Episode 375 Deep Dive: John Deeb | Governance Theater - Why Your AI Policy Is Really Just a Set of Documents

Summary

---

KBKAST Episode 375 Deep Dive: John Deeb | Governance Theater – Why Your AI Policy Is Really Just a Set of Documents

Date: July 8, 2026
Host: Carissa Breen, KBI.Media
Guest: John Deeb (Field CTO, APAC, Workato)

Episode Overview

This episode takes a strategic look at the current state of AI implementation in enterprises, focusing on the growing problem of "governance theater." John Deeb, an expert deeply involved in helping organizations take AI agents from proof of concept to production, discusses why so many AI initiatives stall and how superficial governance efforts are failing to provide real control. Carissa and John explore themes of risk, cost, observability, and the lessons enterprises need to learn as they navigate the volatile AI landscape.

---

Key Discussion Points & Insights

The AI Adoption Surge and Its Pitfalls

• Abandonment of AI Initiatives
  • In 2025, 42% of AI initiatives were abandoned, jumping sharply from 17% in 2024 ([00:00], [07:34]).
  • Causes: Governance concerns, lack of control, and observability gaps when moving AI from POC to production.
• Comparison to Previous Tech Shifts
  • Governing cloud and web transformations was challenging; with AI, the stakes are higher because it now controls both decision and action ([01:50]).

Global Divide and Organizational Attitudes

• Adoption Paces
  • US organizations are moving faster on AI than those in Australia and APAC, but the wave is coming ([03:05]).
  • AI is now a board-level topic; concern about being left behind pervades the C-suite ([04:48]).
  • Competitive pressure is a key motivator ("Kodak moment" analogy, [04:48]).

From Experimentation to Production

• Risk of Inaction vs. Action
  • The risk of not engaging with AI is felt as greater than the risks posed by AI itself ([04:13]).
• Experimentation Issues
  • Many orgs try AI, hit governance/security roadblocks, and abandon efforts ([06:23]).
  • Lack of security consideration during early experiments can nullify successful pilots ([06:23]).

Cost and "Tokenomics" Concerns

• AI Cost Spirals
  • Recent months have seen customers alarmed by rapidly increasing costs; "700% increases in cost, month on month" ([11:03]).
  • Some AI deployments are more expensive than human labor ([08:54]).
  • Example: Companies blowing their annual AI budget in a single month due to runaway token use ([10:05]).
• Efficiency Solutions
  • Need for architectural discipline: Not everything needs AI, integrate AI with deterministic (non-AI) processes to cut costs ([08:54], [12:20]).
  • Integration platforms (like Workato) can offload routine tasks from AI to reduce token spend ([14:11]).

The Centrality of Architecture & Integration

• Control and Execution Plane
  • Integration is becoming the control plane for AI, giving organizations strategic levers over decision, action, and observability ([15:03], [15:24]).
  • Vendor/ecosystem neutrality and connectivity are essential to maximize enterprise AI value ([15:24]).

Governance Theater vs. Real Control

• Superficial vs. Real Governance

  • Many orgs settle for "governance theater": steering committees, policies, and principles that look like control but lack true enforcement ([22:47]).
  • Real governance demands observability, auditing, permissions, and actionable monitoring at the level of the agent ([22:47]).
  • "A policy is really just a set of documents" without these controls.

• The Need for Observability

  • True control comes from being able to know exactly what decisions AI agents make and why, ideally with audit trails ([22:47]).
  • Many orgs lack enterprise-level observability, and the market is just starting to catch up ([30:30]).

Challenges of Scaling AI Governance

• From Humans to Agents
  • Traditional models for managing tens of thousands of human users do NOT scale to thousands of AI agents ([25:32]).
  • AI agents act differently—they will exploit any permission they're given, unlike humans who may ignore accidental access ([26:20]).
• Central Control Plane as Solution
  • To govern AI at scale, orgs need a central control plane for: identity, permissions, policies, consistent monitoring, orchestration ([26:20], [29:38]).

Accountability, Responsibility, and the Business/IT Divide

• Who Owns the AI Agents?
  • As companies scale up, it becomes increasingly unclear if tech or business units (or even vendors) are accountable for agent actions ([32:28]).
  • Ownership of agent actions, monitoring, and management remains a thorny, unresolved issue ([32:57]).

Productivity vs. Security Liability

• Point of Liability

  • AI agents become security liabilities when they can take “meaningful action without the appropriate controls” ([34:32]).
  • Unlike humans, agents are more likely to systematically use every permission they receive ([35:55]).

• Architecture Patterns

  • Hybrid approaches: let AI agents handle exceptions, use deterministic logic for routine work; keep sensitive actions gated behind human approval ([37:02]).

The Moving Target of AI Strategy

• Constant Change and Iteration
  • Rapidly shifting priorities make traditional long-term strategic planning impossible ([17:46], [21:30]).
  • Organizations need adaptive, iterative strategies. Waterfall approaches don't work; agility and real-time responses are now necessary ([20:50], [21:30]).
  • Most companies are still scrambling to even develop a coherent AI strategy ([19:15]).

---

Notable Quotes & Memorable Moments

• On AI Initiatives Failing

  “42% of AI initiatives were abandoned in 2025, up from 17% in 2024. So customers are doing more, but they're not taking you all the way through to production. And a lot of that is because of the concerns around governance, the lack of observability, the lack of control in how these agents operate in production.”
  — John Deeb [00:00], [07:34]

• On “Kodak Moments” and Board Pressure

  “I think most organizations are worried about that Kodak moment where if they don't adopt and understand the impact of AI in their business, in their market, then they're going to be left behind.”
  — John Deeb [04:48]

• On Cost Management and Efficient Deployment

  “Not everything needs to be done by AI. There's a mixture of AI decision analysis and context and processing and then the stuff you can do through traditional deterministic ways. And that balance is what helps you also manage token efficiency.”
  — John Deeb [08:54]

• On Governance Theater

  “Governance theater is about creating that appearance of control. ... The challenge with that is that creates a feeling of security that we've mitigated that risk because we have a policy. But real governance is about observability, enforcing permissions, auditing approvals, having audit trails… otherwise it's really just a set of documents.”
  — John Deeb [22:47]

• On Scaling Governance for AI Agents

  “To really monitor thousands of agents, you need to build a central control plane. ... In the human model you can monitor certain patterns and then identify an individual human ... That's not how agents will operate. You need central control to be able to scale the governance of those agents to thousands.”
  — John Deeb [26:20]

• On Security Liability

  “The moment it becomes a security liability is when it can take meaningful action without having the appropriate controls around it. ... An agent will know its capability and will leverage what it can.”
  — John Deeb [34:32], [35:55], echoed by host in closing [39:17]

---

Important Timestamps

• 00:00 — Opening data: 42% of AI initiatives abandoned (John Deeb)
• 01:50 — Higher stakes for AI than past tech shifts
• 03:05 — US vs. APAC adoption rates; C-suite/board involvement
• 04:48 — “Kodak moment” and competitive risk
• 06:23 — Orgs failing at the experimentation-to-production transition
• 07:34 — Explicit abandonment stats and reasons
• 08:54 — AI cost anxiety, tokenomics, and wasted investments
• 12:20 — Detailed example: Where AI is wasting tokens unnecessarily
• 15:24 — Integration/control plane as future of AI management
• 22:47 — Deep dive on "governance theater" vs. real governance
• 25:32 — Challenge of governing thousands of agents vs. humans
• 26:20 — Central control plane for scalable agent governance
• 30:30 — Observability gaps and the challenges of monitoring agents
• 32:28 — Ownership and accountability challenges as AI scales
• 34:32 — Transition from productivity tool to security liability
• 35:55 — Agents more likely to use all capabilities—unlike humans
• 37:02 — Human-in-the-loop for exceptions, deterministic for routine
• 37:44 — Future outlook: Excitement, risks, and the need for governance-first experimentation

---

Core Takeaways

• Abandonment is rampant: Governance failures are the biggest blocker to AI production, not technical incapacity.
• Control planes and integration layers are becoming essential for organizations that wish to scale AI safely.
• Cost is now a front-line concern, pushing organizations to use hybrid approaches and architectural best practices, not brute-force AI everywhere.
• "Governance theater" is a real risk: Documents and policies without underlying enforcement mechanisms give only the illusion of control.
• The only real governance is observable, enforceable, and auditable: If you can’t see it and act on it, it’s just paper.
• Rapid iteration, adaptability, and humility are necessary; the AI landscape is changing week to week.
• Accountability for agent actions and monitoring must be centralized, even as responsibility is distributed.
• AI agents will use all capabilities granted—configure permission prudently!

---

Closing Reflection (Carissa, paraphrased [39:17]):
"An agent will know its capability and will leverage what it can. Don’t assume an agent won’t go looking where your people never bother. If you're a CISO, this is the week to review what your agents can actually reach: because, as John put it, without that, a policy is really just a set of documents."

---

Transcript

[A]

42% of AI initiatives were abandoned in 2025, up from 17% in 2024. So customers are doing more, but they're not taking it all the way through to production. And a lot of that is because of the concerns around governance, the lack of observability, the lack of control in how these agents operate in production.

[B]

From KBI Media, I'm Carissa Breen and this is KBCast. My guest today, Deeb Field, CTO for APAC at Wakato. John spends his weeks inside the enterprises actually trying to take AI agents from proof of concept into production. We talk about why 42% of AI initiatives got abandoned last year, why so much AI governance is theater rather than control. And the question nobody has answered yet. We never managed to govern 50,000 human beings and users. How do we govern thousands of autonomous agents? If today's episode lands for you, do two things. Hit follow and send it to one person in your network who needs to hear it. That's how shows like this grow. Alrighty, let's get into it. Okay, so John, everyone is now on the hunt, as we know, to deploy AI agents. But are we repeating sort of the same mistake we made with the cloud day? And I asked that question because as we know, history repeats itself and people always want to have movers and shakers first, you know, cab off the rank, advantage, all those sort of things. But what do you sort of think about that, given your role and then obviously the ability to govern that? I mean, I know it's a big question, but start wherever your mind goes.

[A]

Yeah, so I'm lucky. I've been through a couple of these big transformative technology changes in business systems. And you're right, there's a lot that's similar to Internet Web 2.0 Cloud, but I think the stakes are much higher with AI. What was typically storage and processing capability or capacity is now becoming decision and action in enterprise systems. So I think the stakes are higher. The challenge we saw in some of those earlier technology shifts were governance was lagging, security, observability. Now the impact of that is just going to be a lot bigger when you start to create the ability for these agents to make decisions and take action in those systems.

[B]

So, John, what I really want to then focus on a little bit more perhaps is given your role and your pedigree, do you what are customers sort of saying to you day to day, what are some of their. They apprehensive about things? Are they concerned about things? I'm keen to sort of get into that a Little bit more. Because obviously what I'm hearing from folks like yourself down in Australia versus what I'm hearing from people here, it is quite different.

[A]

Yeah. I've spent a week in the US just recently and I totally agree, it's really different. The speed at which customers are adopting AI in the US is faster. However, AI is the hottest topic. It's on everyone's mind. It's transcending CIOs and CTOs. It's moving into board and CEOs in the customers that we're talking to across the spectrum. And so although there may not be as big an investment in terms of dollars in the models on a sort of company by company basis in Australia yet, we think that's around the corner. And we definitely see customers cautious about adopting AI. And this all comes down to risks, concerns around what can be done with information and data and decisions. And we're now also starting to see cost become a concern. So you're right to sort of suggest that there's a divide right now, but we definitely see the wave hitting Asia Pacific and Australia in general.

[B]

So is anything swing on the day? And they said, yes, we're concerned about the risks, but the risk of not, you know, understanding AI and implementing AI and using AI and all that sort of stuff becomes a greater risk. Would you sort of agree with that point? Because right now what we're also seeing, which you would know is like competitive advantage of if companies aren't using it, they're going to fall behind. They fall behind, they lose customers and have customers that go bankrupt. And we understand what happens then. So what would be your view then on is the risk of not doing it greater than trying to be sort of afraid of it?

[A]

I think most organizations are worried about that Kodak moment where if they don't adopt and understand the impact of AI in their business, in their market, then they're going to be left behind. And so totally, in fact, more than any other technology shift, we're seeing organizations, as I said earlier, like from the top down boards, CEOs are looking at AI to try and understand what will it do to that industry, to that business, to that, to that company, and ensuring that they don't get left behind. And I think that's created certainly a lot of experimentation. It's created a lot of initial investment in proof of concepts and value. And that's why we're starting to see that the last 12 months have been all about experimentation and now we're into that action phase where organizations are starting to Identify exactly how they're going to use AI to transform their processes.

[B]

And then would you also say, from what I'm hearing from folks like yourself in this space, they're saying in order to sort of wrap your head around it, like yes, it's one thing listening to this podcast or reading some articles, but it's another to just start, look at how models are working, start to implement it into your organization. Do it subtly though. You don't have to do a whole like overhaul straight away, but you just got to sort of get amongst it. Do you think that people aren't really doing that? They're trying to understand stuff from afar. Whilst that makes sense, you're not sort of understanding the full fidelity of this capability.

[A]

There is definitely that. There's the, the easy stuff now is that some of that personal productivity driving efficiency, whether it's helping create documents or build out reports or analyze data and you know, there's statistics going around. One says that only 6% of orgs fully trust AI agents to manage end to end processes. The 10 times 100 times improvement that AI's promising hasn't hit mainstream yet. And that's because everyone is still experimenting. And part of the challenge with the experimentation is from a security standpoint that's not being considered. So when an experiment is ready to go into production, we need to go back and revisit governance security. And that either slows it down or actually makes it null and void. So the experiments need to start focusing in on that governance and security early. Otherwise even a successful outcome doesn't necessarily mean that goes into production.

[B]

So when you said it's like null and void, so do you think people would say, well we tried it, it doesn't work, it's not for us, so we're just going to like abandon ship. Is that sort of some of the theory behind it?

[A]

Yeah. 42% of AI initiatives were abandoned in 2025, up from 17% in 2024. So customers are doing more, but they're not taking you all the way through to production. And a lot of that is because of the concerns around governance, the lack of observability, the lack of control in how these agents operate in production.

[B]

Okay, that's quite a lot then. So especially the jumper. Then also 42% of abandoning AI because then there's even people in the industry saying, well companies are now. I mean this is slightly off track, but I think it's important to. What we are talking about is people are now saying to me, kb, we are even questioning the ROI now on AI. And I've even sat down with the co founders of Datadog in New York City like two weeks ago, and they were saying, now in some instances, AI is becoming even more expensive than an actual physical worker. So do you think those sort of elements are being looked at in terms of like strategic board level folks to say, does this make sense for our business? Yes, we need to leverage it. But perhaps there's different ways, different paths that we can go about it rather than trying to immerse ourselves in it. This is just very recent sort of conversations I'm having.

[A]

Yeah, that, that actually aligns with what we're seeing as well. Only recently have customers started to get very nervous about cost. And that's partly because some of the frontier labs have changed the way they price. But also people are using AI in places that maybe they don't need to, and therefore the way they use AI can be quite cost inefficient. And so that's one of the things that we talk to customers about, is looking at how you use AI. Where is it best placed? You know, at Workato, for example, we, we are all about integration and connecting to systems and then enabling AI to leverage that connectivity layer. And our point of view is that AI doesn't need to determine how to map a purchase order from one system to another every single time. You can do that deterministically. That saves on token cost. So not everything needs to be done by AI. There's a mixture of AI decision analysis and context and processing and then the stuff you can do through traditional deterministic ways. And that balance is what helps you also manage token efficiency.

[B]

Okay, I want to get into this a little bit more. So are you. You've probably seen online that there are companies out there, I think here in the US that they blew their entire AI budget in like one month in terms of like tokenomics. Obviously it's spiraled out of control. So what I'm hearing, and maybe you agree, people get real nervous about what's the bill going to look like at the end of the month? Because people are just going above and beyond and going a little bit ham in terms of trying to automate everything, for example. So do you think part of that 42% is definitely related to the economics around how much it costs as well as people are spiraling out of control with this? It is, yes. It's somewhat hard to control people internally as well, because that's how we have shadow AI and all these other things that we're, we're seeing. So do you think that is a huge factor? And then if so, do you think that now the pendulum will start to swing back in the middle because people have maybe over indexed on AI, for example.

[A]

So I don't think the 42% was because of the tokenomics and cost because that cost element is quite recent. And to your point, we're starting to hear a lot of stories about, you know, 700% increases in cost, month, month on month, et cetera. So I think that impact will, we'll start to see that which will then to your earlier point, that will create even higher expectations of AI ROI before anyone starts to look to deploy into production. But I do want to come back to your point. Using AI for everything and not considering best practice around the right architecture, we think that would become a bigger component of deploying architectures and agents on the right architecture. And that's something that we definitely see now with both cost, but also security and governance. So with the right architecture you cover, you cover all those elements in an agent deployment.

[B]

And then you mentioned before, John, that companies are leveraging AI in places they don't need to. So, and I know you just mentioned it, but what is that? Can you give me an example? Or what is that? Is it sort of a common theme that you're seeing across businesses? Or what do you think perhaps people are sort of, perhaps wasting tokens on that they shouldn't be?

[A]

Yes. So you think about the way you process tokens and consume them. A lot of it has to do with the amount of data that you're passing through to the model and what you're asking the model to do. And so if for example, you're processing thousands of orders a day and you want the model to help take that order from one core system to another and make a decision on potentially which systems to update or there's definitely a role for AI in how you process end to end processes in the organization. What we're saying is at some point AI doesn't need to reinvent the, for example, the mapping of that data and it could be hundreds of fields to the target system, but instead could leverage MCP and protocols that can perform that action. And what AI is doing is determining which action to take. And then in our case, for example with Wakato, the integration layer performs the action without having to consume tokens. And that's the balance, that's getting the architecture right, getting the best practices and patterns right so that you can use AI in the places where it's going to be advantageous but continue to leverage that sort of more basic processing power for the components that don't need the high powered AI reasoning and processing.

[B]

Yeah, because I know that Wakato, you guys really specialize in obviously integrating the MC space. Do you think customers don't know? Okay, well maybe we can really reduce our tokens because we're leveraging MCP to do it in terms of doing this. But again everything takes time and you need the right people to understand what it looks like. So is that sort of a concern at the moment, would you say?

[A]

Yes, and actually it initially was. The concern was more on the security and governance aspect of things. And it just so happens that by applying that architecture you also get cost gain. So again, thinking back to you're going to automate your entire order to cash process, you're going to leverage AI to make decisions throughout that process, but you'll leverage MCP to perform actions on core enterprise systems. And that model allows you to govern, audit all the changes that you're making to your underlying systems through that process and continue to take advantage of where AI is strong, but leverage these, this control and action planning to actually perform the functions in a deterministic way.

[B]

And John, you also argue then that integration is coming the new to your point before control plane for AI. So does that mean whoever controls the integration, for example, effectively controls the future of like enterprise AI and like how does that sort of look in your eyes in a little bit more detail.

[A]

So we definitely believe that the control plane is tied to the execution plane as well and connectivity is at the heart of that. A platform that can connect to your existing systems, applications, processes, et cetera, is at an advantage in being able to expose the broader capability of the organization to enterprise AI. So that control and execution plan is that combination of being able to take different ecosystems from all over different parts of business and industry to bring together. And we think a part of that is vendor neutrality or ecosystem neutrality. And that's something we strongly believe organizations have invested in major technology ecosystems and we offer that bridging that control and execution plane that connects their enterprise AI, their models, their agents to all those backend systems. So yeah, I think to maximize value on enterprise AI, you need that control and execution plan to deliver that.

[B]

And do you think each week or each couple of weeks things change because in terms of priorities or what's important because was the, you need to implement guardrails and now it's about harnesses and then on the loop, in the loop, all the stuff that everyone's been Talking about, like, each time I'm talking to someone like you, the conversation just seems to shift and it shifted quite significantly, even week by week now. So is it just going to be people just trying to keep their head above the water, see what works, trying to understand where they've got to go next? Because depending on who you ask, who you listen to, it's very varying opinions out there. So there's not sort of one, one size fits all. And I know historically that hasn't been the case. It's just that it's very sporadic now in people's views. So with your customers or people that you're speaking to out in the field, would you say that their concerns and what they are worried about changes a lot? Like, because everything's changing so quickly anyway. But how are people starting to prioritize perhaps which way and how they navigate this moving forward?

[A]

Yeah, that is, I think, one of the biggest challenges for a cio, a cto, even a CEO today is just how much things are changing and how much noise is out there. And all that noise is about what they should be doing next. And that's quite. That's really overwhelming. And I think part of the conversations we're having is to really, and as I mentioned earlier, to really take a step back and look at what is the overall objective of an organization's AI strategy, how that then drives policy and governance. And one of the things that we continually stress is about trusted action. So if we ask a lot of leaders today, what's their biggest concern is trusting the model, trusting the people using the model. These are the things that we need to address now. Otherwise, regardless of the solutions we're acquiring or building, without that trust, there's going to be a real hesitancy and concern around taking things into production. But to your point, I think every week there's a new topic. I think tokenomics is recent costs and now become a little. A little more concerning than what they were a month ago. And that will, I think, continue to grow. But at the foundation of it is trust, which means having the right governance and control and execution plan in place.

[B]

So going back to your comment around AI strategy, do you think companies really have one or do you think that now they're like, well, we better get one, because if we don't, like, we may not have a very strong business at the end of this. So do you think people are scrambling to sort of even get a strategy?

[A]

Yes, I think there's a real. There's a real desire to have clarity at a Time when there's so much uncertainty with what this technology can actually deliver. I think we've seen organizations focus more on policies that drive governance and use. So again, going back to wanting to have control over how AI is being used in an organization, I think that's definitely an area that a lot of customers are investing time and effort and building out proof of concepts. But to think of them being able to define a strategy and really lock it down, there is just too much change going on right now for that. Hence it's really going back to those building blocks of observability. Can I see what people are doing regarding AI control? Can I manage what they have access to, what data they can share in the model, what actions they can perform in the applications and then execution? Can I connect my backend system so that AI can leverage them so they're the fundamentals that we're working with customers on? I don't think anyone can be sure exactly what's going to happen next. And therefore defining that strategy has to be somewhat in real time.

[B]

And do you sort of see it as a bit more iterative? As we mentioned already, like things do change week to week in terms of priorities? What's come out like people blowing their budgets? That wasn't a topic last month. Now it is. Do you think that people are just going to have to be a little bit more adaptive to, hey, it might be this week, but next week we change. So it's not going to be the whole like back in the day we used to do waterfall approaches and we had the static project for two years because nothing changed much, but now it does. It's more aggressive and volatile than ever before. So how do you think that's sort of sitting with companies that are perhaps a little bit more traditional in large corporations that they would take a long time to make a decision and now they may have to make decision a lot faster.

[A]

It's very challenging for those companies. I think that's part of the change. The uncertainty and what it's created in the market is a desire to try and stay ahead of the curve when it comes to the progress being made by or these technology companies. So I agree. I think there's a. You cannot burn anything in stone right now in regards to a strategy because things are changing so quickly.

[B]

We'll come back to that after a quick word from our sponsor. If you're working in AI, machine learning or data science, you're likely already handling sensitive information proving your security and compliance posture. That's Where Vanta comes in in Vanta helps AI driven teams fast track compliance. Think SOC2 ISO 27001 GDPR with minimal disruption to development. Visit vanta.com kbcast V-A-N-T-A.com kbcast to learn more. The other thing I'm hearing about in the space, and maybe you know a little bit more about this than me, what's your sort of view on like AI governance theater? Apparently it's people that are looking like they're doing governance, but in reality that they're not. But it looks that way.

[A]

Yeah, I think it's kind of human nature to, to want to think you're in control. I think governance theater is about creating that appearance of control. There's steering committees, there's, you know, AI principles and policies, everything trying to reduce any kind of risk that will create either reputational risk or exposed data that shouldn't be exposed, et cetera. The challenge with that is that creates a feeling of security that we've mitigated that risk because we have a policy. But real governance is about observability, enforcing permissions, auditing approvals, having audit trails, being able to monitor at the point where and agent acts, because without that then it's really just a set of documents and organizations will need at some point to be able to go back in time and say this is exactly what happened at this point in time and this is why that decision was made. Here's the reasoning of the model, et cetera. So I think the AI governance theater is maybe there to provide some feeling of confidence, but the reality is you need those underlying capabilities to really deliver on governance.

[B]

And would you also say, because this concept area is new and it's changing all the time, is it hard for people to confidently say, in terms of a governance strategy, yes, this is what we should do, or this is what we should do right now. But things may change because, for example, the mythos thing, I was speaking to executive down in Las Vegas at an event recently and they were like, look, because of this, it's completely thrown everyone off again. So do you think that companies will have to just be comfortable with this is the strategy right now and our governance approach. But that may change next week because who knows what's going to happen? Who knows what's going to come down the turnpike, for example?

[A]

Yeah, there needs to be a level of flexibility built into any of those policies, I think to that point as well, for AI enterprise, AI to be impactful and to deliver that kind of impact, it's promising to that's going to expose it to more challenges in security and governance. It means giving AI capability. And you know, it's a different model. You can govern humans. The way you govern humans is very different to the way you will govern autonomous agents. And I think people are discovering that and therefore coming back to those fundamentals of what I mentioned earlier. Audit trails, permissions, approvals, observability, they're the things that will really drive towards building confidence at the when you're monitoring hundreds of autonomous agents in an organization.

[B]

Okay, so then on that note, most companies then comfortably explain what the human users are doing across their environment. Like even traditionally speaking, I mean, I've worked in enterprise, you've got 50,000 people. It's very hard to control what each of these people are doing. But then how can we realistically expect to then govern like thousands of autonomous AI driven actions Then, for example, and I know people are saying, oh, we use AI to then govern the AI and there's all these sort of theories, but if we can't even do it back in physical like human identities doing this stuff, and now we've just gonna let loose with agents are doing who knows what, haven't we just poured more fuel on the fire because we were struggling with just the human side of it and now we've sort of increased it by thousands of AI agents, for example.

[A]

Yeah, and that's what I was alluding to earlier about the way you monitor human humans is quite different to the way you will monitor agents. To really monitor thousands of agents, you need to build a central control plane. And that's because you need to manage identity, permissions, policies, all your monitoring capabilities need to be enforced consistently regardless of how many agents are running. I think in the human model you can monitor certain patterns and then identify an individual human that might be performing things differently. And then you can monitor them specifically to understand what they are doing. That's not how agents will operate. You need central control to be able to scale the governance of those agents to thousands. And part of that too is connectivity, centralizing the orchestration of those agents. Otherwise you'll end up with AI sprawl or these disconnected agents with inconsistent permissions operating outside of the security visibility. So a central control plane is critical to setting up a foundation to scale the number of agents operating in an organization.

[B]

And then what's your view on what I'm hearing as well? That for example, if you're like, okay, agents want to do things super efficiently, get that? Then apparently sometimes if there's got multiple Agents like, they'll even overpower another agent to be like, well, I've got to get there first. And then that causes other issues. Like this is now what I'm hearing again. So how a company sort of gone about that, for example, because there could be multiple teams that are trying to attack the same problem, but looking at it very differently. And therefore another agent overrides another agent, then becomes a bigger problem than where they started.

[A]

I'll go back to what I said earlier, which is unless you can monitor it and capture it, then you're not going to know these things are happening. And so observability is key. And this control plan concept and control and execution plan are critical because they both tie into providing the capability of the agent. So for example, you want to create an agent that is taking care of customer inquiries on the website, and one of those requires updating multiple backend systems. Well, you want to ensure that that agent has the right level of security to make changes to those systems. And what we found is that's actually been a cause of concern where an agent is given too much access and can perform maybe a deletion where they shouldn't be able to. And that's because the underlying control and execution plane doesn't have the correct security policy and security capabilities to enforce credentials that are only relevant to that agent. And so there are examples where organizations aren't getting the right security visibility into the agent and therefore not the right level of commission. So I'd come back to that central control plan, an execution plan where you can control exactly what an agent can do and that limits then, you know, agents overstepping their mark and overstepping the boundary of what they're, what they're meant to be able to perform.

[B]

So then coming back to the observability piece, so obviously most companies out there or vendors will say, like, you know, you can't secure stuff without, you know, understanding it. Like from an observability point of view. Right. Which is pretty much the undertone of what you discussed here. Doesn't that seem very rudimentary though? And so how many customers aren't doing that? Because then how can you start to deploy agents? Because then you don't know what they could be doing because you can't understand what they're and you can't see it. So would you say that people are a little bit more mature in their approach by saying, well, we've got the observability, we understand at a fidelity what's happening. And they were comfortable now to start to deploy these agents because we can rein them back, because we've got the visibility over it. Say there are companies out there that haven't quite understood that and then they've just started to, well, let's start the whole AI agents thing and let's see how we go and we'll try to clean the mess up afterwards.

[A]

What we're seeing and part of the, I guess, value proposition of a company like Wakato is providing that central control and execution plane, which is a, from a technology standpoint, is catching up to the capabilities of the models. So I don't think customers have done that because the technology is behind the model capability and the agent capability so they can monitor individual agents. But thinking about agents across multiple ecosystems, different technologies, that's the point we're making around a central control plane. And that's a relatively new area where organizations are starting to look at not only how do I manage my proof of concept agent that I'm building, but what happens when I start to deploy hundreds of agents across the organization. So I do think there's a more immediate need of an individual monitoring which is going to turn into an enterprise level control plane requirement for organizations. I hope that answers your question.

[B]

So what you're saying is at the moment, in terms of like enterprises, you said before, central control plane, it's relatively new concept for some of these people. In terms of the capability in which you're offering, for example, are people still trying to understand how this looks, what, how it works, what it does for their business? Would you say so there's still like a learning element there is what I'm hearing.

[A]

Yeah, I think they're still working out, what do I need to feel comfortable and trust these agents? How's that deployed? How's that governed, who owns it? I think there's all these questions that are relatively new to organizations that are thinking about mass deployment of agents across their processes. And so to, you know, your earlier point around just how much things are changing, you know, this is an area that we think it's going to become a higher priority as organizations start to see benefit in agent deployments.

[B]

And then I guess the other side of that would be accountability, responsibility. There's conversations now, people saying, well, who owns it? Is it the tech side that owns it or is it the actual business side? Because it's their agent, it's doing stuff in their area. Is it the vendor that owns it now? Do you think that is going to create a problem for people? Because at the end of the day if something goes wrong, everyone's. Everyone A, blames a vendor easily because no one wants to blame themselves, or B, it's very easy to start to blame someone else rather than yourself. So is that, do you see that becoming a future problem?

[A]

We see organizations struggling with that. I think the traditional lines of IT and business, we've always talked about bringing that together and the impact it has on business as, as a partner. But yeah, I think organizations haven't had to deal with that yet because like we said earlier, the streamlined and mass deployment of agents and organizations hasn't occurred. It's not happening today. But that is. Is definitely a challenge that organizations will need to address. What does, who owns the agent? Who owns the agent? Actions, monitoring, management. And we've got to consider it back to the earlier point is how does that fit into a central plane? So you might be able to distribute the accountability of the agent, but you can't distribute the monitoring and management of it itself. So there's some things for organizations to navigate some of these challenges as they start to deploy these things on mass.

[B]

And so then I'm curious then to understand at what point then does a AI agent stop becoming a productivity tool and then start to become more of a security liability? Because sort of everything that we sort of discuss like, yes, it's productivity, but there's still a lot of liability there. I know it's still early days and we're figuring it out, but do you think that we're kind of at this stage now where people are trying to understand not necessarily the value, but more like where does it sit? Because it is, from what I'm hearing, people like yourself, there is some hindrance there for these companies.

[A]

And the challenge is the more you give your agent in terms of its capability, the probably the bigger security liability it will be if you're not managing and controlling it properly. I'd say that the moment it becomes a security liability is when it can take meaningful action without having the appropriate controls around it. And so just like a new employee, when a new employee joins, you wouldn't give them access to systems that they don't need. And so the difference there is that an AI agent is going to be more aware of what it can and can't do in terms of permissions, because it's obviously approaching it in a very different way to a human. Therefore accidentally giving it access to a system that it didn't need is going to become a lot more prevalent than it may be if you'd done that to a human. So Once you start to give agents capability and allow it to take meaningful action in places it shouldn't without the appropriate controls around it, that's when it becomes a true security liability.

[B]

So when you say places it shouldn't, is that by accident? Is it on purpose? Is it with the right intention that, oh, well, we think this AI agent's going to do the right thing? And, you know, I can put my feet up a little bit earlier on a Friday because the machine's doing the work for me. Where does that theory come from, would you say?

[A]

I think if you consider the way that we've been doing things in the past, you know, we, as I mentioned earlier, a new hire might be given access to a system that they don't need access for and they may never know it and they may never try to access it. Whereas with an agent, you won't get away with that because an agent will know its capability and will leverage what it can. And so I think there's a level of detail and again, bringing in the right, potentially bringing in a human in the loop when it comes to certain actions that are high risk for. But understanding and analyzing what capability, what meaningful actions an agent can perform is going to be critical. We're not going to get away with providing too much access and not having to worry about someone using it when it comes to agents.

[B]

And they also extend to your points around it being deterministic. So if there's thousands of things where it's like, okay, that's fine, it's fine, I've seen it before. Yep, it's all good to go ahead. Oh, no, there's something I haven't seen. Therefore John gets notified. John says, yep, it's all good. Thanks for notifying me.

[A]

Yeah, that that pattern is, I think we're going to see a lot of that, that kind of bringing in a human on exception. But there's a reality to, again, making sure that the AI only processes it what it needs to. And some of that deterministic happens outside of AI because it's not needed. And so getting that balance right was probably the point I was making earlier around getting the right architecture.

[B]

So, John, sort of final question, what do you think then moving forward? And I know we've sort of discussed it, one week it's that, and next week it's something else. But you're obviously out in the field, you're talking to people and customers. What do you sort of sit back on the, I don't know, Monday morning or Friday afternoon and think about where's the technology space going to go moving forward.

[A]

It's incredibly exciting. Like I mentioned earlier, we've seen a number of transformational changes in technology over the last 40, 50 years, but this one has the potential to, as I mentioned earlier, drive decision and action in these autonomous agents that can really transform and change the way we interact with businesses through that technology landscape. And so I think the gap between what it could potentially do and what it can do today is still really big. And it's one of those things where organizations do need to get exposure to the challenges and risks and concerns. They need their governance capabilities to be ahead of their technology deployments of these agents. Otherwise they will put themselves at risk. And so I think education, testing things out, doing these proof of concepts, but in factoring in security governance in those early is going to be an advantage for organizations as opposed to building this amazing agent and then having to go back and figure out how to secure it. And so that experimentation which is now moving into serious deployments, it's going to be easier when you're factoring in security and governance early. And that's what I think a lot of organizations are starting to do. And now, as you mentioned, thinking about cost as well as an element of that.

[B]

That was John Deeb, everybody. The line I can't shake from that conversation is that an agent will know its capability and will leverage what it can. All those systems your people have access to but never touch. Don't assume it agent won't go looking for your sizo listening to this. Make this week the week you review what your agents can actually reach. Because as John put it, without that a policy is really just a set of documents. I read every reply. If you've got some thoughts on this one, send me a message on LinkedIn. Kb castle cyber for the c suite.