Grown up agents, stalled pilots, and the bill nobody warns you about
By Kris Van Hauwermeiren ·
Agents now act instead of advise, yet most pilots still stall. And the real cost is not the token price, it is the human time to check what the agent did.
This week brought two stories about AI agents that, read together, tell you almost everything you need to know about where we are. On one side, agents are clearly growing up. On the other, most projects that use them still go nowhere. And hiding underneath both is a cost that almost nobody talks about. Let me unpack all three, in plain terms, and what they mean if you run a smaller company.
Agents are finally growing up
For a long time, AI could answer a question but not actually do anything. That is changing fast. The new wave of AI agents can carry out multi step tasks across your real tools, your email, your calendar, your files, your chat. This week, for example, one large company rolled out an agent based assistant from a pilot of a thousand people to more than thirty thousand staff. The tools can now read a request, plan the steps, and act across connected systems, not just reply with text.
This is a real shift. We are moving from AI that advises to AI that acts. For a business, that is a much bigger deal, and a much bigger responsibility, than a chatbot that drafts an email.
So why do most pilots still stall?
Here is the uncomfortable other half. While the technology grows up, most company pilots still stall and quietly get shelved. We have seen the numbers in earlier research: the large majority of AI projects deliver no measurable result. The agents got better, but the success rate did not jump with them.
The reason is the one I keep coming back to. A pilot fails when a company takes a powerful new tool and points it at the old way of working. The agent can act across your process, but if that process is a tangle of human handoffs and unwritten rules, the agent has nothing clean to act on. It is the faster horse problem again. The tool is new, but the track underneath it is the same old one.
Same technology, different track: a tangled process stalls, a clean one lets the agent move.
The pilots that work share a simple pattern. They pick one clear, repeatable task. They name one person who owns it. They agree what a good result looks like, and where a human signs off. Small, clear, and owned beats big, vague, and impressive every time.
The bill nobody warns you about
Now the part that surprises people. The headline on agents this week was that they are getting cheaper, a lot cheaper. New models claim to run agent tasks for a fraction of the old price. That sounds like pure good news. It is not the whole story.
When an agent only answered questions, a wrong answer was cheap to ignore. When an agent acts, a wrong action has to be caught, checked, and undone. And the cheaper it becomes to generate work, the more work it generates, which means more to review. The real cost of an agent is not the token price. It is the human time to check what it did, and to clean up when it gets something wrong.
Cheap generation, expensive oversight.
This is the trap. A team sees a low price per task and lets the agent loose on everything. Then the hidden bill arrives: hours of people checking output, fixing mistakes, and rebuilding trust with a customer who got a wrong answer sent with full confidence. Cheap generation, expensive oversight. If you do not plan for the second part, the first part will cost you more than it saves.
What this means for a smaller company
You might think all of this is a big company problem. It is not. For a smaller company, the stakes are actually higher, because you have less slack to absorb a mistake and less time to waste on a pilot that goes nowhere. But the same facts also work in your favour, if you are sensible about it.
- Start with one task, not your whole business. Pick something repeatable and low risk, and prove it works before you widen it.
- Give it an owner. One person who defines a good result and signs off on the output. An agent without an owner is how pilots quietly die.
- Count the full cost, not the sticker price. Add the human time to check the work. If that time is larger than the time saved, the task is not ready for an agent yet.
- Keep a human at the point where a mistake would hurt. Let the agent do the routine flow, and keep people on the decisions that touch a customer, a payment, or your reputation.
The honest summary
Agents are growing up, and that is genuinely exciting. But a more capable tool does not fix a broken process, it just breaks it faster. And a cheaper tool is not a cheaper outcome if you have to pay people to check everything it does. The companies that win with agents this year will not be the ones that deploy the most. They will be the ones that pick carefully, redesign the work around the agent, and count the real cost with eyes open. That careful, honest approach to putting AI to work is exactly what we help companies with at Binary Rabbits, because the goal was never to use the most AI. It was to get a better result.
About the author
Kris Van Hauwermeiren is the founder of Binary Rabbits and a Tech Director with 25+ years of enterprise experience in digital transformation, AI and technical leadership.