When an AI project stalls, the model is rarely the reason. The choice of work usually is.
The pattern is familiar. Someone sees an impressive demo, the team picks the most visible process in the company, and three months later there is a pilot that everyone admires and nobody uses. The agent was capable enough. The work it was given was the wrong first job.
So when a team asks me where to start, I don’t begin with tools. I begin with a filter.
Four questions before anything else
I run every candidate process through the same four questions. A process only qualifies as a first job for an agent if it passes all of them.
Does it run often? Something that happens a few times a year gives you no way to learn. You want work that happens daily or weekly, so you see enough cases in the first month to know whether it holds.
Can someone check the result in under a minute? This is the one teams skip, and it decides everything. If checking the agent’s output takes as long as doing the work, you’ve saved nothing. If a person can glance at it and say “right” or “wrong”, you can afford to review every case at the start and gradually stop.
Is a wrong result cheap and reversible? A request routed to the wrong queue costs a few minutes. A payment sent to the wrong account costs a lot more. Start where a mistake is an inconvenience, not an incident.
Does one person own the outcome? If nobody owns the result, nobody notices when quality slips, and nobody decides when the agent is ready for more. One name, one owner.
When I put a list of candidates through this filter, the winner is almost never the exciting one. It’s usually something dull: sorting incoming requests, preparing a first draft of a recurring report, matching documents against a checklist, summarising what changed since yesterday. That’s exactly why it works. Dull work has clear rules, frequent repetition and a cheap way to check the outcome.
Then replay yesterday
Passing the filter is not the same as being ready. The second step is a test I now treat as non-negotiable: give the agent yesterday’s real work before it touches today’s.
Take a day of closed cases. Give the agent only what was visible when each case arrived, nothing the team learned afterwards. Let it make its choice for every case, or flag it for a person. Then compare its choices with what the team actually did.
This does three things a demo never does.
It uses your data, with all its mess. Real requests are shorter, vaguer and more creative than any example someone writes for a demo.
It gives you disagreements to look at, and that’s where the value is. Not every disagreement matters equally. An agent that sends a billing question to the general queue is a nuisance. An agent that files an outage report as a low-priority question is a problem. Sort the disagreements by consequence, not by count.
And it turns a vague question, “is the agent good enough?”, into a specific decision: which kinds of cases it can handle alone, which it should only prepare, and which stay with a person for now.
What the owner gets out of it
At the end of this exercise, the process owner doesn’t have a pilot. They have a scope.
Something like: the agent handles these three types of requests on its own, prepares a draft for these two, and flags everything else. Every case still gets a quick human check for the first weeks. When the check keeps saying “right”, the check gets lighter.
That scope is what makes the difference between an experiment and daily operations. An experiment asks whether AI can do the work. Daily operations asks which part of the work it does, who checks it, and what happens when it’s wrong. You can only answer the second question if you chose the work carefully and tested it on real cases first.
Why not start big?
I understand the pull of the ambitious first project. It’s easier to get budget for a transformation than for sorting requests.
But the first process does more than save time. It teaches the organisation how to work with an agent: how to write down the rules, how to review output, who owns the result, when to widen the scope. Those habits transfer to the next process and the one after it. A team that has done it once on dull work moves to harder work much faster. A team whose first project was too big often never gets a second one.
The same reasoning applies to my own setup. The work I handed over first was the work I could check at a glance. Only after those checks became routine did I trust the system with more.
Where to start this week
If you’re choosing your first process, you don’t need a vendor or a strategy deck to start.
Write down five candidates. Run each through the four questions. Pick the dull one that passes all four. Pull last week’s closed cases for it, and replay one day.
By Friday you’ll know more about what an agent can do in your organisation than a polished demo will ever tell you. And you’ll have something better than an impressive demo: a first job the agent can actually keep.