Your bot's first week — what to hand it, what to hold back

The most common way bookkeepers fail with AI isn't picking bad tools — it's handing the bot too much on day one, watching it wobble, and concluding the whole thing doesn't work. A first week designed for a win looks completely different. Here's the plan I use on every new fleet, including my own.

Pick the boring client on purpose

Your first bot book should be your *least* interesting client: stable transaction patterns, one or two bank accounts, a feed where the same fifteen vendors repeat all month. Not your messiest book — that's an unfair test. Not your biggest — that's an expensive one. You're not proving the bot can do bookkeeping; you're proving it can do *your* bookkeeping, and boring books teach that fastest.

The three jobs of week one

1. Statement pull. One account, one bank site. The bot logs in, downloads the current statement, names it to your convention, files it where your close expects it. This is the perfect first job because the output is binary — the PDF is there or it isn't — and it exercises the scariest-feeling part (a bot on a bank website) on a read-only task.

1. Feed clearing with training wheels. Write rules for the ten most common vendors the way you'd brief a junior: "this fuel card is always vehicle expenses," "transfers between these two accounts always match." Then let the bot clear *only what a rule covers* and leave everything else untouched. Day one it might clear 60% of the feed. That's a win — and every question you answer becomes a new rule.

1. One reconciliation it's allowed to finish only if perfect. The standing rule from my fleet: a bot completes a reconcile only at a $0.00 difference. Anything else gets left with the difference computed and a note. This is the guardrail that makes the whole system trustworthy — the bot literally cannot force a bad tie.

What to hold back — for now

Judgment calls. Owner draws vs. payroll, deposits that might be loans, anything a client would need to explain — those stay yours, and honestly they stay yours forever. The bot's job on ambiguity is to ask, not decide.

Client communication. Week one, nothing the bot produces reaches a client without your eyes on it. Not because the work is wrong, but because your name is on it and trust is built in one direction only.

What week one actually proves

By Friday you'll know three things you can't learn from a demo: how the bot behaves when a login throws 2FA at it, what your rules look like written down (most bookkeepers have never externalized them — it's clarifying), and how it *feels* to review work instead of doing it. That last one is the real product. The mechanical layer was never the job; it was just in the way of the job.

Week two, you add the second account. Week four, the second client. The fleet grows one boring, verified step at a time — which is exactly how you'd train a human, and exactly why it holds.

*This first week is literally what Guided Setup is — we build it on your real client file, together.*