AI as a colleague, not a magic trick
In a small company, AI isn’t a miracle button. It works like a sharp junior who needs to be onboarded and supervised.
There are plenty of promises around AI and very few practical guides. Companies either expect a language model to solve everything, or write it off after the first made-up answer. Both are a shame. What works best for me is to think of AI as a new colleague: fast, well-read and with no context.
Where AI really helps
- First drafts: product descriptions, customer replies, quotes, internal documentation.
- Research and summaries: going through long source material and pulling out what matters.
- Code and scaffolding: tests, migrations, data conversions, the repetitive parts of an application.
- Data cleanup: standardizing the formats of addresses, names or categories that were filled in by hand for years.
Where I leave it out
AI doesn’t make decisions. It doesn’t sign contracts, approve prices or take responsibility for the result. Every output that leaves the company has a person who checked it and stands behind it. The same goes for code: if nobody has read it, it doesn’t go to production.
Onboarding instead of prompts
The most common mistake is writing a new prompt from scratch every time. You wouldn’t brief a colleague differently every time either. So for recurring tasks I prepare a standing context: how the company writes, what products it has, what must never be promised and what a good result looks like. When something goes wrong, I don’t just fix the output, I fix that context too. The AI then improves much like a person you give feedback to.
Small rollout, fast feedback
I always start with one process. Not “we’ll bring AI into the company” but “AI will write the first draft of replies to complaints.” After two weeks I look at how much time it saved and how many corrections it needed. If the numbers add up, I keep going. If not, I stop, and no harm done.
AI makes a good team faster. With badly set-up processes, all it gives you is faster chaos.