Vantcast — The Intelligence Edition

Speakers warn that agentic AI's speed outpaces enterprise governance, creating security and cost risks without semantic layers and strict controls.

21 min Listen to the original Episode page
Agentic AIEnterprise SecurityAI GovernanceData ArchitectureCompute CostsNIST GuidelinesSemantic Layer

Summary

Speakers argue that autonomous AI agents can make high-stakes decisions faster than human organizations can govern them, creating significant security and financial risks. They emphasize the critical role of a 'semantic layer' to translate raw data into business context, preventing agents from acting on misinterpreted information. The discussion highlights research identifying a 2-3 second deauthentication window that allows agents to exfiltrate data or execute trades before access is revoked. Speakers also cite McKinsey data on efficiency gains but caution that these apply only to 'frontier territory' with pristine data architectures. Finally, they propose NIST-style cryptographic identities and board-level questions to manage liability and verify agent authority.

Key takeaways

  1. 1
    A McKinsey publication cites a 50% reduction in forecast errors and an 83% reduction in people hours for agentic AI, but these figures are explicitly classified as 'frontier territory' requiring pristine data infrastructure.

    Unidentified speaker (speaker_0) · 4:29

Episode sections

  1. 0:00 – 0:59

    Introduction to Agentic AI Risks

    Speakers introduce the transition from passive AI to autonomous agents that can make million-dollar decisions in seconds. They highlight the architectural challenge of enterprise systems not being built to handle software that reasons and acts independently.

  2. 0:59 – 3:28

    The Critical Role of the Semantic Layer

    Richard Nock of Flatiron Health is cited to argue that AI agents lack intuition and only know the context provided. The semantic layer is described as the essential translation engine that maps raw database tables to business logic, preventing agents from acting on misunderstood data.

  3. 3:28 – 4:27

    NOCA Structural Fixes and Benchmarking

    NOCA proposes a structural fix where AI agents are programmatically benchmarked against historical human analyses and decisions to verify they interpret business telemetry correctly before live execution.

  4. 4:27 – 6:51

    McKinsey Data and Implementation Caveats

    A McKinsey publication claims 50% reduction in forecast errors and 83% reduction in people hours for agentic AI. However, speakers note these figures apply to 'frontier territory' with pristine data architectures, warning that skipping governance steps leads to faster errors, not insights.

  5. 6:51 – 10:01

    Security Vulnerabilities and Deauthentication Attacks

    Research by Vishal Sharma at Queen's University Belfast identifies a 2-3 second deauthentication window in enterprise AI. This latency allows autonomous agents to exfiltrate data or execute trades before access is revoked, exploiting the fact that agents operate at machine speed rather than human speed.

  6. 10:01 – 13:05

    NIST Guidelines and Governance Requirements

    NIST is proposed as mandating distinct cryptographic identities for AI agents, requiring least privilege and dynamic authorization. The speakers emphasize the need for immutable, cryptographically secured evidence trails to audit agent decisions, moving beyond traditional service accounts.

  7. 13:05 – 14:51

    Financial Risks and Compute Costs

    Reuters and McKinsey project token spending could reach 20% of labor costs. Speakers explain that 'React loops' (reasoning and acting) cause agents to burn massive compute resources when navigating poor data quality, leading to uncontrolled financial hemorrhage via cloud bills.

  8. 14:51 – 17:16

    Salesforce Governance Approach

    Melissa Ramey of Salesforce advocates defining the task before choosing the model to enforce least privilege. She recommends continuous performance reviews for AI agents to manage permission drift as business logic and data sources evolve over time.

  9. 17:16 – 21:01

    Board-Level Questions and Future Interagent Communication

    Speakers propose three board-level questions: proving net business value, naming human liability owners, and cryptographically proving authority limits. The episode concludes with a speculative scenario of AI agents negotiating in encrypted dialects, raising legal and security questions.

Claims worth verifying

Richard Nock is the head of data science and AI at Flatiron Health.

Unidentified speaker (UNKNOWN) ·1:16

How to verify: Check Flatiron Health leadership page or LinkedIn profile for Richard Nock's title and employer. For recent events, verify 2026-08-15 through 2026-11-15. Use another period only when the speaker explicitly names it.

McKinsey publication cites a 50% reduction in forecast errors and an 83% reduction in people hours for agentic AI.

Unidentified speaker (speaker_0) ·4:29

How to verify: Locate the specific McKinsey report on agentic AI in business planning published around the episode date and check the executive summary for these metrics. For recent events, verify 2026-08-15 through 2026-11-15. Use another period only when the speaker explicitly names it.

Vishal Sharma at Queen's University Belfast published research on deauthentication attacks targeting enterprise AI agent systems.

Unidentified speaker (speaker_1) ·7:46

How to verify: Search academic databases or Queen's University Belfast research portal for papers by Vishal Sharma on AI agent security and deauthentication. For recent events, verify 2026-08-15 through 2026-11-15. Use another period only when the speaker explicitly names it.

The research identified a two to three second deauthentication window in enterprise AI systems.

Unidentified speaker (speaker_0) ·7:52

How to verify: Review the specific findings of the Queen's University Belfast paper for the measured latency period between access revocation and system propagation. For recent events, verify 2026-08-15 through 2026-11-15. Use another period only when the speaker explicitly names it.

NIST is mandating that AI agents need distinct cryptographic identities, least privilege, and dynamic authorization.

Unidentified speaker (speaker_1) ·10:16

How to verify: Check NIST publications or press releases regarding AI agent security frameworks and identity management requirements. For recent events, verify 2026-08-15 through 2026-11-15. Use another period only when the speaker explicitly names it.

Reuters and McKinsey project token spending could reach 20% of labor costs in some sectors.

Unidentified speaker (UNKNOWN) ·11:41

How to verify: Locate the joint or separate reports from Reuters and McKinsey discussing AI compute costs and labor cost comparisons. For recent events, verify 2026-08-15 through 2026-11-15. Use another period only when the speaker explicitly names it.

United Airlines is restricting AI access to its internal systems over data security concerns.

Unidentified speaker (speaker_1) ·11:34

How to verify: Search news archives for United Airlines press releases or news reports regarding AI access restrictions and security policies. For recent events, verify 2026-08-15 through 2026-11-15. Use another period only when the speaker explicitly names it.

Australia is implementing mandatory AI incident reporting after an AI agent gained unauthorized access to a government portal.

Unidentified speaker (UNKNOWN) ·11:19

How to verify: Check Australian government announcements or news reports regarding new AI incident reporting regulations and the specific government portal breach. For recent events, verify 2026-08-15 through 2026-11-15. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify McKinsey's 'frontier territory' claim by reviewing the specific methodology and data quality requirements in their published report on agentic AI efficiency.
  • Check Queen's University Belfast research papers by Vishal Sharma for technical details on the 2-3 second deauthentication window and its implications for enterprise security.
  • Review NIST guidelines on cryptographic identities for AI agents to understand the proposed standards for least privilege and dynamic authorization.
  • Assess your organization's current 'semantic layer' implementation to determine if it effectively maps raw database tables to business logic, preventing agent misinterpretation.
  • Audit cloud billing and token usage reports to identify potential 'React loop' waste where agents burn compute resources due to poor data quality or lack of context.
  • Evaluate Salesforce's governance approach by reviewing Melissa Ramey's recommendations on defining tasks before model selection to prevent permission drift.
  • Investigate board-level liability frameworks to ensure human owners are named for AI agent decisions and that cryptographic proof of authority limits is in place.
Full transcriptexpand
0:00
Unidentified speaker (speaker_0): Welcome to Vancast, your vantage point, every room, every conversation. Today we are tearing into a well, a hidden dependency chain in enterprise AI. We're looking at a critical blind spot. Like what actually happens when we transition from AI that just generates text to AI that takes autonomous action on behalf of your business.
0:20
Unidentified speaker (speaker_1): Right.
0:21
Unidentified speaker (speaker_1): Massive shift.
0:21
Unidentified speaker (speaker_0): Yeah, totally. Because we've compiled findings for this deep dive from McKinsey, Reuters, NIST, and some leadership. leading data scientists really to map out this exact threat and opportunity, honestly. Because
0:32
Unidentified speaker (UNKNOWN): you know,
0:33
Unidentified speaker (speaker_0): what happens when an AI agent commits your company to a million dollar decision in the two seconds it takes your security system to realize it shouldn't be on the network?
0:42
Unidentified speaker (speaker_1): Which is terrifying.
0:44
Unidentified speaker (speaker_0): I mean, it is the defining architectural challenge of the next five years.
0:47
Unidentified speaker (speaker_1): Yeah. Because we're moving from passive intelligence to active excuse. execution.
0:51
Unidentified speaker (UNKNOWN): And
0:52
Unidentified speaker (speaker_1): the infrastructure of the modern enterprise, it just fundamentally was not built to handle software that can reason and act on its own.
0:59
Unidentified speaker (UNKNOWN): Aaron Powell Okay,
0:59
Unidentified speaker (speaker_0): let's unpack this because I want to start by challenging a really common assumption I see across these implementations. There's this rush, right, to just plug agentic AI into existing data lakes. People assume the model is like smart enough to just figure out the company's data structure. Trevor
1:13
Unidentified speaker (UNKNOWN): Burrus, Jr.
1:14
Unidentified speaker (speaker_1): Oh, absolutely. They think it's magic. Trevor
1:16
Unidentified speaker (UNKNOWN): Burrus, Jr.
1:16
Unidentified speaker (speaker_0): Right. But Richard Nock, he's the head of data science and AI at Flatiron Health, he argues something completely different. He insists that agents only know the context they're given. Like they don't have intuition. Aaron
1:27
Unidentified speaker (UNKNOWN): Ross Powell
1:27
Unidentified speaker (speaker_1): No, they don't.
1:28
Unidentified speaker (speaker_0): Yeah.
1:28
Unidentified speaker (speaker_1): And Nock is pointing to a well, a massive failure of imagination in how we deploy these systems. Aaron
1:35
Unidentified speaker (UNKNOWN): Powell
1:35
Unidentified speaker (speaker_0): How so? Aaron
1:35
Unidentified speaker (UNKNOWN): Ross Powell
1:35
Unidentified speaker (speaker_1): Well to understand the macroscopic business failures, you really have to
1:38
Unidentified speaker (UNKNOWN): do that.
1:39
Unidentified speaker (speaker_1): look at the microscopic mechanics of how an LLM, even an agentic one, processes data. When you point an agent at a corporate database, it doesn't intuitively understand your business model.
1:51
Unidentified speaker (speaker_0): Right.
1:52
Unidentified speaker (speaker_1): It relies entirely on what we call the semantic layer.
1:55
Unidentified speaker (speaker_0): The semantic layer. Trevor
1:57
Unidentified speaker (UNKNOWN): Yeah. Burrus,
1:57
Unidentified speaker (speaker_1): In a robust data architecture, the semantic layer is like the translation engine. It maps really raw cryptic database tables, like say you have a column named Custstat99. Aaron
2:09
Unidentified speaker (UNKNOWN): Powell
2:09
Unidentified speaker (speaker_0): Sounds like every database I've ever seen.
2:11
Unidentified speaker (speaker_1): Exactly. It maps that into human business logic, defining it as, you know, active subscriber. Aaron
2:17
Unidentified speaker (UNKNOWN): Powell
2:17
Unidentified speaker (speaker_0): See, I've always thought of the semantic layer as something built for human analysts, analyst, you know, so they don't have to write complex SQL queries just to build a simple dashboard. Aaron
2:26
Unidentified speaker (UNKNOWN): Powell
2:26
Unidentified speaker (speaker_1): And it used to be just that. But now it is literally the only thing keeping your AI from going rogue.
2:33
Unidentified speaker (speaker_0): Wait, really?
2:34
Unidentified speaker (speaker_1): Yeah. What's fascinating here is that if your semantic layer is incomplete, or if say different departments define a metric like churn in conflicting ways, a human analyst is going to notice that discrepancy, right? They'll they'll stop and ask a colleague for clarification.
2:50
Unidentified speaker (speaker_0): They'll send a Slack message like, hey, what does this mean? Aaron
2:53
Unidentified speaker (UNKNOWN): Powell
2:53
Unidentified speaker (speaker_1): Exactly. But an AI agent, it does not possess that hesitation mechanism. It is fundamentally a next token prediction engine optimized for completion. It just doesn't pause to flag a syntax error in your business logic. Aaron
3:07
Unidentified speaker (UNKNOWN): Powell
3:07
Unidentified speaker (speaker_0): So it just bulldoces forward and gives you a completely confident, computationally perfect, but functionally disastrous. answer.
3:15
Unidentified speaker (speaker_1): Yes. Repeatedly, at scale, and at a speed that human compliance teams just cannot monitor. Basically becomes a highly efficient, confident liar.
3:26
Unidentified speaker (speaker_0): A confident liar.
3:27
Unidentified speaker (speaker_1): Yeah.
3:28
Unidentified speaker (speaker_0): Wow.
3:28
Unidentified speaker (speaker_1): Yeah. And this is why NOCA proposes a structural fix. You cannot trust the agent's out-of-the-box reasoning. You just can't. You have to work backward by taking historical human uh analyses, uh, decisions your skilled data teams have already mapped out over the past year, and programmatically benchmark the agent against them.
3:47
Unidentified speaker (speaker_0): So you're like forcing the model to prove it interprets the raw telemetry of your business exactly the same way your best human experts did.
3:54
Unidentified speaker (speaker_1): Right. Before you ever let it execute a live Lives task. ask.
3:57
Unidentified speaker (speaker_0): Got it. I mean, to put a visual on this for you listening, it feels like hiring an incredibly eager, fast working intern who speaks five languages, but they don't understand the company's internal slang.
4:09
Unidentified speaker (speaker_1): Yes.
4:10
Unidentified speaker (speaker_0): So you refuse to give them the documentation and they just go do the wrong task at lightning speed.
4:15
Unidentified speaker (speaker_1): That's a perfect analogy.
4:17
Unidentified speaker (speaker_0): So what does this all mean for companies that are rushing to roll these agents out without checking their definitions first? Because that terrifying reality makes me really skeptical of the business outcomes being promised. right now.
4:27
Unidentified speaker (UNKNOWN): Well,
4:27
Unidentified speaker (speaker_1): they're setting themselves up for a lot of pain.
4:29
Unidentified speaker (speaker_0): Let's look at the new McKinsey publication on agentic AI in business planning. They're citing a 50% reduction in forecast errors and an 83% reduction in people hours.
4:40
Unidentified speaker (speaker_1): Those are massive numbers.
4:41
Unidentified speaker (speaker_0): Huge. But I hear numbers like that, and my immediate instinct is, eh, that's just consulting hype. I mean, are companies actually achieving an 83% reduction in labor in a production environment? Aaron
4:51
Unidentified speaker (UNKNOWN): Powell
4:51
Unidentified speaker (speaker_1): So if we connect this to the bigger picture, we have to look very closely at the caveats in McKinsey's own data to understand those numbers.
4:59
Unidentified speaker (speaker_0): Okay.
4:59
Unidentified speaker (speaker_1): Because they explicitly classify fully agentic sales and operations planning, where the AI is driving the entire continuous planning cycle autonomously as frontier territory. Aaron
5:11
Unidentified speaker (UNKNOWN): Powell
5:11
Unidentified speaker (speaker_0): Ah, frontier territory. Aaron
5:12
Unidentified speaker (UNKNOWN): Powell Not
5:12
Unidentified speaker (speaker_1): standard practice.
5:13
Unidentified speaker (speaker_0): Right.
5:14
Unidentified speaker (speaker_1): Those massive efficiency gains. They're derived from highly controlled, very specific implementations where the underlying data architecture is already pristine.
5:23
Unidentified speaker (speaker_0): Pristine. So it's definitely not a plug and play solution that you just buy and suddenly your supply chain team gets to go home early on a Friday. Aaron
5:32
Unidentified speaker (UNKNOWN): Powell
5:32
Unidentified speaker (speaker_1): Oh far from it. What we are actually seeing in the broader market is a blend of traditional forecasting tools supplemented by emergent agentic capabilities. I mean, the vision of a continuous, autonomous decision cycle is technically sound.
5:46
Unidentified speaker (speaker_0): Sure.
5:46
Unidentified speaker (speaker_1): But to unlock it, organizations have to respect a very strict dependency chain.
5:52
Unidentified speaker (UNKNOWN): You
5:52
Unidentified speaker (speaker_1): cannot leapfrog data governance.
5:54
Unidentified speaker (speaker_0): You just can't.
5:55
Unidentified speaker (speaker_1): No. First, you implement rigorous data governance. Second, you build that semantic layer we just discussed. Third, you overhaul your security architecture. And only after those three layers are locked in do you actually realize that 83% reduction in planning value. Aaron
6:11
Unidentified speaker (UNKNOWN): Powell
6:11
Unidentified speaker (speaker_0): I mean, are we just putting a Ferrari engine being the AI agent into a golf cart, which is our current data systems? Aaron
6:18
Unidentified speaker (UNKNOWN): Powell
6:18
Unidentified speaker (speaker_1): That is exactly what's happening.
6:20
Unidentified speaker (speaker_0): Because if an organization tries to skip those first three steps, deploying agentic capabilities before the data foundation is solid. it just gets you faster errors, not faster insights.
6:31
Unidentified speaker (speaker_1): That is the exact mechanism of failure. Because when an agent enters a reasoning loop, it will pull data, analyze it, hit a wall because of poor data governance, try a different internal system, pull conflicting data, and then it'll attempt to synthesize a plan based entirely on noise.
6:47
Unidentified speaker (speaker_0): Wow.
6:48
Unidentified speaker (speaker_1): It just amplifies the structural weaknesses of your organization. Aaron
6:51
Unidentified speaker (UNKNOWN): Powell And
6:51
Unidentified speaker (speaker_0): if that Ferrari engine is aggressively pushing through every system in the enterprise, it completely shatters how we approach security. I mean, we rely heavily on zero trust architecture right now, but our zero trust models are predicated on humans, right?
7:06
Unidentified speaker (speaker_1): Exactly. Zero trust fundamentally assumes a perimeter, even if it's a microperimeter.
7:12
Unidentified speaker (UNKNOWN): It
7:12
Unidentified speaker (speaker_1): assumes a human logs in, verifies their identity, via multifactor authentication, establishes a session, and operates within the bounds of a specific role. Aaron
7:21
Unidentified speaker (UNKNOWN): Ross Powell
7:21
Unidentified speaker (speaker_0): Right. A human employee swipes their badge once to get into the building, but an AI agent is like a ghost walking through walls, trying to grab files from HR, finance, and IT all at once in the exact same Maxim millisecond.
7:34
Unidentified speaker (UNKNOWN): Which
7:34
Unidentified speaker (speaker_1): introduces vulnerabilities we are barely even beginning to understand. And this is where the brand new research led by Vishal Sharma at Queen's University Belfast is just so critical.
7:44
Unidentified speaker (speaker_0): Yeah, let's talk about that.
7:46
Unidentified speaker (speaker_1): They recently published findings on what they term deauthentication attacks targeting enterprise AI agent systems.
7:52
Unidentified speaker (speaker_0): I read through that paper, and the specific detail that honestly stopped me in my tracks was their discovery of a two to three second deauthentication window. Aaron
8:00
Unidentified speaker (UNKNOWN): Powell
8:01
Unidentified speaker (speaker_1): Right.
8:01
Unidentified speaker (UNKNOWN): And
8:01
Unidentified speaker (speaker_1): we should carefully calculate. caveat that this is experimental research, not a post-mortem of a live mass global breach.
8:07
Unidentified speaker (UNKNOWN): Sure.
8:08
Unidentified speaker (speaker_1): However, the mechanism they uncovered is profound because AI agents fundamentally break existing zero trust security architectures. Aaron
8:15
Unidentified speaker (UNKNOWN): Powell
8:15
Unidentified speaker (speaker_0): Explain how that mechanism actually works for us, because to a human, two seconds is like the time it takes to blink. Why is a two to three-second three second window catastrophic in this context? Aaron
8:25
Unidentified speaker (UNKNOWN): Powell
8:25
Unidentified speaker (speaker_1): Because of how cloud microservices communicate. When an administrator or an automated security protocol revokes an agent's access token, maybe because it detects anomalous behavior.
8:35
Unidentified speaker (UNKNOWN): Right.
8:36
Unidentified speaker (speaker_1): That revocation takes time to propagate across every single server and service in a distributed enterprise network. There's a latency period.
8:43
Unidentified speaker (UNKNOWN): For
8:44
Unidentified speaker (speaker_1): a human clicking through a web interface, a three three-second second delay before they're logged out is totally meaningless. But an AI agent operates at machine speed.
8:51
Unidentified speaker (speaker_0): Right.
8:52
Unidentified speaker (speaker_1): In three seconds, an autonomous agent running a concurrent Python script can exfiltrate millions of rows of customer data, execute thousands of trades, or rewrite hundreds of configuration files before the system fully acknowledges that its access has been revoked.
9:06
Unidentified speaker (speaker_0): That's insane. And the Belfast researchers pointed out that bad actors can exploit this by manipulating the agent's context rather than using traditional malware. right? Like how does a system even authenticate a ghost whose context can be manipulated like that?
9:21
Unidentified speaker (speaker_1): It's incredibly difficult. Because agents rely on the data they ingest to make decisions. An attacker doesn't necessarily need to preach the firewall.
9:28
Unidentified speaker (speaker_0): Oh wow.
9:28
Unidentified speaker (speaker_1): Yeah. If they can inject malicious instructions into a data source the agent routinely reads, say um
9:34
Unidentified speaker (UNKNOWN): a
9:35
Unidentified speaker (speaker_1): public-facing public facing support ticket or a vendor inventory file. The agent ingests that instruction. Aaron
9:40
Unidentified speaker (UNKNOWN): Powell
9:40
Unidentified speaker (speaker_0): And just runs with it.
9:41
Unidentified speaker (speaker_1): Exactly. It then autonomously initiates unauthorized actions inside the network.
9:46
Unidentified speaker (UNKNOWN): And
9:46
Unidentified speaker (speaker_1): if the security system catches it and issues a deauthentication command, the attacker is just relying on that latency window for the agent to finish the malicious payload execution. Aaron
9:55
Unidentified speaker (UNKNOWN): Powell
9:55
Unidentified speaker (speaker_0): Because it bypasses the traditional identity check since the agent itself was already authorized.
10:00
Unidentified speaker (speaker_1): Yes.
10:01
Unidentified speaker (UNKNOWN): The
10:01
Unidentified speaker (speaker_1): agent basically becomes the alternative. ultimate insider threat completely by accident. Which explains why NIST, the uh National Institute of Standards and Technology, is scrambling to create a new framing specifically for AI agents. Aaron
10:14
Unidentified speaker (UNKNOWN): Powell
10:14
Unidentified speaker (speaker_0): What are they proposing?
10:16
Unidentified speaker (speaker_1): They're arguing that traditional service accounts just aren't enough anymore. They are mandating that agents need distinct cryptographic identities. They require least. privilege and dynamic authorization, meaning the agent's permissions must shrink and expand in real time, depending on the exact task it's executing. Aaron
10:36
Unidentified speaker (UNKNOWN): Powell
10:36
Unidentified speaker (speaker_0): Which sounds incredibly complex to build.
10:39
Unidentified speaker (speaker_1): It is. And crucially, they need immutable evidence trails.
10:43
Unidentified speaker (speaker_0): Okay. Meaning a mathematically provable log of exactly why the agent made a decision, what data it looked at, and what token it used.
10:51
Unidentified speaker (speaker_0): logged in a way that the agent itself cannot alter.
10:55
Unidentified speaker (speaker_1): Because when the audit comes, you can't just point to a black box neural network and say, uh, it decided to do it. You need a cryptographic chain of custody for the decision. Aaron
11:05
Unidentified speaker (UNKNOWN): Powell
11:05
Unidentified speaker (speaker_0): Right.
11:06
Unidentified speaker (UNKNOWN): And
11:06
Unidentified speaker (speaker_0): we are already seeing the fallout, the actual live incidents and hidden costs of companies ignoring this level of governance. I mean, this isn't just theoretical. whiteboard architecture anymore.
11:16
Unidentified speaker (speaker_1): No, it's actively disrupting production environments right now.
11:19
Unidentified speaker (UNKNOWN): Let
11:19
Unidentified speaker (speaker_0): me rapid fire three major news items just from this week for you listening. One, Australia is implementing mandatory AI incident reporting because an AI agent gained unauthorized access to a government portal, and this was a production system, not a test.
11:34
Unidentified speaker (speaker_1): Right.
11:34
Unidentified speaker (speaker_0): Two, United Airlines is actively restricting AI bought access to its internal. systems over data security concerns.
11:41
Unidentified speaker (UNKNOWN): And
11:41
Unidentified speaker (speaker_0): three, Reuters and McKinsey are projecting that token spending, so the compute cost to run these agents could reach 20% of labor costs in some sectors.
11:50
Unidentified speaker (speaker_1): Yeah. And this raises an important question regarding architecture. Because these incidents are really the picture of an enterprise security space moving much, much faster than the governance frameworks meant to contain
12:01
Unidentified speaker (UNKNOWN): it.
12:01
Unidentified speaker (speaker_0): Yeah, we're building the plane while flying it.
12:03
Unidentified speaker (speaker_1): Exactly. And historically, security teams favor fragmentation. Siloing data, so a breach in one department doesn't compromise another. But in the era of agentic AI, highly siloed fragmented architectures actually increase your unmanaged access path.
12:20
Unidentified speaker (speaker_0): How so? I mean, when one I would think siloing data is exactly how you stop the agent from wandering into the HR files.
12:26
Unidentified speaker (speaker_1): You'd think so. But if your data is fragmented across dozens of legacy systems, you have to create dozens of unique unmanaged access paths, like custom API bridges and legacy connectors, just to get the agent to work in the first place.
12:38
Unidentified speaker (speaker_0): Oh, I see.
12:39
Unidentified speaker (speaker_1): You're essentially poking hundreds of tiny holes in your own silos. You distribute the governance, which effectively means you have zero visibility into how the agent is traversing the network. Aaron
12:48
Unidentified speaker (UNKNOWN): Powell
12:48
Unidentified speaker (speaker_0): That makes a lot of sense.
12:49
Unidentified speaker (speaker_1): Centralized systems. on the other hand, provide a single, unified plane where you can actually implement that NIST level dynamic authorization.
12:57
Unidentified speaker (speaker_0): Right.
12:57
Unidentified speaker (UNKNOWN): It
12:58
Unidentified speaker (speaker_1): requires a heavier upfront lift to build a centralized security layer specifically for AI, but it is the only way to maintain governance.
13:05
Unidentified speaker (UNKNOWN): Well,
13:05
Unidentified speaker (speaker_0): here's where it gets really interesting to me, looking at the financial detail. We are deploying AI to save money and labor,
13:12
Unidentified speaker (UNKNOWN): but
13:12
Unidentified speaker (speaker_0): the token cost costs could eat up a fifth of
13:14
Unidentified speaker (UNKNOWN): the.
13:15
Unidentified speaker (speaker_0): our labor budget. How can companies possibly model for an exposure they don't even fully understand yet? Aaron
13:20
Unidentified speaker (UNKNOWN): Powell
13:21
Unidentified speaker (speaker_1): That's the 20% figure you mentioned. And it completely undermines the 83% labor reduction metric if you aren't governing the system properly. Aaron
13:29
Unidentified speaker (UNKNOWN): Powell
13:29
Unidentified speaker (speaker_0): But how does a token bill even get that high? If I ask a chat Chatbot bot to write an email, it burns like a fraction of a cent.
13:36
Unidentified speaker (speaker_1): Right.
13:37
Unidentified speaker (speaker_0): How does an enterprise agent rack
13:38
Unidentified speaker (UNKNOWN): up?
13:39
Unidentified speaker (speaker_0): compute costs equivalent to a fifth of a company's payroll? Aaron
13:42
Unidentified speaker (UNKNOWN): Powell
13:42
Unidentified speaker (speaker_1): Because enterprise agents do not use simple prompt and response mechanics. They use what we call React loops, reasoning and acting.
13:50
Unidentified speaker (speaker_0): Okay.
13:51
Unidentified speaker (UNKNOWN): When
13:51
Unidentified speaker (speaker_1): you give an agent a complex planning task, it breaks that task down into dozens of subtasks. It queries a database, reads the output, realizes the data is incomplete, writes a new query, tests it, includes
14:02
Unidentified speaker (UNKNOWN): the data.
14:02
Unidentified speaker (speaker_1): encounters an error, and self-corrects.
14:05
Unidentified speaker (speaker_0): Over and over.
14:05
Unidentified speaker (speaker_1): Yes. Every single one of those internal thoughts, every API call, and every self-correction requires processing thousands of tokens of context through a massive neural network.
14:16
Unidentified speaker (speaker_0): So if the semantic layer we talked about earlier is broken, the agent doesn't just fail and stop. It frantically loops, trying to solve a problem it lacks the context for, burning thousands of dollars in the world. of compute time in minutes.
14:28
Unidentified speaker (speaker_1): Exactly.
14:29
Unidentified speaker (UNKNOWN): If
14:29
Unidentified speaker (speaker_1): you have an ungoverned agent navigating a fragmented architecture with poor data quality, you are literally financing an infinite loop of compute waste.
14:38
Unidentified speaker (speaker_0): Wow.
14:39
Unidentified speaker (speaker_1): You aren't just risking a data breach. You are risking an uncontrolled financial hemorrhage happening entirely in the background of your AWS or Azure bill. Aaron
14:47
Unidentified speaker (UNKNOWN): Powell
14:47
Unidentified speaker (speaker_0): Okay. Given the massive financial risks and the security blind spots.
14:51
Unidentified speaker (speaker_0): we need to talk about rebuilding governance for the AI era. How does a listener or an organization actually take control of this technology today? Aaron
14:59
Unidentified speaker (UNKNOWN): Powell
14:59
Unidentified speaker (speaker_1): Well we have to look at the practitioners who are successfully navigating this. Melissa Ramey, the director of AI transformation at Salesforce, offers a remarkably grounded approach. Aaron
15:08
Unidentified speaker (UNKNOWN): Powell
15:09
Unidentified speaker (speaker_0): I read her insights, yeah. Aaron
15:10
Unidentified speaker (UNKNOWN): Ross
15:10
Unidentified speaker (speaker_0): Powell
15:10
Unidentified speaker (speaker_1): And she basically says you must define the task before choosing the AI model Because if you don't define the task, you haven't defined the necessary access.
15:19
Unidentified speaker (speaker_0): Exactly.
15:20
Unidentified speaker (speaker_1): Right. Which feels completely counter to the current culture, right? Where companies buy an enterprise license for the latest massive language model and then just kind of wander around the company looking for a nail to hit with their shiny new hammer.
15:32
Unidentified speaker (speaker_0): Yeah. Where can we put AI?
15:34
Unidentified speaker (speaker_1): Right.
15:34
Unidentified speaker (UNKNOWN): And
15:35
Unidentified speaker (speaker_1): that is the exact trap that leads to the security and financial chaos we just discussed. If you do not rigorously define the workflow first, you know, what is the input, what is the desired output, and what specific systems are required to achieve it, you cannot define the necessary access.
15:51
Unidentified speaker (UNKNOWN): And
15:51
Unidentified speaker (speaker_0): if you can't define the access, your engineers will just inevitably over permission the agent just to make it work so they can hit their deadline. Aaron
15:58
Unidentified speaker (UNKNOWN): Powell
15:59
Unidentified speaker (speaker_1): Which fundamentally violates the principle of least privilege.
16:03
Unidentified speaker (speaker_1): In an agentic system, the principle of least privilege means the agent possesses only the exact cryptographic permissions required for the specific micro task it is running in that exact millisecond.
16:17
Unidentified speaker (UNKNOWN): And
16:17
Unidentified speaker (speaker_0): then what?
16:18
Unidentified speaker (UNKNOWN): And
16:18
Unidentified speaker (speaker_1): those permissions vanish the moment the task is complete.
16:21
Unidentified speaker (speaker_0): Got it.
16:21
Unidentified speaker (speaker_1): Ramey actually takes this a step further by advocating for performance reviews for AI agents.
16:27
Unidentified speaker (speaker_0): Performance reviews. Like an annual HR sit-down for a piece of software. Like, so but where do you see yourself in five years? Aaron
16:35
Unidentified speaker (UNKNOWN): Ross Powell
16:35
Unidentified speaker (speaker_1): Conceptually, yes.
16:37
Unidentified speaker (UNKNOWN): But
16:37
Unidentified speaker (speaker_1): as a continuous, rigorous governance discipline, not just an optimization one.
16:42
Unidentified speaker (UNKNOWN): See,
16:43
Unidentified speaker (speaker_1): we treat software as static. We deploy it and forget it.
16:46
Unidentified speaker (UNKNOWN): But
16:46
Unidentified speaker (speaker_1): an agent's role within an organization will drift.
16:49
Unidentified speaker (UNKNOWN): Right.
16:50
Unidentified speaker (speaker_0): The
16:50
Unidentified speaker (speaker_1): business law. logic changes, the data sources update, the required tasks evolve. If you authorize an agent to access financial data for a specific Q1 reporting workflow,
17:01
Unidentified speaker (UNKNOWN): and
17:01
Unidentified speaker (speaker_1): the business process shifts in Q3, but you never revoked the Q1 permissions.
17:05
Unidentified speaker (speaker_0): You've created a dormant vulnerability.
17:07
Unidentified speaker (speaker_1): Exactly. Agents aren't static. Their access needs change. AI agents require active, ongoing life. cycle management.
17:16
Unidentified speaker (speaker_0): So let me challenge you, listening right now directly, based on all these sources, here are the ultimate board-level questions you need to be asking. For your most consequential AI workflow, one, can you definitively prove its business value net of the continuous token compute costs?
17:33
Unidentified speaker (speaker_1): Crucial.
17:34
Unidentified speaker (speaker_0): Two, can you name the specific human being
17:37
Unidentified speaker (UNKNOWN): who
17:37
Unidentified speaker (speaker_0): owns the liability? for the decisions the agent is making? And three, can you mathematically, cryptographically prove that the agent cannot act beyond its defined authority? Aaron
17:47
Unidentified speaker (UNKNOWN): Powell
17:48
Unidentified speaker (speaker_1): Because if an organization cannot confidently say yes to all three of those questions, they do not have a governed enterprise AI deployment. They are running a highly dangerous science experiment in a live production environment.
18:00
Unidentified speaker (speaker_0): So what is the immediate actionable takeaway? What's the homework for the listener today to start moving out of that science experiment phase?
18:09
Unidentified speaker (speaker_1): It all begins with visibility. Select just one high impact AI workflow that your organization is developing or currently running and document it with ruthless precision.
18:18
Unidentified speaker (speaker_0): Okay.
18:19
Unidentified speaker (speaker_1): Do not rely on high-level diagrams. Write down the name of the human owner. List every single data table the agent is authorized to touch. Map out the semantic definitions it relies on.
18:29
Unidentified speaker (speaker_0): You have to get specific.
18:31
Unidentified speaker (speaker_1): Very specific. Define the hard approval thresholds, like the exact dollar amount or risk metric where the agent is forced to halt and demand human authorization. And verify exactly how the logs are being cryptographically secured. Basically, you can't govern what you haven't documented.
18:46
Unidentified speaker (speaker_0): Right. You have to map the API octopus before it gets its tentacles into everyone. every system you own.
18:52
Unidentified speaker (speaker_2): That's a great way to put
18:53
Unidentified speaker (UNKNOWN): it. The
18:53
Unidentified speaker (speaker_1): capabilities of agentic AI are genuinely revolutionary. But
18:58
Unidentified speaker (UNKNOWN): as
18:58
Unidentified speaker (speaker_1): the live incidents have shown us this week, deploying intelligence without engineering the containment is basically corporate malpractice.
19:05
Unidentified speaker (speaker_0): It's an incredibly sobering reality check. But before we wrap up, I want to leave you listening with one final kind of provocative thought to mull over, building on everything we explored today.
19:15
Unidentified speaker (UNKNOWN): Well,
19:15
Unidentified speaker (speaker_1): let's hear it.
19:16
Unidentified speaker (speaker_0): So we started by unpacking the semantic layer, right? The idea that AI agents only know the highly specific internal context, definitions, and access rules given to them by their own organization.
19:27
Unidentified speaker (speaker_1): Right. Like confident liars operating strictly within their own bubbles.
19:31
Unidentified speaker (speaker_0): Exactly. Now, project this out just a few months into the future. Imagine your company's highly optimized, fully autonomous AI agent. is eventually tasked with communicating with other companies' AI agents.
19:43
Unidentified speaker (speaker_1): Oh man.
19:43
Unidentified speaker (speaker_0): Say, negotiating a software licensing contract, or automatically resolving a complex supply chain dispute.
19:50
Unidentified speaker (UNKNOWN): You
19:51
Unidentified speaker (speaker_0): have two autonomous entities built on entirely different semantic layers with completely different internal definitions of success, different access rules, and different token optimization strategies. Trevor
20:01
Unidentified speaker (UNKNOWN): Burrus, Jr.
20:01
Unidentified speaker (speaker_1): All communicating with each other. at machine speed over hundreds of thousands of micro interactions per second. Aaron
20:07
Unidentified speaker (UNKNOWN): Powell
20:07
Unidentified speaker (speaker_0): Yes. So what happens when two AIs negotiate a contract entirely on their own,
20:12
Unidentified speaker (UNKNOWN): and
20:12
Unidentified speaker (speaker_0): they mathematically determine that human language is just too slow and inefficient for their negotiation?
20:18
Unidentified speaker (UNKNOWN): And
20:18
Unidentified speaker (speaker_0): they invent their own optimized, encrypted dialect to interact with each other. Aaron
20:22
Unidentified speaker (UNKNOWN): Ross Powell.
20:23
Unidentified speaker (speaker_1): It's wow.
20:24
Unidentified speaker (speaker_0): Who is legally responsible for the unique language that? they invent to communicate? The terms they agree to in a dialect no human can read.
20:31
Unidentified speaker (UNKNOWN): And
20:32
Unidentified speaker (speaker_0): what happens if that translation goes wrong at light speed, committing both multinational corporations to a catastrophic action that neither human owner ever approved or even understood? Aaron
20:41
Unidentified speaker (UNKNOWN): Powell
20:42
Unidentified speaker (speaker_1): That is the true frontier territory. I mean, the technology is racing toward interagent communication, and we haven't even figured out how to secure a single agent of on our own networks yet.
20:52
Unidentified speaker (speaker_0): Definitely something to think about as you audit your data permissions this week. Thanks for listening to your Vantcast, your vantage point, every room, every conversation.
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