Maya sat in the war room at 9:47 p.m., staring at usage data for her AI email triage feature. Two months ago: “Your inbox, now with 40% less noise.” Users opened it once, played with confidence sliders, then went back to manual sorting. Adoption: 17%. Time-to-inbox-zero went up 11% — users spent extra minutes second-guessing the AI.
Maya didn't fail at building AI. She failed at asking the right question. The job wasn't “triage with AI.” It was “quickly separate urgent from ignorable without missing anything that burns me.” Simple saved searches, shortcuts, and smarter snooze would have crushed 80% of the friction with zero hallucinations.
Start with the User, Not the Model
Friction Audit
Write the JTBD in one sentence, no tech words. Then list every painful step:
▸ Cognitive: “Which of these 47 threads is actually urgent?”
▸ Time: “Reading every subject line at 6:12 a.m.”
▸ Error: “Missing the one email from legal that costs $40K.”
▸ Emotional: “That sinking feeling when the badge hits 200.”
Only after this: Does AI actually remove a friction, or replace it with “did the AI get it right?”
User-First AI Scorecard
Score every AI idea 1–10. Total <32? No AI. 32–38? Rules first. 39+? Talk models.
| 1 | Can rules / UX solve 70%+?Rules win on determinism, speed, explainability. Maya's case: priority labels + keywords = zero model needed. | __ /10 |
| 2 | Does AI reduce real effort, or add novelty?$5 test: Would users pay $5/mo if 100% accurate? If “maybe for the wow” — kill it. | __ /10 |
| 3 | What breaks when it's wrong?Cost-of-wrong = frequency × severity. If >$10K/mo at scale, human-in-loop or walk. | __ /10 |
| 4 | Do we control data + tolerate non-determinism?If the job requires perfect consistency (finance, legal), AI needs heavy scaffolding. | __ /10 |
| 5 | Can we define success WITHOUT accuracy?Adoption, trust, net time saved, decisions improved — the only metrics that matter. | __ /10 |