Latest / Master Your AI Future | AI Executive / Why Is the Human-Agent Operating Model Becoming the Next AI Advantage?
Transcript
- 📍 📍 📍 Welcome to the debate Right now, uh, there are enterprise systems where a piece of software is, is independently negotiating vendor contracts or approving expenses. Yeah, or routing highly sensitive customer data. Exactly. But if that autonomous system makes a million-dollar mistake today, there is a very good chance the company's organizational chart has absolutely no idea who to fire. Oh, completely. We are crossing a massive frontier in business here. We're moving away from AI as just, you know, a personal productivity tool. It's not just a fancy auto-complete that helps you write an email faster anymore. Right. We are stepping directly into the execution layer of the enterprise. And this shift is creating an entirely new discipline, which is called the human agent operating model. What's fascinating about this moment is that the technology itself is no longer the main bottleneck. I mean, we have the cognitive models. The models are there, yeah. Right. The real constraint now is organizational design. When you look at the research across the major consulting firms, it reveals this staggering divide. Executives who are treating AI as just a new flavor of automation are seeing, well, pretty minimal returns. But those who are fundamentally redesigning their operating models around these systems, they are capturing massive enterprise value. Which brings us to the core tension we're examining today. As we restructure these companies for autonomous execution, how do we actually govern the AI? I look at this shift, and I think we have absolutely no choice but to treat these agents as discrete digital workers. You have to give them an explicit identity, a specific set of system permissions, and a designated human manager. And see, I look at that exact same shift, and I see a massive trap. A trap? Yes. Treating an AI agent like a worker or a digital employee is a fundamental category error. When you do that, you inevitably force a next-generation capability into a legacy org chart. We shouldn't be assigning job titles to algorithms. We need to treat them purely as cognitive capabilities, skills like analysis, synthesis, execution, that are dynamically woven into horizontal value streams. Well, let's unpack the mechanics of that because the stakes are incredibly high. If an AI system is going to execute a real workflow, say approving a commercial loan or processing supply chain invoices, it is generating real operational risk in real time. Sure, risk is inherent. Right, and governance has to scale linearly with autonomy. You cannot audit an abstract capability floating somewhere in the cloud. You can only audit an entity. That means every single agent requires a specific digital identity. I don't know if I agree with that premise. Well, think about it mechanically. It needs a login, it needs an auditable boundary of action, and crucially, an accountable human business owner whose name sits right next to it in the directory. If you just diffuse AI into the ether of a workflow without isolating it as a governable actor, you destroy the entire chain of accountability. See, to me, putting a human manager in charge of an AI agent is like forcing a supercomputer to print out its calculations so an accountant can manually check them with an abacus. An abacus? Come on. It is. You're tethering a gigabit technology to a human bottleneck. When you anthropomorphize an AI, when you give it a title, a manager, and treat it like a human worker, you end up doing exactly what the research warns against. You just layer new technology onto fragmented legacy processes. But- Instead of creating artificial reporting lines for bots, organizations need to map out the entire end-to-end value stream and then allocate work dynamically based strictly on proficiency and cost. Look, I get the theory, but let's look at what happens when things actually break in the real world. The market is facing a severe crisis of trust right now. There's a staggering stat from recent enterprise surveys from Deloitte. Two-thirds of senior managers expect multi-agent AI to materially reshape their work within just two to three years. Yeah, but only about a quarter of them feel prepared. Exactly. And that readiness gap isn't because they lack the cloud infrastructure. It's because people fundamentally do not trust systems they cannot identify. This brings up PwC's concept of staged autonomy. Right. You build trust by explicitly defining an agent's identity, its owner, and its escalation rules. If you're listening to this and you work in middle management, you're probably thinking, "Great, I'm gonna have to supervise a fleet of bots that hallucinate half the time." Which is exactly why you shouldn't supervise them like employees. But that's exactly why treating an agent without a defined identity is so dangerous. It's like giving an anonymous contractor a corporate credit card. You would never do it in the physical world, so why on earth would we do it in the digital world? I come at it from a completely different way. The assumption that slapping a login ID onto a piece of software somehow solves your governance problem is, well, it's fundamentally flawed. True governance shouldn't be attached to an anthropomorphized bot. Then what is it attached to? It must be embedded deeply in the runtime operation of the workflow itself. It's what EY Parthenon calls an agentic-first workflow. You don't manage the agent, you manage the process. You build what's known as purposeful friction directly into the architecture. Wait, I need you to break down what you mean by purposeful friction, because to me, that just sounds like a fancy corporate rebrand for a human bottleneck. How is that actually different from a manager checking an employee's work? It's a great question because the distinction is everything. A human bottleneck implies a manager reviewing every action a worker takes just because they are the supervisor. Okay. Purposeful friction means you deliberately engineer human review only at points of genuine ambiguity, material risk, or consequential trade-offs. Give me an example of how that works in practice. Okay, let's say an AI is processing vendor discounts and procurement. The system processes 1,000 routine 5% discounts automatically. No human ever sees them. Right, basic automation. Exactly. But the workflow is architected so that if the AI calculates a discount over 15%, or if it flags a massive anomaly in the vendor's history, the workflow automatically routes that specific decision to a human risk officer. Mm-hmm. You don't require human approval for the mundane steps, and you certainly don't need a human acting as a full-time supervisor to a digital worker. Placing human intervention at structurally defined points of risk is infinitely more efficient Okay, I understand the theory of managing by exception, but I'm just not convinced that embedded workflow rules can perfectly anticipate every edge case. They don't have to perfectly anticipate them. They just route the exceptions. But business operations are inherently messy. But- When an AI agent encounters a situation entirely outside its parameters, like a true black swan event, it needs to escalate. If that agent isn't a discrete entity with a designated human owner, who does it escalate to? A generic workflow can't pick up the phone and negotiate with an angry supplier. But the syst- We treat identity, permissions, and ownership as core operating requirements precisely because things go wrong. If you rely solely on runtime governance without entity-level accountability, you create an environment where everyone is supposedly responsible, which usually means absolutely no one is. Look, the moment you assign a specific human owner to an agent, you inherently lock that agent inside a departmental silo, and this gets to the absolute core of where these systems actually generate returns. I don't think you're locking them in. You are, structurally. If governance dictates how agents operate, workflow architecture dictates where they create enterprise value, and the data shows that seventy to seventy-five percent of all potential value is trapped between organizational silos. Right. It's trapped in the handoffs. The coordination delays, the emails waiting for approvals across different departments. Precisely. So op- Think about a massive automotive operations transformation. Researchers identified over two point four billion dollars of digitally enabled value potential in one case. Wow. Yeah, but here is the critical mechanism. Nearly two billion dollars of that value wasn't found in making one specific department faster. It was associated with enterprise value streams, meaning horizontal cross-functional processes like going from an engineering incident directly to a manufacturing resolution. So how does the AI actually unlock that without an owner orchestrating it? Under the old model, procurement spots a parts issue, they email engineering, engineering runs a CAD analysis, they send it back to finance for cost approval, and you lose three weeks. If you treat AI agents as employees reporting to human managers, you just put a slightly faster bot in procurement and another bot in engineering. You're just paving the cow path. You reinforce those vertical silos. Exactly. But if you treat AI as a cognitive capability, it can instantly analyze the supply chain disruption, run the engineering variants, and surface the financial impact across all three departments simultaneously. True transformation requires orchestrating these capabilities horizontally, completely ignoring the traditional org chart. I see why you think that, but let me give you a different perspective. That horizontal agility sounds beautiful in a supply chain context, but what happens in highly regulated environments? Regulation doesn't negate the need for horizontal flow. In banking or healthcare, how do you manage exception risk across those silos without a designated agent and a designated human holding the bag? Let's look at Boston Consulting Group's banking transformation targeting thirty to fifty percent workflow automation. Which is massive. We're talking about releasing millions of hours of human capacity. But in a bank, you're dealing with strict capital requirements, regulatory compliance, and immense risk. You cannot simply let a cognitive capability float horizontally across the order-to-cash process without discrete, auditable nodes of accountability. But building for the exception path shouldn't limit the entire system. If the system incorrectly flags a hundred legitimate transactions for money laundering, the regulators do not want to hear about your beautiful horizontal value streams. They demand an auditable trail of exactly who executed the action and who authorized the parameters. A discrete digital worker model provides that. It creates a traceable identity that moves through the systems with specific access rights, just like a human. I completely understand the regulatory anxiety. I do, but treating it like a human is exactly what limits the economic return. The research explicitly notes that most companies fail at this because they just layer AI onto these legacy, highly defensive workflows. It's not defensive; it's compliant. Rebuilding the model means you stop thinking in terms of a vertical hierarchy that requires one human to, quote, "hold the bag" for one agent. In an agentic-first workflow, the accountability's built into the outcome and the exception path itself, regardless of which department the data touches. If you force a bank's AI to operate identically to a human bank teller, checking one box at a time and passing it to a human manager, you limit the technology's capacity to the exact speed of the old silo. And I'm definitely not suggesting we limit the speed. I'm suggesting we limit the blast radius of a failure. When an agent spans multiple functions, say it takes customer intake in retail banking and connects it directly to credit approval and risk management, the regulatory liability doesn't magically disappear just because you've mapped out a new workflow. No, but the liability's managed at the workflow level. And this transition logically brings us to the CFO's perspective. How do you track the cost of this horizontal execution? If you insist on treating agents as fluid capabilities rather than individual digital workers, you create a massive financial blind spot. How is it a blind spot if you are measuring the total outcome? Because you lose the ability to apply financial discipline. Let's look at the framework of AI FinOps. When you deploy an agent today, the cost isn't just a flat software license. You have token costs for every prompt, cloud compute infrastructure, API integration costs, governance overhead, and the cost of handling exceptions. Sure. Those are the inputs. And you have to apply continuous, rigorous discipline to those inputs. You measure the cost of the agent, you measure its utilization, and then you decide whether to scale it or shut it down. This self-funding investment engine is only mathematically possible if you isolate the agent as an operational unit. If you smear the AI across a horizontal workflow as a capability, you completely lose the ability to track the true cost stack of the execution I'm sorry, but I, I just don't buy that. Uh, uh, uh, let me tell you why. Tracking an individual agent's cost stack as if it's like a discrete employee, it is a complete illusion. It distracts leadership from actual enterprise value Distracts them? Let's talk about the, the, the concept of, quote-unquote, time saved, right? There's a brilliant maxim in the EY research: time saved is only potential value. Now, think about it mechanically. If your highly auditable, discrete, quote-unquote, digital worker saves a mid-level manager ten hours a week, but that manager just spends those ten hours sitting in more internal meetings Well, obviously, you have to manage the human's time better the enterprise has captured absolutely zero financial value. In fact, you've lost money because you're paying a token cost for the AI. The correct metric is not the individual agent's return, it is the monetized outcome of the entire workflow But how do you accurately measure the efficiency of a workflow without measuring the specific agents driving it? You measure cycle time. You measure the straight-through processing rate. Let's unpack that term. Straight-through processing means how often does a complex transaction, like a mortgage approval, go from the initial customer application all the way to final funding without a human being ever having to intervene Right. That's the holy grail of automation Exactly. You measure the cost per outcome. Let's look back at that banking case you brought up from BCG. They projected a massive one hundred and fifty percent five-year ROI, and a full payback within two years Yes You do not get a one hundred and fifty percent ROI by meticulously tracking the API token cost of an individual chatbot. You get it because you redesigned the entire mortgage workflow to require fewer human handoffs, and then you capture that release capacity. You either reduce headcount or you explicitly redeploy those employees to revenue-generating activities. Sustainable returns depend on converting productivity into actual growth at the systemic level, not the agent level I don't disagree that ultimate enterprise value requires capturing that freed up capacity. You absolutely must capture the productivity savings and reinvest them selectively. But my argument is about the baseline prerequisite for that measurement Which is what? If you cannot isolate the execution cost of the autonomous action from the human action, your cost per outcome metric is a complete black box. Think about it mechanically. If an agent hallucinates or gets caught in a logic loop running thousands of API calls overnight, racking up massive infrastructure costs, how does the CFO audit that if the agent isn't a discrete entity? Well- AI FinOps requires granularity. You need to know exactly which agent generated the runaway cost, who owns it, and whether its specific output justifies its daily compute expense. But the compute expense is just one tiny input into the broader value stream. If you hyperfocus on the agent's individual P&L, you risk shutting down a capability that might have high token costs but drastically reduces downstream human rework. But you still need to know it costs that much. Sure, but if the incidence of resolution time for a critical supply chain failure drops by forty percent, the enterprise wins massively. It doesn't matter whether a discrete digital worker or a woven fabric of cognitive capabilities did the heavy lifting. The focus has to remain on the horizontal output. Well, this fundamental tension is exactly why this debate is so critical for executives right now. We are really talking about two totally different ontologies for the future of the enterprise. From my perspective, the human agent operating model absolutely requires structured, auditable digital identities. Mm-hmm. As AI takes on more complex execution, autonomy naturally generates risk. By treating agents as discrete digital workers with explicit permissions and accountable human owners, we ensure that governance and financial discipline can scale safely right alongside that autonomy. And I maintain that viewing agents through the legacy lens of employees is just gonna replicate existing bureaucracy and code. To capture the massive value that is currently trapped between organizational silos, we have to treat agents as cognitive capabilities. By dynamically weaving these capabilities into horizontal agentic first value streams, organizations can actually redesign work for outcomes rather than just adding artificial digital bodies to an outdated org chart. You know, despite our completely different approaches to governance and measurement, there is a profound area of convergence here. We both entirely agree on the fundamental thesis of the data. Deploying advanced AI without fundamentally redesigning your operating model is destined to fail. Oh, absolutely. The technology alone will never save a broken process, and the current readiness data should be a massive wake-up call for everyone. If you look at recent enterprise research from Accenture, it highlights a truly shocking reality. They found that only about three percent of surveyed UK organizations are actually fully ready for advanced agentic AI. It really is staggering. It is. And this is despite their estimation that a staggering eighty-two percent of working hours could potentially be enhanced by this technology. The gap is huge because only a fraction of employees have actually seen their major processes redesigned around these tools. That gap between the 82% potential and the 3% readiness is exactly where enterprise value goes to die. And there's another point of complete agreement between us, the shifting nature of human value. Right. As routine execution, data coordination, and basic synthesis are automated, whether you call them digital workers or cognitive capabilities, the distinctly human contributions become the organization's most valuable scarce assets. Yes. Things like complex judgment, genuine empathy, high-stakes relationship building, and managing that purposeful friction we talked about. That becomes the core of human work. The goal isn't just to automate tasks, it's to elevate human capability to the points where it materially changes the risk profile or the business outcome. So as we wrap up, it's incredibly clear that the next great AI advantage won't belong to the company that simply deploys the highest number of autonomous agents. It will belong to the company with the best, most rigorous organizational design. It all comes down to architecture. Can you seamlessly orchestrate human judgment and machine execution to generate actual measurable financial returns? Think back to that physical perimeter of trust we used to rely on. You know, the ID badge, the desk, the clear lines of authority. As those physical boundaries dissolve into digital workflows, we invite you to look at your own enterprise. If AI agents became a material source of execution capacity in your operations tomorrow, would your operating model clearly define who decides, who acts, and who ultimately remains accountable when the system breaks? That is the question every executive must answer.