So in this doubling, token spending maybe 10-20% of the total, but all the rest is on the usual. It's on people writing
- 18:38
- Unidentified speaker (speaker_1): code, people writing new applications, people unlocking new value, people changing workflows, unlocking data. And yes, it is all enabled by tokens, but tokens are not the only thing in that picture.
- 18:49
- Unidentified speaker (speaker_0): That's good. Let's talk about
- 18:50
- Unidentified speaker (UNKNOWN): it.
- 18:51
- Unidentified speaker (speaker_0): a little bit about software and you know, for our listeners, uh the IBM software unit now accounts for about forty-six percent of total IBM sales compared to around thirty-two percent when Arvind took over as the CEO. So Arwin, this has just gone so well since the acquisition. Profits have improved. You know, when you look at the world of AI, you know, which assets out
- 19:11
- Unidentified speaker (speaker_0): of the ones that you have right now are do you think are most most prone or most exposed to that spend?
- 19:19
- Unidentified speaker (speaker_1): Yeah. Uh so Anruk, I'm going to give a slightly uh longer arc on this, just a couple of years more. If I go back to 2018, uh, the year we made the offer to buy Red Hat, but closed in 2019, as you said at the beginning, uh, software was about 22% of IBM, and it is now 46%. This is a very
- 19:39
- Unidentified speaker (speaker_1): conscious strategy because we believe that this provides us the largest mode and the most innovation for our clients.
- 19:46
- Unidentified speaker (UNKNOWN): So
- 19:47
- Unidentified speaker (speaker_1): if I look at software, I kind of put it into a few buckets. Bucket one if the value of the software largely lies in how an end user interacts with it, that means a lot of it is around design, around UI, around how easy. it is to use, et cetera. I think all that software is subject
- 20:06
- Unidentified speaker (speaker_1): to risk in an AI world. Other extreme. We are moving money, we are uh worrying about uh retail deposits, we're worrying about um uh payments. Here, efficiency really matters because you're trying to get tens of billions of transactions done per peak off.
- 20:27
- Unidentified speaker (speaker_1): hour. Just think about if you wanted to use an LLM to do that. For a moment, let me not even question the computer science of is it probabilistic or is it deterministic? Let's assume it can do it. But spending a trillion cycles of compute per transaction as opposed to a million cycles of compute per transaction
- 20:47
- Unidentified speaker (speaker_1): is six orders of magnitude more expensive. So I think that class of software is nowhere near being subject to any kind of danger from the gen AI world.
- 20:59
- Unidentified speaker (UNKNOWN): Then
- 21:00
- Unidentified speaker (speaker_1): you're left with a couple in the middle. I think the closer you are to its deterministic, its high volume, as well as its mission critical, the less likely it is going to be subject to disruption. Core databases. I mean, AI doesn't store your data. in a way that you can retrieve it instantly. So all those things, I think, are not at all
- 21:19
- Unidentified speaker (UNKNOWN): to
- 21:19
- Unidentified speaker (speaker_1): be disrupted. Now you're left with the only one category, which is a business application.
- 21:26
- Unidentified speaker (UNKNOWN): So
- 21:26
- Unidentified speaker (speaker_1): the kind that was purely interactive, we're not gonna name companies, but you can see the impact on them. And I think the market, at an intuitive sense, has understood that. And then there's the a question mark. like if it's uh categories uh that are closer to maybe CRM or a few others like that.
- 21:46
- Unidentified speaker (speaker_1): I actually suspect there's core business logic in there. So those are not going to be subject to uh as much disruption as the pure interaction ones, but they're not as deep in being uh a payment system which is a completely completely, I think, not disruptible. So as we go across this
- 22:06
- Unidentified speaker (speaker_1): spectrum, most of what IBM has by design is in the non-disruptive bucket. When I look through our software, we had about 2% that lay in that interactive bucket.
- 22:21
- Unidentified speaker (speaker_1): We have a product called TriRega, which does uh lease management and a few things. So clearly there is some business logic, but the bulk lies in how people interact with it and it's an easy repository. Uh yeah, I think those could get uh disrupted with AI, but our gain from the other side, because
- 22:40
- Unidentified speaker (speaker_1): now volumes will go up on the underlying databases and data movement, would more than make up for this. At the end of the day, also, all of these agents have to run. in something. And that's a big uh tailwind for us on our Red Hat and OpenShift business. So
- 22:59
- Unidentified speaker (UNKNOWN): I
- 22:59
- Unidentified speaker (speaker_1): sort of gave you a think a broad brush across the categories uh which are going to be disrupted and which are less likely to be disrupted.
- 23:07
- Unidentified speaker (speaker_0): Perhaps um, you know, maybe uh dive a little into OpenShift. open shift And I, uh you know, I was really intrigued when we did the podcast. with Red Hat CEO. But I would love to get your opinion on where you are where do you think the world is in terms of that shift or that movement and you know what kind of a tailwind is it in the next several
- 23:27
- Unidentified speaker (speaker_0): years.
- 23:28
- Unidentified speaker (speaker_1): I think we are less than 10% of our journey into that. So if I take it to a big picture uh view on and Rog,
- 23:36
- Unidentified speaker (speaker_1): why is cloud attractive? Cloud is attractive because you get a lot of innovation. You get a much lower cost of operations and complexity because a lot of the underlying complexity is taken care of. Somebody worries about scaling, somebody worries about uh backup, somebody worries about attaching to storage. So all those things that
- 23:56
- Unidentified speaker (speaker_1): came from the traditional heterogeneous or sort of on-premise on premise server world, those things are taken care
- 24:02
- Unidentified speaker (UNKNOWN): of.
- 24:03
- Unidentified speaker (speaker_1): of.
- 24:03
- Unidentified speaker (UNKNOWN): So
- 24:04
- Unidentified speaker (speaker_1): to me, at a macro level, what is OpenShift? OpenShift is a platform in which people can deploy all of their applications and it takes care of all these things. Okay. So if it takes care of all these things, then you'll say, well, virtualization offered the same, but containers is even more flexible. So containers
- 24:24
- Unidentified speaker (speaker_1): is a superior architecture in terms of security, in terms of resilience, in terms of patching, all those things compared to VMs. Because VMs, you didn't get rid of all those things. In the container world, patching is simply restart the container and you're done. So there's some advantages at a deep technical level, as well as the scaling that
- 24:44
- Unidentified speaker (speaker_1): that comes. All of this, a lot of credit has to go to uh CNCF, the foundation, as well. as to Google, as well as to Red Hat for all the work uh that they did 10 years ago on getting Kubernetes. That's maybe just a proxy name for the core uh contribution, but there's a
- 25:03
- Unidentified speaker (speaker_1): lot more than Kubernetes. Okay. So now where are
- 25:08
- Unidentified speaker (UNKNOWN): we?
- 25:08
- Unidentified speaker (speaker_1): So in order to get some of those benefits that I talked about, the world went to virtualization. But virtualization tackled maybe half the things. It certainly tackled storage and backup and all those. Didn't really tackle how to scale well, didn't really tackle how to patch and how to secure well, et cetera. The container architecture gives you a way to get through
- 25:28
- Unidentified speaker (speaker_1): all that, which is why clouds also very heavily use container architectures. So, but if the enterprise wants a bit of flexibility, I'm going to do some things in my own data center, whether for software. sovereignty or for brand protection. I'm going to do some things in cloud one. I'm going to do some things in cloud two. OpenShift gets the benefit that it gives you a common way to
- 25:48
- Unidentified speaker (speaker_1): deploy across all of those. And in the on-premise side, I'd say it's probably the only way you can do it, unless you want to kind of do it yourself.
- 25:57
- Unidentified speaker (UNKNOWN): So
- 25:57
- Unidentified speaker (speaker_1): that is the tailwind that's behind OpenShift. And I think we're only 10% of the world. way in to this journey before it kind of becomes the default standard for all the things that I'm describing.
- 26:10
- Unidentified speaker (speaker_0): And in the c in the third quarter call, you also talked about application modernization being a big deal about it. Could you elaborate on that, please?
- 26:18
- Unidentified speaker (speaker_1): Certainly. So as people are looking to how do I begin to modernize an existing application? You kind of get three choices. Choice one,
- 26:29
- Unidentified speaker (UNKNOWN): I
- 26:30
- Unidentified speaker (speaker_1): don't really want to modernize it. I want to freeze dry it and run it. In that case, maybe virtualization works fine because you kind of freeze dry it and you're going to run it. But that means in all likelihood, in two, three, four years, you're going to sunset that thing because you're not improving
- 26:43
- Unidentified speaker (UNKNOWN): it.
- 26:44
- Unidentified speaker (speaker_1): Then sometimes people will say, I want to be able to use, and now with AI, you can do it really quick. Let me decode. decompose this and kind of rewrite it. Yeah, maybe 60% can stay the same, but I want to make a lot of improvements and flexibility in the other 40%. So now the question is: what scaffolding do you run that in? That's the big tailwind
- 27:04
- Unidentified speaker (speaker_1): for OpenShift because now you can say, I'm going to take that 40% and put it into a new container, and that's how I'm going to deploy it going forward. These are really big. tailwinds for us as we go forward. And then we're doing a lot of work to make sure that OpenShift is also a great place for you to write your agents and
- 27:24
- Unidentified speaker (speaker_1): run your agents when you're going to do that yourself, as opposed to just use third-party agents.
- 27:30
- Unidentified speaker (speaker_0): Great Arvind. Let's dive into the services business, please. And this is an area where uh you know we'd would love to learn from you because on a conceptual basis, when you think about, you know, custom software development or implementation of software, whether it's uh any packet software, you know, AI
- 27:50
- Unidentified speaker (speaker_0): should be able to reduce the cycle time or shrink the thing. And yet, even after you know, two years or so, I'm still seeing services companies headcount going up. Now, the revenue is not going up, I get that, because of the macro reasons we discussed. But why are they not right sizing their organization? Because you know,
- 28:09
- Unidentified speaker (speaker_0): if my custom coding is shrunk so much, you don't need, I mean, you're not going to make that much money. So please tell us how the services business is going to evolve next several years.
- 28:20
- Unidentified speaker (UNKNOWN): So
- 28:21
- Unidentified speaker (speaker_1): I'm going to put it in my usual words. There are tailwinds and there are headwinds.
- 28:25
- Unidentified speaker (UNKNOWN): So
- 28:25
- Unidentified speaker (speaker_1): what are the tailwinds? that are creating more demand for services? I need to get my data ready for AI. I need to make sure that I can go secure all my systems because if I'm getting it ready for AI, I want to also make sure that I'm not opening myself up to unintended harm. I got to harden my perimeter. So these are
- 28:45
- Unidentified speaker (speaker_1): all tailwinds for services. You talked about I can now get AI tools that can replace a part or a large part of what people were doing before. So I'll give you some examples. When we are doing uh SAP implementation, you spend a few weeks or months up front trying to document all the processes that
- 29:05
- Unidentified speaker (speaker_1): people have. So now you can transfer that process into an ERP system. Used to take 30 people, six months, interviews, documents. documentation, not all of it is perfectly documented, so you actually have to sit down with people. Using AI tools, there is no way that I don't see at least a 50%, if not even a larger,
- 29:25
- Unidentified speaker (speaker_1): improvement in total elapsed time. So that's a headwind for the amount of people that you use. So we can have this kind of debate because now we are going to speculate out into the future.
- 29:38
- Unidentified speaker (UNKNOWN): How
- 29:38
- Unidentified speaker (speaker_1): much of the total work done by these teams is in those areas where you can see significant reduction in elapsed human time? Certainly the coding task. But how many people are doing testing using AI tools? How many people are worrying about the acceptance, which is still very much a human task? So I
- 29:58
- Unidentified speaker (speaker_1): think right now you're kind of seeing a 10, 10-20% 20% improvement. Personally, I believe this is going to be more like 50 to 60% all in. And then the question is going to become now the complete unknown. Does that mean if that cost and time and risk goes down, will people say, hey, my appetite for getting more work done goes
- 30:18
- Unidentified speaker (speaker_1): up? So I want to get more work done? Or are they going to say, I'm just going to take this to the bottom line and stay the same? And my guess is we'll get a spread. and a spectrum across. And I would hazard a guess that in our own services business, certainly, if I go out a few years from now, we'll have
- 30:37
- Unidentified speaker (speaker_1): fewer people
- 30:39
- Unidentified speaker (speaker_1): than we have today.
- 30:41
- Unidentified speaker (speaker_1): That would be my supposition.
- 30:43
- Unidentified speaker (speaker_0): Yeah, I mean, this is how I've explained to Into people you know, go back 20 years and see there is in a simplistic world, there is a shiny new technology and there is an old technology. The old technology, you see deflation going in, the new technology, you have, you know, a lot more uh uh financial power to run with that. And over time, your portfolio keeps on changing, you know, whether that was
- 31:03
- Unidentified speaker (speaker_0): digital, whether that was cloud services, or etc. But it seems now that the compression on the project sites could be so much that, yes, we will see the turn, but there will come a time when you will have compression in revenue or or or we are already there. I don't know that. But where I've not seen is the
- 31:22
- Unidentified speaker (speaker_0): headcount shifting, and that's kind of confusing to me.
- 31:25
- Unidentified speaker (speaker_1): Yeah. Our headcount has actually gone down a tiny amount in services. I wouldn't say anything significant. But we actually manage our business that way. We can see one side of the business, like how do you maintain? older applications? How do you uh do all that work? We can see that that revenue stream is
- 31:45
- Unidentified speaker (speaker_1): decreasing because people are applying much more efficient techniques using the modern technology to do part of that work.
- 31:52
- Unidentified speaker (UNKNOWN): Then
- 31:52
- Unidentified speaker (speaker_1): there is a side, but I need to deploy AI more to get the benefits. That side is increasing. So we see that in the two sides of our own business, and you're completely correct. Except your past cycles kind of took 10 to 15 years to play out. I suspect this one will take five, not ten to fifteen. So it's going to go
- 32:12
- Unidentified speaker (speaker_1): faster.
- 32:13
- Unidentified speaker (speaker_0): No, that's a very fair point. And the other part is the strategic advantage of larger vendors over smallers, because this is probably the only industry that I can see, given the massive trillion plus TAM, that it has such a disaggregated market share by all of the others. But will having some expertise in AI
- 32:33
- Unidentified speaker (speaker_0): make it a lot more concentrated, or will the bigger companies take market share? Or how do you think that playing
- 32:39
- Unidentified speaker (UNKNOWN): out?
- 32:39
- Unidentified speaker (speaker_1): I actually think it's going to be an and.
- 32:42
- Unidentified speaker (speaker_1): So
- 32:44
- Unidentified speaker (speaker_1): take the world of enterprise tech, if you don't mind me going there and then coming back to services.
- 32:49
- Unidentified speaker (UNKNOWN): If
- 32:50
- Unidentified speaker (speaker_1): I go back to the late 90s and early 2000s, there were probably a dozen database vendors.
- 32:56
- Unidentified speaker (UNKNOWN): Of
- 32:57
- Unidentified speaker (speaker_1): those dozen, they've all evaporated. We are left with three majors
- 33:02
- Unidentified speaker (speaker_1): Oracle, Microsoft, IBM.
- 33:04
- Unidentified speaker (UNKNOWN): The
- 33:04
- Unidentified speaker (speaker_1): cloud guys have their own, which they write based on open source. But all the other database vendors, we won't take their names, have kind of disappeared.
- 33:13
- Unidentified speaker (UNKNOWN): Now,
- 33:14
- Unidentified speaker (speaker_1): they've been replaced with alternate data technologies, things all the key value stores and some of the JSON type data stores have but though have been added to replace that. So now let's go to the services world. I think that we'll have consolidation because only some will have the capacity to invest
- 33:34
- Unidentified speaker (speaker_1): in the AI tools. They need to augment their people,
- 33:37
- Unidentified speaker (UNKNOWN): and
- 33:37
- Unidentified speaker (speaker_1): then they got to invest. in these platforms. And it's not a one-time, it's a continuous investment that keeps
- 33:42
- Unidentified speaker (UNKNOWN): going.
- 33:43
- Unidentified speaker (speaker_1): That, by the very nature, always causes consolidation. However, services also has the other aspect that there are people who are domain experts. They understand how to do it for a law firm, or they understand how to do it for a consumer uh packaged goods company. So then you'll get people coming up there. where
- 34:03
- Unidentified speaker (speaker_1): actually their barrier to entry comes down because they can use AI tools as opposed to having armies of people. So I think you'll get consolidation on the majors,
- 34:14
- Unidentified speaker (UNKNOWN): and
- 34:14
- Unidentified speaker (speaker_1): then you'll get a new expansion because it's easier for people to become domain experts without having thousands of people, but maybe only hundreds of people. So I think both those dynamics are going to go on.
- 34:27
- Unidentified speaker (speaker_1): But it's a good thing because there's a lot of value then for the buyer from people who are domain experts who can help them on their fundamental core business.
- 34:36
- Unidentified speaker (speaker_0): Okay, fair point. Perhaps we jump and talk about how
- 34:41
- Unidentified speaker (speaker_0): AI is being used internally at IBM. And you know, if there are any examples you could share, that would be amazing.
- 34:47
- Unidentified speaker (UNKNOWN): Um,
- 34:48
- Unidentified speaker (speaker_1): I love my first example, which is drawn from the HR. world. And I kind of begin with a simple example. Uh for our employees here, at least in the US, but in other countries also, one of the things people love to do when they have a little bit of confidence and some savings is to buy a house. Okay.
- 35:08
- Unidentified speaker (speaker_1): So you need to go to a bank to get a mortgage for most people.
- 35:12
- Unidentified speaker (UNKNOWN): That
- 35:12
- Unidentified speaker (speaker_1): means the banks immediately is going to ask for. I need to verify your employment, I need to verify your history and all that. So what does the employee do? They go to their manager.
- 35:21
- Unidentified speaker (UNKNOWN): The
- 35:22
- Unidentified speaker (speaker_1): manager has no intention of doing all this. So they go to their HR partner. The HR partner has no intention of doing all this. They call up their HR back office somewhere and say, hey, do this for this person. So now think about all the steps. So 17 steps later, the HR back office looks up the complete. directory, looks up the HR record, looks up the
- 35:41
- Unidentified speaker (speaker_1): payroll database, looks up something else, and then writes a letter saying all this is there, and then they can either email or physically send that letter.
- 35:52
- Unidentified speaker (UNKNOWN): So
- 35:52
- Unidentified speaker (speaker_1): we asked ourselves a question.
- 35:54
- Unidentified speaker (UNKNOWN): If
- 35:54
- Unidentified speaker (speaker_1): you are on our network, we actually know exactly who you
- 35:57
- Unidentified speaker (UNKNOWN): are.
- 35:58
- Unidentified speaker (speaker_1): Unless I can verify that you're on a rock and you've done your uh two or three-factor three factor authentication. you're not on our network. So we said, why could we create an agent where you simply say, I want an employee verification letter? The only question the agent should ask you is email or physical mail to what address? The
- 36:17
- Unidentified speaker (speaker_1): agent has the right to look up all these systems. It doesn't need the human.
- 36:21
- Unidentified speaker (UNKNOWN): So
- 36:21
- Unidentified speaker (speaker_1): you then ask the question: how many other HR activities are there which we could do? like
- 36:27
- Unidentified speaker (UNKNOWN): this?
- 36:28
- Unidentified speaker (speaker_1): So, over a three-year journey,
- 36:32
- Unidentified speaker (speaker_1): 94% of all HR activities by volume
- 36:37
- Unidentified speaker (UNKNOWN): are
- 36:37
- Unidentified speaker (speaker_1): done using a single agent now.
- 36:40
- Unidentified speaker (UNKNOWN): And
- 36:41
- Unidentified speaker (speaker_1): this agent has access to
- 36:43
- Unidentified speaker (UNKNOWN): all
- 36:43
- Unidentified speaker (speaker_1): kinds of PII, but it has all the guardrails to make sure that none of that PII escapes the agent into one of the outer. outputs that it gives. I would tell you that most of our employees, probably 90%, don't even know what is our HR system. They simply interact with our agent called Ask HR, and
- 37:03
- Unidentified speaker (speaker_1): it does this on their behalf. Think of the advantages in terms of productivity, and this is 24 by 7, and it never really makes mistakes, and we can keep. adding and adding and adding to it. So we actually have a team that is continuously improving this agent. Nobody knows. And they just get the benefit as it goes along.
- 37:23
- Unidentified speaker (speaker_1): So that's one example of what we could do.
- 37:27
- Unidentified speaker (speaker_0): Yeah, this is a very interesting one. So in this case, you created this agent internally, and then it has rights to all the different, you could say, systems of records, whether it's a payroll system or a co. core HR system, et cetera?
- 37:40
- Unidentified speaker (speaker_1): Correct. So we worked with our underlying vendors, and I would tell you that the vendors we eventually ended up with are those who were the most cooperative in saying, we understand that you need these APIs and you need the underlying. Hey, but we get paid according to the number of employees you have, so that's okay. But I get it that I'm not
- 38:00
- Unidentified speaker (speaker_1): going to get the value for adding all these other visits. widgets over time because you're doing your own. But it's a much friendlier underlying set of repositories. So, yes, you're correct. About 10 to 15 different systems, this one had to go touch. But now think about procurement. You have maybe an expert buyer agreeing to
- 38:20
- Unidentified speaker (speaker_1): buy something, Thing. but now why do I need the person who's checking to make sure that that statement of work and the input and the cost and what you agreed to was all the same? You're just verifying information. So why can't I have an agent that does that? It doesn't take away the human expertise, but it takes away a lot of the, I'll call it rote work
- 38:41
- Unidentified speaker (speaker_1): that is done there.
- 38:43
- Unidentified speaker (speaker_1): How about getting more sophisticated on pricing in terms of um uh discounting and so on? So at this point, about 60 of our enterprise processes are in really good shape with these kinds of agents and workflows going on.
- 39:01
- Unidentified speaker (speaker_1): Another 60 are kind of making good progress. To give you an example that is completely different than HR, Are. how about uh IT tech support?
- 39:13
- Unidentified speaker (UNKNOWN): So
- 39:13
- Unidentified speaker (speaker_1): often there is a help desk and you call them and they look up and they give you information. And our reaction
- 39:18
- Unidentified speaker (UNKNOWN): was,
- 39:18
- Unidentified speaker (speaker_1): wait, isn't this much better done with an agent where you the agent actually knows you? The agent can actually get rights to even look inside your system if it's an IBM-owned system, and it can actually make suggestions far beyond and far better than what the human can. And everybody's now getting that same level of expertise as opposed to.
- 39:39
- Unidentified speaker (speaker_1): previously, you always say, hey, I want Joey because I know Joey is much better than somebody else. And so that is the advantage you get. So instead of having a team of a couple of hundred people on the IT help desk, it's actually the 20 experts who are helping the agent get better
- 39:56
- Unidentified speaker (speaker_1): day by day. And we tend to be conservative. If the agent is not confident of its answer, we tend to bump it up to a human.
- 40:04
- Unidentified speaker (speaker_0): So it seems that would help both in the customer satisfaction and your margin structure on that end.
- 40:10
- Unidentified speaker (speaker_1): Well, uh, how much have our margins gone up in the last
- 40:13
- Unidentified speaker (UNKNOWN): six years?
- 40:14
- Unidentified speaker (speaker_0): No, that's fair. Um Armin would love to learn about frontier models, open weight models, and uh small language models. I mean, uh as you have are talking to your enterprise. clients, you know, what is their appetite to use the all these different models and how should one think about
- 40:34
- Unidentified speaker (speaker_0): it?
- 40:35
- Unidentified speaker (speaker_1): I think the appetite is high. What people get worried about is the source of the good overweight models. So if they're only from a certain country, the people are worried because not for rational reasons, I'll say for irrational reasons. Now, I'll make a forward-looking prediction and then come back to
- 40:55
- Unidentified speaker (speaker_1): the logic on why. I think
- 40:59
- Unidentified speaker (speaker_1): in three years from now, 80% of all enterprise token usage will be on open weight or a model that the enterprise chooses where it runs. So, because it could also it could run maybe on premise, but you could also choose to run it in a cloud as long as you have access to GPUs and the right
- 41:18
- Unidentified speaker (speaker_1): infrastructure.
- 41:19
- Unidentified speaker (UNKNOWN): So
- 41:20
- Unidentified speaker (speaker_1): making a rather bold prediction that within three years, 80% of all tokens go
- 41:24
- Unidentified speaker (UNKNOWN): there.
- 41:25
- Unidentified speaker (speaker_1): When we were testing for the sake of patching open source, we used frontier models, we used open web weight models, we used models from China, we use models from the US, we used a whole variety, not so much small language models except for documentation. The cost of running these models, the
- 41:45
- Unidentified speaker (speaker_1): open weight, as opposed to the frontier, is anywhere from one-fifth to one-tenth that of the large, and they were about 96% as good for the tasks that we were wanting.
- 41:59
- Unidentified speaker (speaker_1): I think that this is where people have not understood this. When in the history of enterprises, which are now about three centuries old, if I'm offering something 90% as good at 10% of the cost, where is the enterprise going to go? And I think that this is so today it's not going
- 42:19
- Unidentified speaker (speaker_1): to happen because the average enterprise, if I leave the top 10 in the world aside, I'm going to say, look, I don't have the engineers who know how to optimize where. this open web model runs. I don't have the guardrails in place that I can make sure that it doesn't misbehave. I don't have the evals in place. I don't have the hard. So there's a lot of things, but these things are not going to take much more than
- 42:39
- Unidentified speaker (speaker_1): a year to get out there in a way that you can hit the easy button. It's like the OpenShift story. If I told a client, hey, you need to do operators and you need to have a secure Linux and you need to get your Kubernetes and you need to worry about a um a scaling engine and you need to worry about load balancing and you need to worry about having a secure
- 42:58
- Unidentified speaker (speaker_1): backup. They gotta say, look, I I get it, it's possible. I'm not gonna do all of that stuff because that's not our core competence. So I think within a year, people are gonna have these systems that they can say, basically, as long as I have the infrastructure, be it on a cloud or be it on premise, I can go ahead. and do that. So the I think it is driven by the
- 43:19
- Unidentified speaker (speaker_1): gap between Frontier and OpenWait is sort of closing. Part one. The economics are compelling for doing inference-only open weight, and it gives you a lot more protection in terms of leakage because now all your logs you're going to destroy at the end of the day. So there is no leakage of your
- 43:39
- Unidentified speaker (speaker_1): implicit IP either. So I think all those reasons are going to point to that's an ad. 20% of the tokens I think will still be on Frontier. And on the pure consumer case,
- 43:50
- Unidentified speaker (UNKNOWN): who
- 43:51
- Unidentified speaker (speaker_1): cares? That could probably be almost all on Frontier. But for enterprise usage, I think it goes 80 20.
- 43:57
- Unidentified speaker (speaker_0): That's a very good point. And you know, an area that I have not seen that much development in and maybe with you know in between. NVIDIA buying hugging face and more open stuff coming through, that may change is
- 44:08
- Unidentified speaker (speaker_0): sector specific model or models that are dedicated to one vertical specific solution. I think Salesforce has just released one that is very specific to them. So, you know, would it not be possible, let's say, for an auto company to download a generic model and make it more specific to their sector?
- 44:27
- Unidentified speaker (speaker_1): Yeah. So let's take the auto example. Could I take, start with an open weight model? Because the frontier people are going to be reluctant to just give you their model.
- 44:37
- Unidentified speaker (UNKNOWN): So
- 44:38
- Unidentified speaker (speaker_1): if I take an open weight model and I now
- 44:41
- Unidentified speaker (speaker_1): train it around warranty claims data,
- 44:45
- Unidentified speaker (UNKNOWN): now
- 44:45
- Unidentified speaker (speaker_1): suddenly, whenever somebody is at a dealership or is calling them up, if I can get even a 10
- 44:51
- Unidentified speaker (speaker_1): improvement from a generic use case of that, that is worth how much money? These guys all spend tens of billions on warranty claims. If I can get 10% improvement, that's a billion dollars a year straight to the bottom line for probably a hundred million dollar training task. So I think that these things are going to happen, and that is what
- 45:11
- Unidentified speaker (speaker_1): is partly going to drive open weight. So it's not so much a retrain. retraining of the model, people get confused. It's more an augmentation of the original model with this kind of data. So now this is the business model conflict. The frontier people want to charge the high price, so they'd be reluctant to give that to you. They'd want to keep that for themselves.
- 45:31
- Unidentified speaker (speaker_1): The open weight guys might say, hey, I'm making money in other ways. Take it, retrain it, and maybe I can help you run it. Maybe you want some help on that. But I think that this is going to go. This is the same way that operating systems went. Companies, including us, made a lot of money in proprietary operating systems for a long time. And once
- 45:51
- Unidentified speaker (speaker_1): Linux got to escape velocity, let's call it circa 2000, over time it has now become, I think, 70% of all enterprise workloads run on Linux. And by the way, I'd I would assert to you that before the decade is out, it'll be 90%.
- 46:07
- Unidentified speaker (speaker_0): Yeah, I'm very so certain there are more open uh there are more Linux workloads in Azure now than there are uh you know Microsoft, at least. That was that that was actually a long time ago. So I'm definitely sure that's the case. Um, Armin, this is such a great discussion. Any final words uh for our audience as to what should we think about with you know all this discussion around AI building.
- 46:26
- Unidentified speaker (speaker_0): buildup or anything. You know, any any final thoughts that would be amazing?
- 46:30
- Unidentified speaker (speaker_1): Look, I think the AI buildup to me, it's always a question of whenever there is a new technology, there is almost invariably overinvestment. By the way, at a macroeconomic level, I could sit back and say a lot of this got written about circa in the late 90s when we you talked about
- 46:51
- Unidentified speaker (speaker_1): what is a common good as opposed to what is a unique good. Building out all this infrastructure is going to pay off for the economy at large. Now, those who are providing the capital and the liquidity for this investment, some fraction of them are going to be disappointed. Not the majority, but some
- 47:11
- Unidentified speaker (speaker_1): fraction of those people, because of various reasons, when you go that fast and that quick, not everything is going to work out. But the vast majority will get used eventually for a great benefit for the economy and for economic growth. I personally come down to if we can keep our GDP growing north of
- 47:30
- Unidentified speaker (speaker_1): 3%,
- 47:32
- Unidentified speaker (speaker_1): it's worth taking a lot of risk because the impact of that over 10 years is so much more than all the risks. that are being talked about. And I think this is maybe the only thing we have that could get us there. So I'm a big believer in what AI is going to do for the economy. All that said,
- 47:52
- Unidentified speaker (speaker_1): uh, the underlying is always a little bit messy, maybe to use that word.
- 47:56
- Unidentified speaker (speaker_0): Great, Arwen. Thank you so much for your time. This has been such a fun discussion.
- 48:00
- Unidentified speaker (speaker_1): Always good talking to you, Anurag.
- 48:02
- Unidentified speaker (speaker_0): For our listeners, we want to thank you for tuning in. If you like the episode, please subscribe and leave a review and check back to hear conversations with the leading disruptors in the tech landscape.
- 48:12
- Unidentified speaker (UNKNOWN): If
- 48:13
- Unidentified speaker (speaker_0): you want to learn more about our research, including deep dives on topics like AI, check our work on the Bloomberg Terminal at BI Co. We also want to thank Aditya Thaditya Samani for his help in editing this podcast. This is your host, Anurag Anuragrana, Rana, signing off.
More from Tech Disruptors
Honeywell CTO on AI and Physical-World Autonomy
Episode discussion: Honeywell CTO outlines AI-driven autonomy strategy, citing edge computing and deterministic guardrails for industrial applications
Speakers’ claims: Mustafa Okur interviews Suresh Venkatarayalu, CTO and President of Forge at Honeywell, regarding the company's pivot from automation to autonomy. Venkatarayalu explains that Honeywell has simplified its portfolio to focus on commercial buildings and industrial process automation, leveraging its Forge platform to integrate AI with deterministic control systems. He describes a three-phase AI strategy involving internal knowledge models, partnerships with Nvidia for open-source models in secure hardware, and hybrid architectures for data sovereignty. The discussion highlights use cases in reducing false alarms in ethylene plants and building management systems, as well as collaborations with Google, LS Electric, and Dangote to expand ontologies and meet cybersecurity mandates.
Figma on Why Design Gains Value in the AI Era
Episode discussion: Kris Rasmussen on Figma's AI Strategy: Design Context as a Moat and Hybrid Pricing Models
Speakers’ claims: Kris Rasmussen, Chief Architect at Figma, discusses how AI agents serve as interface glue rather than replacements for traditional UIs, emphasizing that design systems provide critical context for better AI outputs. He argues that as code becomes commoditized, value shifts to design, with Figma leveraging proprietary user data to fine-tune open-weight models for specific workflows. Rasmussen highlights bottlenecks in AI adoption due to security and backend expertise gaps, noting that human curation remains essential. He also outlines Figma's infrastructure strategy, which avoids building foundation models from scratch, and predicts a shift toward hybrid seat-based and consumption-based pricing models.