Use AI where it earns the trust. Don't use it anywhere else.
Where AI earns its keep, where it doesn't, and the honest tradeoffs the pitch decks skip past.
AI has been the easiest thing to oversell I’ve ever seen. Even tools I love — especially tools I love — get pitched in ways that overstate what they can do, hide their failure modes, and skip the inconvenient bits.
That’s a problem for the buyer, and it’s a problem for the field. The teams that lose trust in AI today are slower to adopt the good AI work that’s coming. I’d rather be honest now and earn the option to do more of this work later.
What “integrity in AI” looks like, on a project
Saying no to a feature. If the AI piece of a proposal won’t reliably do the thing the client thinks it’ll do, I tell them — even when it shrinks the engagement.
Failure modes on the table. Every AI feature has a class of inputs where it’ll be wrong, slow, or weird. Those classes get named, monitored, and surfaced in the UX. Nobody should be surprised by them three months in.
Evals before launch. A feature without an eval suite is a feature shipping on vibes. The eval suite is the thing that lets the team know — quickly, cheaply — when a model update changes behaviour.
Cost ceilings, hard. Every AI feature ships with a spend cap and a real alert. The team shouldn’t be wondering on the 28th of the month whether this month’s OpenAI bill is going to make their CFO miserable.
Human-in-the-loop where it matters. Letting an LLM autonomously do something that affects a customer’s money, contract, or trust, is a call to make carefully — not by default.
Privacy and data lineage are not afterthoughts. Where is the prompt going? Where’s it being logged? What’s it learning from? These are answered before we wire anything up.
What I won’t do
- Build a chat widget on the homepage because “everyone has one.”
- Bolt AI to a step in the workflow that worked fine without it.
- Pretend an LLM is deterministic to make a pitch easier.
- Ship an AI feature without observability and an off-switch.
The upside of being honest
The teams I’ve done my best AI work with are the ones who agreed up front that some experiments would fail and that failing them quickly and cheaply was the whole point. That posture is how AI work actually compounds in a business — instead of producing one feature, a press release, and a quiet retreat.