ManagementSeptember 05, 2026 · 9 min read
A 30-day plan to put AI into your business management
Week by week, with a success criterion for each and what NOT to do at each stage. Short on purpose: a data project that goes past a month without delivering value usually never finishes.
Data projects die from scope, not difficulty. The version that works is short, delivers a useful answer in week one, and only grows once someone actually uses it. The plan below assumes one person spending a few hours a week, not a team.
The success criterion for the month is not "everything connected". It is one person asking on their own, every day, without help.
Week 1 — one question and two sources
- Pick ONE expensive, recurring question. Usually "where did the sales come from" or "what did the acquired customer cost".
- Connect only the sources that question needs. Two is enough: sales and origin.
- Reproduce a number you already know — last month’s revenue. Matching it is the milestone of the week.
- DO NOT connect everything. An extra source at this stage only delays the first answer.
Week 2 — the assistant comes in
- Connect the assistant to the base and spend the week reading only. No write permissions yet.
- Ask the chosen question every day. Repetition is what builds trust and reveals data holes.
- Always ask for the filter used. Every discrepancy investigated this week is worth three months of later auditing.
- DO NOT build a dashboard yet. An early dashboard freezes the question before you know which one is good.
Week 3 — the live page and the routine
- Now publish the page with the numbers you look at daily — the ones that survived two weeks of questions.
- Take that page into the weekly meeting and settle doubts there, by asking.
- Switch on two alerts: stalled source and out-of-range drop. More than that becomes noise.
- DO NOT automate actions yet. Reading has to be trustworthy first.
Week 4 — act and expand
- Turn one analysis into action: an audience, a contact list, a budget cut.
- Define the write rule: what the AI does alone and what requires confirmation.
- Choose the next source based on what was missing in this month’s questions, not on a wish list.
- DO NOT expand the user base before the routine stands with one person.
How to tell whether it worked
| Week | Good sign | Warning sign |
|---|---|---|
| 1 | A known number matched | Nobody knows which number is the right one |
| 2 | A discrepancy appeared and was explained | Nobody asked anything |
| 3 | The meeting used the page | The old spreadsheet kept circulating |
| 4 | An action came out of a question | Reports only, no decision changed |
The most serious warning sign is in week 2: if nobody asked anything, the chosen question did not really hurt. Change the question, not the tool.
Three mistakes that waste the month
- 01Starting with the hardest source. The PDF-only one can wait; it eats the whole month and delivers at the end.
- 02Delegating the rollout to someone who does not ask the questions. Whoever feels the pain has to be at the keyboard in week 2.
- 03Measuring success by coverage. Ten sources connected and nobody asking is worse than two sources in daily use.