Your first agent should do the boring job

Ask a leadership team which process an agent should take on first and the answers are rarely small. Handle complaints end to end. Run procurement. Manage the sales pipeline. Each would matter, each makes a good slide, and each is a poor first choice for the same reason: nobody can tell, quickly and cheaply, whether the agent did it right.

The first agent that actually reaches production is usually dull. That is not a lack of ambition. It is how the second and the third get approved.

Why the impressive candidate stalls

Nobody can check it. If judging whether the agent handled a case correctly takes an expert half an hour, the evaluation set never gets built, the error rate is never known, and the risk review has nothing to approve. The project stays a pilot because nobody will sign off a number that does not exist.

The process is several processes. Broad tasks are usually done differently by different people, and the differences are where the judgement sits. An agent built on top of that encodes one version at speed and surprises everybody who did it another way.

Every mistake is expensive. A wide scope means the agent touches money, customers and commitments early. Each of those needs an approval gate, and an agent where almost everything needs approval is a slower way of doing the work by hand.

Failures cannot be debugged. When something goes wrong in a broad task, the cause can be anywhere. A narrow agent fails in ways that can be found, explained and fixed within a week.

What a good first task looks like

The criteria are unglamorous, and most of them can be checked in a spreadsheet before anybody talks to a supplier.

High volume. Hundreds of cases a month, say, rather than a handful. Volume is what pays for the build, and it produces enough evidence to widen the bounds later.

A checkable outcome. Somebody can look at the result and say right or wrong in a minute or two. Matching an invoice to an order is checkable. Settling a commercial dispute is not.

A process that exists. Written down, or at least done the same way by the people who do it. If two experienced people would handle the same case differently, settle that first.

Reversible actions. Preparing, classifying, drafting, routing and updating internal records come before anything that pays, promises or contacts a customer. The irreversible steps can follow later, behind a person’s approval.

A known current cost. How long a case takes today, and who does it. Without that, nobody can say afterwards whether anything improved.

An owner who wants it. A line manager whose team does the work and who will read the escalations. Not a sponsor three levels up.

The matrix worth drawing

Plot the candidates on two axes: how many cases arrive in a month, and how quickly a person can check a result. The top-right corner is where the first agent belongs. The bottom-left is where the flagship ideas usually sit, and they are better served by fixing the process than by automating it.

Infographic

Where to look for the first agent

Easy to checkHard to check

A checklist will doEasy to verify, but too rare to pay for a build.

First agentVolume pays for it, and checking keeps it honest.

Leave it with peopleRare, judgement-heavy and expensive when wrong.

Define the outcome firstWorth automating once right and wrong can be told apart.

Few casesMany cases

The exercise takes a process owner and an analyst a few days, not a programme. If it cannot be done because nobody knows the volumes or how long a case takes today, that is the first finding, and it is much cheaper to discover in a spreadsheet than halfway through a pilot.

The boring agent earns the interesting one

A narrow agent in production produces things a slide cannot: a measured error rate on real cases, a decision log, a record of how long approvals take, and a team that has lived with the system. That evidence is what a risk committee needs before approving a wider scope, and it is the only thing that makes the flagship idea credible later.

It also builds something less visible: people who know how to work alongside an agent. They learn how to read an escalation, when to reject a draft, and what to note when something looks wrong. That skill carries over to the next agent. A stalled flagship pilot teaches nobody anything.

Starting with the impressive task tends to produce the opposite: a long pilot, no number, and a sceptical organisation by the time somebody proposes a second attempt.

How we approach it

In AI agents engineering the first step is scoping one task — narrow, high-volume and with a checkable outcome — and we say so plainly when the process is not ready to be automated. A narrow first agent in production behind approval is typically a matter of weeks. Widening its bounds takes longer, and should.