Insights
Individual use of AI is widespread by now. Most people use it to make their own tasks easier: drafting emails, summarising documents, speeding up analysis. Useful, but this is the tip of the iceberg. It makes individuals faster without changing how the organisation works, and it rarely shows up in results.
One layer deeper sit the teams working with agents: an agent that processes incoming invoices from mailbox to ERP, prepares the weekly management report, or handles standard customer requests end to end. These systems do not advise, they act. They carry out multi-step work and hand back finished outcomes instead of suggestions.
And then there is a small group of companies attempting the real step: embedding agents in their processes, data flows, and decision making, to eventually steer the business on it. Most companies are nowhere near this yet. And of the ones that are, many have not set it up properly.
Where we see agent initiatives stall, it is almost never the agent. It is the organisation the agent is dropped into.
An agent that acts makes demands a chatbot never made. It needs data it can reach and trust, a process clear enough to hand over to a machine, an owner who is accountable for the outcome, and clarity on what it may decide on its own. Take one of those away and the agent will still run. It just stops delivering.
Implementation feels easy, and that is exactly the danger. Switching an agent on takes days. Building an organisation in which that agent delivers value is the real work. The second part gets underestimated precisely because the first part has become so accessible.
In our conversations with management teams and investors, the same patterns keep coming back.
Agents get implemented without the discipline that is standard for any other investment. Many companies are still in trial mode. There is no defined value upfront, no measurement, and no moment where someone decides to stop or scale. The pilot has become the goal instead of the means.
A problem gets sketched and the answer has to be AI. What is missing is an understanding of what the organisation needs to make that true: which data, which process clarity, which capabilities, which decisions.
And when an agent gets implemented anyway, without those conditions in place, the outcome is predictable. The agent lands on a process that is unclear, full of exceptions, and dependent on knowledge that only exists in people’s heads. It automates confusion and makes the mess move faster. The solution was chosen before the work was understood, and the gap surfaces months later as a stalled initiative.
The agent acts, but nobody owns what it does. Governance tends to be binary: either the agent is allowed to do nothing and delivers nothing, or it is allowed to do everything until the first incident. The layer in between is missing. Explicit agreements on what the machine executes and where human judgment takes over.
The wrong conclusion would be to stop. Keep experimenting, and keep letting people explore what AI can do in their own work. That is how an organisation builds familiarity. But the step that follows deserves real thought, because that is where the nature of the work changes: execution starts running on a combination of human judgment and machine capability. Organisations that learn this way of working early will move faster than those watching from the side.
The ambition can stay. The approach has to change: from trying things out to implementing, from adoption to demonstrated value. And implementing an agent changes more than people expect. Processes get redrawn around the agent, roles shift from executing work to checking and steering it, and data suddenly has to be in order because a machine is acting on it. Whoever treats this as an IT rollout will find out it is an organisational change.
The real added value of an agent is not proven at go-live. It is proven afterwards. So define upfront what the agent should change in performance, cost, cycle time, or customer behaviour, and then actually check whether it happens.
Treat the agent like any other investment. Value defined upfront, visible change as evidence, an explicit moment to scale or stop, and evaluation as a routine rather than an afterthought. That evaluation matters more here than with most implementations, because a successful agent reaches beyond the process it runs in. It frees up capacity, shifts how people spend their time, and changes how parts of the business are steered. Impact of that size deserves to be measured, and initiatives that show no movement deserve to be stopped, because they keep competing for attention and capacity.
The models get better every month, and experimenting with them costs almost nothing. The gap opens one step later: between companies that implement agents with a clear view of the value, the conditions, and the follow-through, and companies that switch them on and move on.
Your agents are ready. Whether your organisation is ready for them is the question worth answering first.