Alex
Runs a home-services business, 5 employees. Comfortable with email and spreadsheets. Has never configured AI. Hates answering the same “do you cover X / when are you free” questions.
Problem framing · barriers · assumptions
Alex needs one reply they trust, not a configured AI
The brief asks for setup. Alex wants fewer repeated questions and has 10 minutes. So I treated setup as a by-product of getting a first reply Alex would actually send. The knowledge base grows from the corrections Alex makes while checking it.
How might we…get Alex to a customer reply they’d send within 10 minutes, so the business knowledge gets built by checking what the AI understood, not by filling in forms?
Who and what we know
Their information
- A website (current, thin on detail)
- A 2021 service guide (outdated service area, a discontinued-looking service, internal costs and a personal phone number)
- Things only Alex knows (usual hours)
Constraints
- 10 minutes, one use case, one business
- Extraction is slow and sometimes wrong
- Live availability is unknown to the platform
- Must be able to pause and return
What stands between Alex and a first win
Ordered by how much damage each does if ignored. Each barrier maps to a visible decision in the prototype.
| Barrier | What Alex is thinking | Design response | See it in |
|---|---|---|---|
| Wrong answers reach customers | “If it says we cover somewhere we don’t, that’s on me.” | Only checked facts feed replies. Unknown things become “I’ll confirm”, never a guess. Every sentence is numbered to its source. | Step 4 draft + sources |
| Private info leaks | “That document has what I pay my crew.” | Internal-looking details are auto-set to Only me and shown as “Kept out”. The send dialog repeats what’s excluded. | Step 2 · Step 4 |
| Setup feels like homework | “I don’t have time to write all this down.” | One source is enough. 4 screens. Only 3 of 12 facts ask for a decision. Prices are parked under Can wait. | Step 1 · Step 2 |
| Blank page and jargon | “What’s a knowledge base? What do I even type?” | Start from things Alex already has (URL, file) or a few typed lines. Plain words: “what I know”, “needs your eye”. The term “knowledge” appears once Alex has some. | Step 1 |
| Sources disagree | “Which one is right? I’m not sure the doc is current.” | Show both side by side with dates, suggest the newer, Alex decides. The suggestion is pre-selected so it’s one click. | Step 2 conflict card |
| Losing control | “I don’t want it emailing people on its own.” | Approve & send shows the exact email and recipient. Auto-send does not exist yet, and the screen says so. | Step 4 dialog |
Three principles that drove the design
1 · First value before completeness
Alex sees a real draft after checking a few things, not after the knowledge is “done”. Missing facts show up as honest gaps in the draft, and Alex fixes them there.
2 · Nothing is used until it’s checked
Extraction is a proposal. A fact needs a clean status and a “can appear in replies” setting before it can reach a customer.
3 · Gaps are shown, not guessed
Unknowns are visible (“Missing”, “I’ll confirm”) and fixable in place, and fixing one changes the draft right away. That’s what makes the knowledge feel worth building.
Assumptions I’m making, and how risky they are
| Assumption | Risk if wrong | How I’d find out |
|---|---|---|
| A1 · Alex will really read the “Looks right” list rather than rubber-stamp it | High Wrong facts get saved | Seed one subtly wrong fact; see if it’s caught |
| A2 · A draft with “I’ll confirm” lines still feels like value | High Whole approach depends on it | Ask “would you send this?” before and after fixing |
| A3 · “Can appear in replies / Team only / Only me” is understood | Medium Private info mis-shared | Ask Alex to explain who sees the crew cost |
| A4 · Auto-flagging private details is mostly right | High A missed one is a leak | Test on real owner documents; track misses |
| A5 · Typical hours are enough to answer “when are you free?” when paired with “I’ll confirm” | Medium Customers expect real slots | Compare customer reaction to draft vs. calendar-backed reply |
| A6 · Owners will come back to update knowledge | Low for v1 Knowledge goes stale | Track returns within 2 weeks; add a nudge if low |
What “good” looks like
≤ 10 min
Median time from start to first approved draft.
100% caught
Of the seeded wrong or outdated facts (the old service area), corrected before anything is sent.
0 leaks
Private-marked items appearing in a customer-facing draft.