AI where it earns its keep
A practical filter for deciding when an LLM belongs in the loop — and when it's just a liability with good PR.
There's a version of AI that's a feature in a roadmap and a version that's a line in a fundraising deck. The difference matters, because shipping the second one as if it were the first is how products get slower, more expensive, and less trustworthy while looking more modern. AI belongs where it earns its keep — and nowhere else.
Two honest modes
In practice, useful AI shows up in two shapes. The first is a custom AI system: the model is doing real work the product couldn't do otherwise — extracting structure from mess, ranking, generating a first draft at scale. The second is LLM-in-the-loop: the model assists a person who stays in control, proposing rather than deciding. Both are legitimate. What they share is that the AI is load-bearing. The trouble starts when it's decorative.
The test
Before adding a model, we ask one question: what does this do that a simpler thing couldn't? If the honest answer is "nothing, but it sounds impressive," that's not a feature — it's a liability with good PR. A deterministic rule that always works beats a model that mostly works, especially anywhere correctness is the point. AI should be the answer to a real constraint, not a reflex.
Be honest about the edges
Every model has a failure mode, and pretending otherwise is how you lose a user's trust in one bad output. Designing with AI means designing for when it's wrong: surfacing low-confidence results, keeping a human in control where the stakes are high, and never presenting a guess as a fact. Being honest about the edges isn't a weakness in the pitch. It's the thing that makes the product safe enough to rely on.
The point
AI is leverage when it's pointed at a real problem and risk when it's pointed at a headline. The discipline is knowing which one you're doing — and being willing to leave it out when the simpler thing is better. That's what "AI where it earns its keep" actually means.