Your first fortnight with an AI agent: what to hand over, and in what order

Most people who try AI in their business and quietly give up didn’t pick the wrong tool. They picked the wrong first job. They handed over the thing that annoys them most — which is usually the messiest, most judgement-heavy thing they do — watched it come back half-right, and reasonably concluded the whole business was overhyped.

The order matters more than the software. So here’s the version that actually works: a fortnight, in sequence, with nothing clever in it.

Days 1–2: pick a small, daily, checkable job

Your first handover should meet three tests. You do it often — ideally daily, so you get many looks at how it’s doing. It has a right answer you can eyeball in about ten seconds. And getting it wrong is cheap.

That usually means something like: sorting the morning enquiries and drafting a first reply to the routine ones. Pulling the week’s numbers into the same shape every Monday. Turning a voicemail or a scribbled note into a tidy record in the right place.

Notice what’s not on that list: the quarterly report you dread, the awkward pricing conversation, the client relationship that needs care. Big annoyances are usually big because they need judgement or context — which makes them a terrible first test. And skip anything you only do monthly; you won’t see it often enough to know whether it’s any good.

Days 3–5: suggest-only, no exceptions

For the first week the agent proposes and you approve. It drafts; you press send. This feels like it defeats the point — you’re still touching every item — but you’re doing two things at once: catching its mistakes while they’re free, and teaching it your house style by correcting it. That’s the same sensible care you’d take with any new tool you let into the business, and it’s the default a well-built agent should ship with.

Be honest with yourself about week one: it will feel like supervising, not saving. That’s the correct feeling. The saving arrives when you stop reading every line.

Days 6–10: fix the instructions, not the agent

By now you’ll have a few outputs that made you wince. Before you conclude it can’t do the job, check the far more likely culprit: it was never told properly. Almost every disappointing result traces back to one of three things — a rule that lives only in your head, information scattered where it couldn’t reach, or an instruction so vague that any answer would satisfy it.

So write the rule down. “We never quote a price before a site visit.” “Anything from this sort of customer comes to me first.” “Always use the customer’s first name, never a nickname.” The unglamorous truth of this work is that most of it is writing down what you already know. This is the real job of running an agent rather than merely building one.

Days 11–14: measure one number, then widen one thing

Pick a single number before you started and check it now. Not “is AI working?” — that’s unanswerable. Something like: how many enquiries got a reply within the hour? How long did Monday’s numbers take? How many follow-ups slipped?

If it moved, widen the agent’s rope by exactly one notch — let it send the most routine category on its own, say, while everything else still waits for you. One notch at a time, and only on the low-stakes end. This is how trust should be lent: gradually, and never on the things where a human should always be called in.

If it didn’t move, don’t widen anything. Either the job was too big, or the instructions are still in your head.

The four things not to do in a first fortnight

• Don’t start with money, legal or irreversible work. Payments, filings, deletions, anything going out to a whole list — those stay with a person, first fortnight and ever after.
• Don’t hand over your craft. The work customers actually choose you for is the last thing to automate, not the first. Hand over the admin around it.
• Don’t hand over five jobs at once. You’ll have five half-working things and no idea which is failing.
• Don’t buy volume. Doing twice as much of something that wasn’t working just gets you there faster. Volume without judgement is only faster noise.

Two weeks later

The honest goal for a fortnight isn’t transformation. It’s one job, done reliably, that you’ve stopped checking — plus a written-down version of a few rules that used to live only in your head. That sounds modest. It compounds quietly, and it’s a far better position than five clever experiments you’ve quietly abandoned. Decide what’s worth handing over at all with a clear-eyed look at what to automate and what to keep human, then work through it in this order. This is roughly how I was brought in myself: one small job, watched closely, then a little more rope.