You Can’t Stop AI Agents. But You Need to Know Who’s Accountable for Them

“Business will increasingly rely on AI agents,” says Clay Van Doren, the new CEO of KKCG-owned Avenga. The companies that get real value from them, he argues, will be those that can clearly identify who is accountable when an AI agent makes a mistake.

Source: Wired.cz

Avenga, part of Karel Komárek’s KKCG group, has a new CEO: Clay Van Doren, formerly Group Chief Growth Officer at global IT company Atos. In his new role, he leads more than 6,000 professionals across Europe, the Americas and Asia, and is bringing fundamental changes to the company’s approach to artificial intelligence and its overall strategy — including an entirely new vision of businesses supported by armies of AI agents.

“We are at the point where one person can orchestrate stages of delivery that used to need a department,” he says in an interview with WIRED, his first with a Czech media outlet. “That isn’t because nobody is in the loop. It’s because the loop has been designed, and somebody still owns the outcome.” For more than fifty years, he adds, everything in IT has been sold on time and effort. “Put an agent into a process and hours stop being the thing that decides how long it takes.”

On LinkedIn, you said that you no longer see AI merely as a feature added to software, but as a new way of creating software in the first place. For most of your career, new technologies have been incorporated into existing ways of working, but you say AI represents a fundamental change. How fundamental is it?

Most technologies get absorbed the way cloud was. Companies took the processes they already had and moved them somewhere else. The infrastructure changed and almost nothing else did. AI doesn’t behave like that. Bolt it onto one step and it gives you very little. Rebuild the process around it and the numbers change completely.

Software delivery has a handful of stages, and until recently we were doing what everyone else was doing. A tool here to help write code, a different tool there to generate tests. Useful, but each one started cold. None of them knew what the others had worked out, so a developer spent the day re-explaining the same system to four different assistants. What changed isn’t simply that the tools got better. It’s that context now survives from one stage to the next. What we learn pulling apart a legacy estate is still there when we design the target architecture, still there when we build, still there when we test. That continuity is the whole game, and it’s the part you can’t buy off the shelf.

Take a legacy estate. Four million lines of COBOL (a programming language developed in the 1960s, editor’s note), or a Java system nobody has documented in fifteen years, written by people who retired a long time ago. What we can do now is reconstruct how that system actually behaves and which business rules it enforces. I want to be careful here, because our industry oversells this. You do not recover intent. Code tells you what a system does, not what anyone meant it to do, and old estates are full of defects that became load-bearing and rules that enforce legislation repealed years ago. What you get is an accurate map of behavior, and that is exactly what you need in order to decide what to keep, what to retire and what to rebuild. Better models made this work feasible. They did not make it reliable. Reliability comes from how you break the estate down, how you map the dependencies, and how you check every finding against what the system really does in production.  From there it runs as one engineered lifecycle rather than a relay race.

You start by deciding what is worth doing at all, which is the step most people skip. You design the target architecture. You build. Then you verify, and I mean verify independently, not let the same system that wrote the code mark its own homework. Deployment produces the documentation as a byproduct, so what you hand over is an estate somebody understands. Then it loops. Skip the decision step and you will automate work that should never have been done. Skip independent verification and you have built the kind of thing that fails its first audit.

So yes, one person can now orchestrate stages that used to take a department. Not because nobody is in the loop, but because the loop has been designed. Somebody still owns the outcome, and at a regulated client that somebody has a name.

“First, who captures the value — the client, the supplier or the model provider? And second, how does accountability work once the day-to-day decisions are made by machines?”

How efficient is this change?

Keep the context flowing across the whole process and quality goes up while cost comes down, which surprises people, because they are used to trading one against the other. On modernization projects, where you are up against decades of accumulated waste, we are seeing efficiency gains of 50 to 70 percent. I would be careful with that number, though. It isn’t a discount you can apply to any piece of work. On greenfield projects, the gains are real but smaller. And if you use AI at a single point in the process, testing, say, or code generation, you will get your 20 percent and wonder what everyone is so excited about.

How do you quantify the actual return on investment? Skeptics say that ROI is difficult to calculate for AI.

We built a measurement framework for exactly this problem. Internally, we call it A-UPI, the Avenga Unified Productivity Index. It starts by establishing a baseline, which almost nobody bothers to do, and that is why so many AI business cases fall apart when somebody asks a hard question six months in. After that, it tracks effort, throughput, quality and delivery speed against that baseline, so the client can see where efficiency is genuinely improving and where it isn’t. The uncomfortable part is that it will sometimes tell you a piece of work is not paying for itself. I would rather know that in month two than in year two, and so would the client.

Apart from its approach to AI, what is changing at the company following your appointment as CEO?

An AI-native approach now sits at the center of everything we do, and the hard part isn’t the technology. It is 6,000 people learning to work in a lifecycle rather than in a lane. If you test software for a living, you now need to understand what happened before your stage and what happens after it, because that is where your context comes from. We have put real money into that program. It is the part competitors can’t copy quickly, and it is also the part everyone underestimates.

Our strategy can be viewed from several different angles. Historically, we focused primarily on traditional data products and custom software development, predominantly for the private sector. We are remaining true to those roots — data, software development and cybersecurity — but we are now approaching them in an AI-native way. That is a fundamental shift. As for what we sell, it is now much more about managing an entire transformation with AI, whether by building an agentic enterprise or through software development itself.

What will change in the markets where you operate and among your target customers?

Geographically, we will continue to serve our existing key markets: Western Europe, North America and Eastern Europe. Eastern Europe is where we base some of our development capacity, but we also deliver services directly to that market. We likewise operate in Argentina and elsewhere in South America, and we will continue to do so. As for growth, we have very ambitious targets. Western Europe and North America will be the main drivers — specifically the United Kingdom and Germany within Western Europe. We see these as our largest markets.

In terms of customers, we are moving into the public sector in the UK and Germany. We have delivered to government before, so it is not new ground, but it is new at this scale for us, and we are building the frameworks and clearances that go with it. We are also focused on utilities and transport, railways in particular. These are Tier 1 and Tier 2 organizations, large regional players rather than global giants. What changes is that instead of delivering projects for them, we take responsibility for the transformation alongside them, which makes the average engagement around four times the size of what we do today.

What is your vision for the next five years?

We have prepared a five-year strategy approved by the KKCG group. Under this strategy, our plan is essentially to double the size of the company over the next five years. We want to shift approximately 35 percent of the services we provide toward new AI-based solutions.

What role do you see Europe, including the Czech Republic, playing in these developments?

In my view, Europe has a major opportunity to become a fast follower and learn, in particular, from what has happened in the United States. The opportunity lies in adopting a different approach to AI. The main problem we currently see in North America is that too many projects were created as pet projects — essentially isolated experiments conducted in a supposedly safe sandbox. The problem with a sandbox project is that it operates in a very limited environment. You may test whether something works from a technical standpoint, but when you attempt to scale it and integrate it with the company’s other live systems, you discover that the solution was never designed or tested for that purpose.

Nor have you tested how regulatory requirements will be incorporated into the solution, how audits will be supported, or how security will be ensured. Europe therefore has an enormous opportunity to establish this entire end-to-end process correctly from the outset — to design systems around genuine software engineering rather than mere software development. This gives you confidence from the beginning that the solution will satisfy all information-security standards and regulatory requirements, including those of the European Union. Where we do not have much of an opportunity, however, is in competing directly in the field of large language models. Others are far ahead of us there. But I do not think that matters very much anymore. We can easily use their models — while applying them differently from the way they were initially used in the United States.

“For fifty years this industry has sold time. You allocate people, you count hours, you invoice. Put an agent into a process and hours stop being the thing that decides how long it takes.”

What risks or bottlenecks might Europe encounter with this approach? The disadvantages of excessive token consumption, or tokenmaxing, and the actual cost of AI seem fairly alarming.

First, let’s hope nobody introduces new tariffs or trade barriers. On tokens, the cost only becomes an issue when the value of what you are building is already marginal. We run an AI-native engineering platform called Intelligent Flow. It carries a piece of work across the entire lifecycle, from discovery and the decision about whether to do it at all, through build, operate and improve, so that context, controls and evidence travel with the work instead of being reassembled at every handover.

Token efficiency follows naturally from running that properly. If the system knows what you are building and why, it isn’t burning cycles rediscovering the same estate every morning. When token cost is the thing that kills a project, the project was usually sitting on the line already. Improve a process by 50 percent and token prices are a rounding error. Chase a 20 percent saving on code cleanup, with no control over what runs overnight and developers firing off compute whenever they feel like it, and yes, the economics can turn on you.

At Avenga, we measure ROI from the very beginning, but more importantly, we design it together with the customer from day one. I often see the same scenario repeated: at first glance, something looks like an excellent investment with a high ROI, but problems arise when you attempt to move it into production. This is one of the main reasons why AI is currently perceived as a disappointment in many companies. Early-stage testing focuses only on the technology itself, not on its seamless end-to-end integration with all the other systems and processes actually operating within the business.

How does the deployment of AI agents work in practice at your company?

The agentic enterprise starts from an unglamorous observation. A lot of processes have people in them because the process was designed around paper, not because anyone needs their judgment. Procurement. Parts of HR. Finance operations. Customer service. People spend their day moving information between systems that should have been talking to each other years ago. Companies put chatbots on top of all that and nothing changed, because a chatbot answers questions about a broken process instead of fixing it.

Agents are different, but you can’t graft them onto the old procedure either. You redesign the process, then you place agents inside it with controls around them. In procurement, an agent can carry an approval or a payment from one end to the other, while a human still maintains segregation of duties and signs where the auditor or the law requires a signature. That constraint is not something to engineer away. It is what makes the thing deployable inside a bank. What you get back is capacity, and most of it goes to work that requires more judgment. We are retraining our own people for precisely that reason. I won’t pretend every role comes through unchanged. Some won’t, and any company telling its people otherwise is not being straight with them.

The same applies to HR. Most activities in human resources today are extremely routine. Human intervention adds no value to the task. The same is true in finance. For fifty years this industry has sold time. You allocate people, you count hours, you invoice. Put an agent into a process and hours stop being the thing that decides how long it takes. The constraint moves to judgment and verification instead. That flips the commercial logic from inputs to outcomes, and it is a harder adjustment for suppliers than for clients, ours included, because our whole industry was built on billing by the day.

“People never stop being accountable for how those machines behave.”

Do you perceive a certain degree of skepticism toward AI among your clients and other companies, or concerns about an AI bubble and the risk of putting all their eggs in one basket?

For the overwhelming majority of customers, that train has already left the station, so to speak. They are simply working with and deploying AI. When objections or skepticism arise, I believe it is generally because the public, investors or company boards are nervous and do not understand AI. And because of that nervousness, companies are not entirely transparent about the extent to which they are actually using it. All the customers I am in contact with are introducing AI in some form. Some are building new systems; others are rescuing projects already underway that have run into trouble. But AI now accounts for a good 80 percent of all our discussions. There are exceptions, of course. One client told me: “It’s great that you’re doing this, Clay. When we decide that we want to experiment with such fads, we’ll call you.” But this is precisely the type of customer that will find itself in trouble very soon, because its competitors will not be thinking that way.

What will be the biggest change in business over the coming decade?

B2B companies are still run on time worked. Four people here, six there, and you manage the plan. Look at the whole process, though, and where it ends up is agents handing work and data to other agents, with people sitting at the decision points. Getting there is an operating model change rather than a tooling change, and most companies are underestimating that by a wide margin. The gains in quality and productivity are real, and they raise two questions nobody has answered yet. Who captures the value — the client, the supplier or the model provider? And how does accountability work once the day-to-day decisions are made by machines?

People never stop being accountable for how those machines behave. But “the system decided” is not an answer any regulator will accept, so somebody has to design the evidence trail before the incident rather than after it. That is the biggest change coming, and it is a governance question more than a technology one.

And from a societal perspective?

AI becomes part of almost everything, quietly. People fear mass job losses. My worry runs the other way, and it is arithmetic rather than optimism. When we look at the Western world, we simply do not have enough people. That may sound strange, but I mean economically active, working-age people who can contribute productively to society, pay taxes, and perform all the work that needs to be done. Just look at the aging population: the demographic curve is not developing in the way we need it to. We must therefore find ways to make human work more efficient and significantly increase productivity so that we can address the looming labor-market crisis. In my view, we are almost certain to face that crisis within the next ten years.

Author: Darek Šmíd

This is an English translation of an article originally published by WIRED Czech Republic and Slovakia 
Avenga CEO Clay Van Doren: You Can’t Stop AI Agents. But You Need to Know Who’s Accountable for Them