Your board is you. Your IT department is also you.#

In a corporation, AI adoption is a fight between two camps. In a small business there are no camps. There’s an owner who in the morning reads that “competitors are already rolling out AI” (FOMO), and in the evening fears that plugging anything in will break what is somehow working (the IT worry). The same emotions, just inside one head. That’s harder, not easier — because there’s no one to argue with, so there’s no one to stop you from doing something dumb.

In small businesses I see the same failure pattern: excitement → one flashy demo → no continuity → the tool dies after two weeks. What works instead:

Five principles that actually work#

1. Start with one painful process, not with a tool#

“We’ll roll out AI” isn’t a goal. “I’ll stop losing 4 hours a week writing proposals” — that’s a goal. Pick ONE process: repetitive, text-based, low cost of failure. Don’t start with invoices or anything that ships straight to a customer without a check.

2. Don’t move the tanker. You’re the pirate ship.#

The business that pays the bills stays untouched. Experiment alongside — on copies of data, on archived cases, not on a live customer. In a corporation, the special-ops team is a separate unit. In yours, that special-ops team is you or one willing person. The rest of the firm doesn’t even need to know you’re testing.

3. Small-scale POC Mirroring#

Same trick, smaller scale. For 2 weeks, do the selected task in two tracks: your way and with AI. Nothing goes live. Count: in how many cases did AI handle it well enough that you just copied the output? Usually you’ll get 30–40% — and a sub-group will emerge that the model does brilliantly. Dig in there. Leave the rest to the human.

4. List the things that could go wrong — before, not after#

Counter-intuitive, but it’s the only “documentation” you have time for. Three questions: where could data leak? what happens if the AI hallucinates to a customer? who checks the output before it leaves the company? Answer those and you’ve got 80% of the “security policy” a corporation spends six months writing.

5. It’s a game of habit, not subscription#

In a small business, AI isn’t killed by a bad vendor. It’s killed by abandonment after the novelty fades. One process, one owner, one fixed moment in the week. Better slow and weekly than fast and once.

Concrete: 30 days step by step#

• Days 1–2 — selection. List the 5 most tedious repetitive tasks. Mark one: lots of text, low risk, you do it every week.
• Day 3 — risk list. The three questions from point 4. Write the answers in a single file. That’s your policy.
• Days 4–5 — tool. One, off-the-shelf (ChatGPT/Claude). Don’t build anything custom. Give it 3 real, well-done examples as a template.
• Weeks 2–3 — mirroring. Do each case in two tracks and log a simple table: ✅ used as-is / 🔧 edited / ❌ trash. Goal: data, not impressions.
• End of week 3 — decision on the numbers. Look at the table. If AI handles a meaningful sub-group — roll out only that sub-group for permanent use, with a human as the last reviewer.
• Week 4 — habit. Set a fixed time and one owner for the process. Only now, when this is locked in, do you pick process #2.

What worries me#

I’ve got plenty of worries myself — mainly about how many firms confuse “I played with a chatbot” with “I deployed AI”. But the same thing works that works for big firms:

Slow is smooth, smooth is fast.