Journey · information model · priorities · tradeoff
One small unit of knowledge, used everywhere
The whole journey rests on a single idea: a fact that knows where it came from, whether it’s been checked, and who may use it. That one model carries the review screen, the privacy controls, the draft, and the future agents.
The journey: four screens, with a pause button that’s always there
1Share
- Alex wants
- To start fast without writing a manual
- System does
- Reads a site, a doc, or a typed note. Slow reads don’t block.
- Recovery
- Unreadable page 3 → “Type it instead”
2Review
- Alex wants
- To be sure nothing wrong goes out
- System does
- Sorts 12 facts: 3 need a decision, the rest “look right”, private ones set aside
- Recovery
- Sources disagree → pick, edit, or leave open
3Saved
- Alex wants
- To know what’s kept and who sees it
- System does
- Shows saved / waiting / private, a visibility setting per fact, and what it powers
- Recovery
- Edit, remove, or add any fact later
4Preview
- Alex wants
- A reply they’d send
- System does
- Drafts from saved facts, numbers sources, hedges gaps
- Recovery
- “Fix this” beside the draft → it regenerates
Pause & return: “Save & finish later” is in every header. Reopening shows a welcome-back note with what’s still open. Open items never block saving or previewing.
Information model: the fact
Fact · example: “Where you work”
- Value
- Riverton, Oakdale, Fairview and Millbrook, plus within 25 miles
- Source(s)
- Website (checked today) · 2021 document (Mar 2021)
- Status
- Sources disagree → resolved by Alex → Looks right
- Visibility
- Can appear in replies · Team only · Only me
- Used by
- Reply assistant now. Booking agent later.
A fact reaches a customer reply only if its status is “Looks right” and its visibility is “Can appear in replies”.
Looks rightRead from a current source, or confirmed by Alex. Usable.
Needs a checkOnly in an old source, or low confidence (e.g. chimney sweeping). Not used until Alex answers.
Sources disagreeTwo sources differ. Both shown with dates. Not used until Alex decides.
MissingNeeded but not found (availability). The draft hedges. Alex can answer or say “later”.
Only meAuto-flagged as internal (crew cost, personal phone). Excluded from every reply.
Why this model: the same fields answer all five brief questions: what was understood (value + source), what’s saved and missing (status), who can use it (visibility), and how agents benefit (“used by” is just a readiness check against the same facts: the Booking agent shows “Missing: availability”).
Priorities: what’s in, and what I deliberately left out
Now · in the prototype
- Three input pathsWebsite, document, typed note. One is enough.
- Review by exceptionDecisions only where the answer changes what a customer hears.
- Per-fact visibilityThree plain levels, with sensible auto-private defaults.
- Draft that traces to sources and hedges gapsPlus in-place fix and regenerate.
- Approve before send, pause and return, edit later
Next · designed for, not built
- Calendar connectionTurns “I’ll confirm” into real slots. Needs availability to be trustworthy first.
- Prices & quotesAlready visible as “Can wait”. Unlocks the quote agent.
- Staleness nudges“Your service area was set 90 days ago. Still right?”
- Learn from sent repliesEdits Alex makes become suggested facts.
Deferred · and why
- Auto-send / autonomy levelsTrust has to be earned first. “Ask before sending” stays on.
- Agent configurationOut of scope. Agents appear only as a preview of what saved knowledge enables.
- Roles and per-person permissionsThree visibility levels cover Alex plus a team of five.
- Multiple businesses, versioning, bulk import
- Inbox integrationThe customer email is a sample.
The key tradeoff: show value with visible gaps, or finish the knowledge first?
The brief wants a useful result while building knowledge gradually. Those pull against each other.
A · Complete first
Require availability, service area and every flagged item before the first draft.
- Pro: the first draft is as correct as it can be.
- Con: up to a dozen questions before any payoff. That’s homework, and a likely drop-off.
B · Draft with visible gaps
Let Alex preview after saving, with unknowns hedged (“I’ll confirm”) and marked. Fixes happen beside the draft.
- Pro: value arrives early. Each fix visibly improves the reply, which motivates the next.
- Con (accepted): a hedged draft is less impressive, and Alex could approve it without fixing a hedge.
Safeguards that make B safe: hedges are amber and numbered “!”. The “Before you send” list names each one. The send dialog counts them (“2 lines say I’ll confirm”). Live availability is never promised.
How I’d validate it
Method: 5 moderated sessions with owners of 2–10 person service businesses using their own website and a real old document, then an unmoderated test of B vs. A with ~40 owners. Seed one outdated fact in each.
Time to first approved draftmedian ≤ 10 min (B), ≥ 25% faster than AIs early value actually quicker?
Seeded wrong fact corrected before send≥ 90%Does review catch what extraction got wrong, or do people rubber-stamp?
Sent with an unresolved “I’ll confirm” unchanged≤ 30%, and nobody surprisedDo hedges read as intended, or are they ignored?
“Would you send this?” (1–5)median ≥ 4 after fixesTrust in the result.
Can explain who sees the crew cost≥ 4 of 5Is visibility understood?
What would change my mind: if more than a third send hedged drafts without a thought, or fewer than 60% reach the preview, I’d make the service area a required answer (a gate on one fact, not all of them) and keep the rest of B.