Choose your next AI job only when the first one runs without the person who built it. If that person takes a fortnight off and the work still goes out on time, you’re ready to move. If the work stalls while they’re away, you don’t have a second job to start. You have a first job to finish.

The first job has to run without its builder

That absence test matters because plenty of first jobs look complete months before they are. The tool works, the output reads well, and the person who set it up still corrects it every Friday afternoon. Nobody counts that quiet correction time, so nobody notices it.

That hidden effort is what stops you moving. Your second job will want the same attention the first one wanted. If the first still needs a minder, the second gets whatever attention is left over, and that is rarely enough.

So watch the first job for a month while you stay out of it. Let the person who built it answer questions, but ask them to change nothing. Anything that needs your hand during that month is unfinished work on the first job.

Three signs a job has settled

A month of quiet watching gives you three things to look for. The first sign is that somebody other than the builder runs it from start to finish without asking for help. Written notes help here, although the real proof is a second person doing the job cold.

The second sign is that corrections have flattened out. In the early weeks you fix plenty, so the rate on its own tells you nothing. What matters is that it levels off and then stays level for several weeks running. If it still swings about, the job is not ready to leave alone.

The third sign is that the step lives in the process rather than in one person’s head. It has a trigger, a slot in the day and a named reviewer. Whoever takes the work over inherits the step with it, which is what naming an owner buys you.

Choose the next AI job for what it tests

Once the first job stands on its own, the temptation is to grab the biggest problem left on the list. Resist it. Your next AI job earns its place by testing a condition the first one never met.

Write down the conditions you still know nothing about. Can the tool cope with messy input, such as a scanned supplier invoice? Does it hold up once a second reviewer joins? Will it behave on text a customer actually reads? Does quality drop once the volume doubles?

Pick one condition, then find a job that puts it under pressure. A small job that answers a real question beats a large one that only repeats what you already know. The prize matters less at this stage than the lesson, because the lesson is what makes the third job and the fourth quicker to choose.

Change one thing at a time

With the condition named, the rest of the work is protecting the comparison. Keep the same tool, the same style of review and the same way of recording results. Change only the one condition you set out to test.

Change the tool, the task and the reviewer together and a poor result tells you nothing at all. You’ll spend a fortnight arguing about which part let you down. Hold the rest steady and the answer arrives on its own.

This is also why a second job in the same corner of the business is often the better choice. The people already know the review step, so the only new thing in the room is the condition under test. Familiar ground makes a clean comparison much easier to read.

Why several jobs at once defeats the point

Holding everything else steady is easy with one job and impossible with three. Running three in parallel feels quicker and almost never is. Each one needs a reviewer, a standard and somebody watching the running cost. Three sets of that land on the same few people, usually in the same week.

The second cost is worse. Mixed results across three jobs leave you unable to explain any of them. You end up with opinions rather than evidence, and evidence is the thing the first job was meant to give you.

One job at a time also keeps the decision honest. When a single job disappoints, stopping it is easy to say out loud. When three are running and one of them looks promising, the weak two tend to drift along unexamined for months.

What carries across from the first job

One job at a time also costs less than it sounds, because a surprising amount carries across. That includes the way you write a job down before you start, the review step, the measure you agreed, the supplier contact and the rules for what people may put into a tool. None of it needs building twice.

So your next AI job should cost less to set up than the first one did. If it costs the same, stop and look at why. Usually it means the first job produced nothing reusable, and that is worth knowing before you commit to a third.

The other thing that carries is credit. People who watched the first job work will give the second one a fair hearing. That goodwill is easy to spend and slow to earn back, so spend it on a job you have a good reason to believe in.

What you have to decide again

Some things never carry across, and treating them as settled is the common error. The reviewer changes, so the standard changes with them. What counts as a good output on an internal summary is not what counts on a reply a customer reads.

Decide four things again for every job: who reviews it, what they compare the output against, what the stop rule is, and what the job may cost each month before you look at it again. Write those down before the job runs. Agreeing what good looks like in advance is what makes the result readable afterwards.

Knowing when to stop adding jobs

Deciding those four again each time is what caps how many jobs you can sensibly run. Every live job takes review time from the same small group of people. When that review time starts crowding out the work it supports, you have reached the edge for now.

So watch the running jobs rather than the new ideas. If corrections start climbing on a job that had settled, the answer is to go back, not forward. Something upstream has changed: a supplier’s format, a new member of the team, a fresh sort of enquiry.

Delaying the next AI job by a month costs you almost nothing. Leaving a settled job to drift while nobody is looking costs you the credit you earned on the first one.