How I work with AI
These are the rules I keep coming back to when I build AI products. Each one comes from something I shipped.
1Start from the process, not the model
Before I think about a model I map the work as it is done today: who does what, how long it takes, where the errors are. On the logistics platform I built, that mapping found two hours of daily manual work that needed an API integration and no AI at all. Lean Six Sigma taught me this before AI did.
2Let rules decide and let the model explain
When an output has to be auditable, I keep the decision deterministic and use the language model for the part it is good at: explaining. In ERIC Discoveries, a rule set flags the anomaly and the LLM writes the report. A client can ask why something was flagged and get the same answer twice.
3Validate before you narrate
A model should never be the source of a number. In PortfolioFit, every figure was typed, validated evidence before it reached a chart or a sentence. When the data was missing, the product said "I don't have that".
4Measure quality two ways
One metric is easy to fool. For ERIC Analyst I use user thumbs-up feedback and human expert review side by side. When the two agree, I trust the number. When they disagree, I have found something to investigate.
5Tune for the segment, not the average
A single global threshold hides the cases that matter. Our predictive maintenance model has a separate decision threshold for each vehicle model, because engines fail differently and the cost of a miss is not the same for every truck.
6Roll out in proportion to the risk
Low-risk capabilities go to everyone. Capabilities that touch client data go to a small group first, then widen as the quality numbers hold. Staged rollouts are slower on paper and faster in practice, because you do not have to roll anything back.
7Nothing ships without a "Go"
Agents can write the PRD, the code and the tests. A person approves the steps that matter. In Agentic Product Manager every release needs a human decision, and a verifier agent rechecks serious review findings before they reach that human.
8Build with the team
I prototype with Claude Code next to the developers. A working prototype settles a requirement discussion faster than a document does, and it tells me what is hard before I promise a date.
9Teach the people around you
An AI roadmap moves at the speed of the least confident team. I ran an eight-session programme for about 40 colleagues, and four working tools came out of it that I did not have to build alone.