AI
Shipping LLM features that survive contact with real users: evals, agents, retrieval and restraint.
AI answers vs good navigation: choose deliberately
Replacing your site's navigation with an AI answer box is rarely the upgrade it looks like. A decision framework for AI search vs structured browse, honestly measured.
Disclosure patterns: telling users when a machine is speaking
AI disclosure that survives legal review and user testing: label placement, export watermarking, model transparency, and wording that informs, not alarms.
Shipping 'sometimes wrong': communicating AI features
Launch copy for features that are sometimes wrong: setting accuracy expectations, changelog phrasing, support macros, and the trust maths of underpromising deliberately.
Our responsible-AI review, in the open
The full review we run before any AI feature ships: risk taxonomy, red-team scripts, disclosure patterns, data-handling checklists — and how we say no to features.
RAG and the staleness problem: freshness is a ranking signal
Stale retrieval makes RAG answers confident folklore. Treating freshness as a ranking signal, surfacing 'last updated' in answers, and designing the re-indexing pipeline.
AI triage for support: routing, not deflecting
AI support triage should route customers, not wall them off. Intent detection that errs toward humans, handoff packages, and the deflection metric trap.
Human-in-the-loop: designing the review queue
Human-in-the-loop AI fails at the queue, not the model. Triage design, keyboard-first reviewer ergonomics, feedback loops into evals, and honest staffing maths.
Pricing AI features: cost curves, credits and honest limits
How to price LLM features without burning margin: cost per action as a unit-economics input, credits vs guardrailed unlimited, fair-use caps, tiering and drift.
Onboarding for AI features: sell the p50, not the demo
AI onboarding fails when it promises the demo's best-case output. Design first-run flows around median performance — calibrated examples, honest scope, progressive trust.
Swapping the model without breaking the product
Model migrations without breaking the product: shadow runs, eval-gated cutovers, tone and format regression diffs, and rollback plans for behaviour you can't unit test.
Citation design: making RAG answers checkable
Citations only work if they're checkable: claim-to-chunk mapping, snippet fidelity tests, anchor UX, and what to do when the source contradicts the summary.
Prompt libraries that survive the team
A well-structured prompt library still dies if nobody owns it. The team side: ownership, review rituals, discoverability, drift control and reuse metrics.