There's a version of "we're adding AI" that produces a slide deck, a committee, and a pilot that dies in Q3. This is not about that version. This is the literal week-one log of what a working AI system looks like when someone builds it inside a real business — the kind of week we run for clients, written down so you can check it against whatever anyone is trying to sell you.
Day one is an audit, not a kickoff
The first day produces a boring document: every task your team did last week that followed a script. Answered the same six support questions. Copied order numbers from one tab into another. Chased three invoices. Wrote quotes from the same template with different numbers. Nobody likes making this list, because the list is where the payroll goes.
The test for each item is narrow: did a human make a judgment call, or did a human follow a procedure? Judgment stays human. Procedure is a candidate. Most businesses find the split is uncomfortably far from what they'd have guessed — the inbox alone is usually a third procedure.
Days two and three: one workflow, built end to end
The mistake everyone makes is building six things to 60%. A working week one builds one thing to done — where done means a specific person stopped doing a specific task, not that a demo ran once in a meeting.
Say it's the inbox. By day three there's an agent that reads every incoming message, answers the procedural ones — where's my order, what are your hours, can I get the invoice again — and files the judgment ones to a human with the context already assembled: who this is, what they bought, what they asked last time. The human opens a decision, not a mystery.
The agent's job description is one sentence: everything that didn't need you, handled; everything that did, prepared.
Day four: the approval layer, or nothing ships
Here's the part vendors skip and the part we consider non-negotiable: nothing the system produces reaches a customer, a supplier, or a dollar without a human approving that specific action. Not a policy document — a queue. Every outbound message sits in it, readable, editable, with an approve button a human has to press.
This sounds like it defeats the point. It's the opposite. The approval queue is where trust gets built at the only speed trust builds — one reviewed decision at a time. In week one you'll approve nearly everything and edit a lot. By week four you're skimming. The system earns autonomy the way a new hire does, with a paper trail.
Day five: instruments, or you're flying blind
- Every message the agent handled, counted.
- Every handoff to a human, logged with a reason.
- Every minute the old process took, benchmarked against the new one.
- Every edit a human made in the approval queue, kept — that's tomorrow's training material.
Numbers from a real week one tend to look like this: the agent fully handles 40–60% of inbox volume immediately. Not 95% — anyone quoting 95% in week one is reading you a brochure. The 40% is worth it on its own, and it climbs, because every human edit in the queue teaches the next draft.
Days six and seven: the handover, and the honest conversation
The last two days are documentation and a decision. Documentation: exactly what the system does, what it refuses to do, who approves what, and how to turn any of it off. If the person who built it disappears, your team can still run it — that's the bar.
The decision is the honest conversation: what's next, and whether there should be a next. Sometimes the audit from day one shows there isn't enough procedure in the business to justify more build. A real team tells you that and stops. The ones that never stop are billing by the month.
What week one is not
It is not a chatbot on your homepage that answers questions about your homepage. It is not an "AI strategy." It is not a model choice — which model matters far less in week one than which workflow, and anyone leading with model names is selling you their reseller margin. It's one procedure off a human's plate, an approval queue in front of anything that touches the world, and instruments that tell you whether it worked.
That's buildable in seven days inside almost any business that has procedures — which is every business with more than three people.