I keep having the same conversation.
A COO or head of operations tells me they ran an AI pilot last year. It worked in the demo. It worked in the test environment. Then it went near the real process and fell over. Now the board thinks AI doesn't work for a business like theirs.
It does. The pilot was aimed at the wrong thing.
What actually went wrong
Here's what usually happened. Someone picked a process that looked simple on a whiteboard. Invoice comes in, gets checked, gets paid. Three boxes and two arrows.
Then the agent met the real version. The supplier who sends invoices as photos. The cost centre that was renamed years ago but still lives in one spreadsheet. The approval that happens in a corridor because the system takes too long. The exception that one person in accounts handles from memory because she's done it for eleven years.
None of that was on the whiteboard. All of it is the process.
The agent didn't fail. It did exactly what it was told, on a process nobody had actually looked at.
The real process is never the documented one
In 20 years of selling and delivering technology I have never once seen a process that matched its documentation. Not in a bank, not in an insurer, not in a logistics business.
The documented process is what people agreed to do at some point in the past. The real process is what they do on a wet Thursday when the system is slow and month end is tomorrow.
People paper over the gaps. That's what good staff do. They know which field to ignore, which supplier needs a phone call, which report is always wrong. That knowledge keeps the business running and it is written down nowhere.
An agent has none of it. So the first time it hits a gap a person would have stepped over, it either stops or it guesses. In a regulated business, a guess is worse than a stop.
Why this gets expensive
Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls.
All three trace back to the same root. If you don't know how the process really runs, you can't price the work, so costs escalate. You can't measure the baseline, so the value is unclear. And you can't say where the decisions are made, so you can't control the risk.
You can spend six months and a six-figure sum finding that out the hard way. Plenty have.
Do it the other way round
Before anyone writes a line of agent logic, find out what is really happening.
Pull the event logs from the systems the process touches. Process mining will show you every route work actually takes, including the ones nobody admits to. Sit with the people who do the job and watch, don't interview. Count the exceptions. Put a cost on the rework.
Then, and only then, decide what to automate.
What you find is usually not what you expected. The step everyone wanted to automate turns out to be fine. The money is leaking somewhere else, in an error rate nobody measured or a handoff that adds four days.
That's why ETT starts most engagements with a fixed-price Business Process Workshop. You come out with a map of what really happens, a number on what it costs you, and a ranked list of what's worth automating. Sometimes the answer is an agent. Sometimes it's a rule change and a better form. We'll tell you either way.
It is a much cheaper way to learn than a failed pilot.
Source: Gartner, June 2025