Agentic AI's real risk: losing the plot

These are the five questions I heard most, from more than one senior executive this year. The wording is mine; the gist is theirs.

These are reasonable questions. But they share a problem.

Each of these questions starts with the technology and works backwards to the business. None of the questions start with the business and work towards the technology. That pattern led me to go and investigate it and this post is about what I found, and where it led.

It could be that these questions came from companies of a certain size, or from a couple of sectors that happen to be more anxious about AI than others, and say nothing about business at large. So I went looking for evidence either way. I came across the McKinsey and MIT studies, Gartner’s forecasts, the shadow-AI reports and the Klarna coverage. Each one, in its own way, turned out to be telling the same story. Here is the trail.

ActivTrak 2026 · Gartner via CSA · Gartner via HRD

So the companies I met were not the exception. They were the general case. Laid side by side with the five questions, the reports stop looking like six separate findings and start looking like one case, with three pieces of evidence.

Exhibit one: the value exists, then goes missing. Eight in ten people say AI makes them more productive. Fewer than four in ten companies can find that productivity in their accounts. Somewhere between the individual and the P&L, the gain disappears. That is not a technology failure.

Exhibit two: nobody blames the models. MIT says the tools never learned the workflow. Gartner says unclear value and misapplied use cases. Not one of these reports says the AI was not clever enough. They say the same thing: the technology was fine, the target was wrong.

Exhibit three: Klarna’s agent did exactly as it was told. It was told to cut cost, and it cut cost. Customers left, quality fell, humans were rehired. The agent did not fail. The brief did.

Put the three together and the pattern behind the five questions is no longer a hunch. Companies are not losing the AI race. They are losing track of what they are for, and then handing that confusion to a machine that executes it faster.

Nobody wants to be left behind. The board reads the headlines. A competitor issues a press release. A vendor books a meeting. The Head of AI gets a mandate: do something with agents, this quarter. The mandate turns up before anyone has written down what problem it is meant to solve.

So three old questions get skipped. They are not AI questions. They are the questions any operations director would have asked twenty years ago.

Only once you have answers to those three does the fourth question make sense. That fourth question is the only one about agents that was ever worth asking: in a process that already delivers our value, in a way we would be proud of, where would an autonomous system do the job better, faster or more reliably?

With that in mind, go back to the five questions that executives asked.

The data backs this up. McKinsey’s high performers are the companies that redesign the workflow rather than drop AI on top of an existing process. They are also twice as likely to have a defined way of measuring impact. MIT found the same thing from the bottom up: the startups that scaled picked one specific pain point and went deep on it. In both cases the AI came second. Understanding the work came first.

That left one question open. Companies skipped the basics through the ERP years, the cloud years and the first chatbot wave, and most survived. Why wouldn’t they survive this one? The reports kept hinting that something was different, so I went looking for what. It was not scale or speed on their own. It was that, for the first time, the software acts. In the earlier waves, software helped people make decisions. It put something on a screen, and a human read it and decided what to do next. In this wave, the software decides and does. It reaches into your systems, takes the action, and moves to the next one, at a speed and scale that no human can review. Earlier tools waited for a person. Agents don’t.

A chatbot answers a question. A dashboard shows a number. An agent acts. It takes a goal, breaks it into steps, calls systems, makes decisions and carries them out, many times a minute, usually with nobody watching each one.

So an agent does not fix a process. It speeds up whatever process you give it. Give it a process that already serves the customer well and reflects how the company wants to work, and you get more of that. Give it a process that already favours cost over quality, or speed over care, and you get more of that too — at scale, with the individual decisions buried in logs nobody reads. Klarna is the public example. I would bet there are hundreds of private ones.

Then there is the second problem, which I see in almost every company: people running ahead on their own.

Every company has a few people who are further along with these tools than everyone else. They are curious, capable and impatient. With no direction from the top, they build things, and good for them. But each of them, build towards their own idea of what good looks like, in their own department, on their own stack. In practice it looks like this:

Each project may show a result on its own but together they pull the company in several directions at once. The seven-tools-and-rising figure in the reports and the 13% governance figure are what this looks like from outside. From inside it looks like a company that is doing a lot of AI and slowly losing track of what it is for.

I do not think “collapse” is too strong a word for where that leads. A company is, at its core, an agreement about how it creates value and for whom. Automate enough different versions of that agreement, fast enough, and there is no longer one company. There are several companies, sharing a logo.

By this point it was clear that the five questions were pointing at an underlying problem: companies had lost sight of what they were for. It would not be right for those executives to simply wait and watch. The 6% best performers are racing ahead, and the gap between them and the laggards will only widen with every quarter spent waiting. Here is what the executives should be doing instead, in this order. The first three steps have nothing to do with AI.

Pick your two or three main value streams: the path from a customer needing something to that customer paying you for it. Map each one as it really runs, not as the org chart says. Who touches it, in what order, using which systems, and how long each step takes. If nobody can do this for a process, that process is not ready for an agent. It is ready for someone to go and understand it.

What did this company promise its first customers? Which of those promises still sets you apart? Walk the map created above and mark each step: does this serve the promise, or is it left over from something else? You will find both. The leftover steps are usually where people most want to deploy agents, because they are painful. But automating a leftover step keeps it forever.

For each step you are considering, name the outcome the customer would notice and the number you measure it with today. Eg: Time to resolve. Errors per hundred. Repeat rate. Margin per order. The moment “productive” and “ROI” are attached to one step and one number, they become answerable. Not before.

Good candidates are steps that are clearly specified, high volume, able to tolerate a mistake because there is a way to catch and fix it, and currently slow because a person has to pay attention rather than because a person has to judge. Steps that depend on nuance, trust or the relationship itself are where Klarna learned its lesson.

Decide who may build agents, against which data, with what review, and how an agent’s decisions get logged and checked. This is not red tape. It is how you let your fastest people run fast without running the company apart. Give them a direction and some boundaries and they become your advantage. Leave them alone and they become the five-copilots problem.

Choose the platform last. By now your process map tells you what to look for. The platform or product should plug into the systems that the step already uses, keep working when the workflow changes, and prove itself on your data. Be sceptical of the brochure. Gartner found only around 130 genuine agent products among thousands calling themselves one; the rest are chatbots or automation scripts with a new label. So before any large amount of money changes hands, make the vendor run their product on one of your real workflows, with your real data, and show you what happened.

Two things about this list. First, steps one to three are things a good operations leader could have done a decade ago. They feel urgent now because agents make the cost of skipping them much higher. Second, the list works whether or not you end up deploying agents. A company that does steps one to three and decides agents are not right for it yet has still come out ahead.

The five questions I keep being asked are not really questions about AI. They are questions about the business, dressed up as AI questions. Take the costume off and each one becomes: what do we deliver, to whom, how, and why does it work?

The numbers say nearly every company is experimenting, fewer than four in ten can see any financial result, four in ten agent projects will be cancelled, and the companies that win are the ones who understood and redesigned the work before they automated it. Klarna says that an agent given the wrong target will hit it precisely.

So my advice to an executive who is unsure; is dull and boring. Do not start with the agent. Start with the customer, the thing that made you worth something to that customer, and the process that connects the two. Fix what no longer fits. Decide what good looks like and how you will measure it. Put rules around the people who are keen to build. Then, and only then, ask where autonomy helps.

Agentic systems are going to be one of the most powerful things a company can point at a problem. The companies that get left behind will not be the ones that moved slowly. They will be the ones that moved fast in a direction nobody had chosen.

So the next time someone asks which agent platform to standardise on, try asking them back: which step in your process would you be happy for it to run a thousand times tonight, unsupervised, and how would you know in the morning whether it had done well?