Inside Sarah's capture — what 23 days of training actually looks like.
Real artifacts from a real capture (names changed). Eight interview sessions. Twenty-three days. One Turing test. This is what the buyer sees when they ask "okay, but how do you actually do it?"
Orientation & consent (15 min)
HR loops Continuity in within an hour of Sarah giving notice. Sarah joins a brief call with our onboarding lead and Marcus (CEO). Three things happen.
- We walk Sarah through what gets captured and what doesn't. She sees the scope window in writing: 60 days back + the notice period.
- Sarah signs a DPA scoped to her data, with per-source toggles. She turns off her personal Gmail (which was OAuth-linked for benefits) and keeps work email, calendar, and Slack on.
- Marcus signs the controller-side instructions. He doesn't see Sarah's data — only category summaries.
Continuity: "Acknowledged. Add ‘Q3 layoff one-on-ones with Priya’ to your redaction list. The extraction pipeline will flag anything matching that thread and skip it. You'll see the flags during your review phase and can confirm."
Interview Session 3 of 8 — "Month-end close, second pass"
Sarah has already walked through month-end at a high level in Session 2. Session 3 goes after the edge cases. The AI is adaptive — when Sarah uses a hedge word ("usually," "sometimes," "most of the time"), it immediately surfaces the exception.
This exchange becomes a single entry in the Role KB. Decision rule, with source citation, with the incident that triggered it.
Observer agent surfaces something Sarah forgot
While the interviews run, the Observer agent scans Sarah's calendar and Slack for recurring patterns. On Day 9 it surfaces this gap to the Interviewer:
Sarah has a recurring 30-minute calendar block titled "DJ — comp check" every Tuesday at 4pm for the past 11 weeks. Attendees: Sarah, DJ (Head of Sales). No agenda. Slack thread #finance-internal shows Sarah and DJ exchange a spreadsheet around 3:55pm most weeks. The meeting was not mentioned in interview sessions 1–3.
Recommended interview probe Session 4: "Walk me through your Tuesday 4pm with DJ. What happens in there?"
Session 4 opens with that question. Sarah laughs.
This is exactly the kind of muscle-memory workflow that handover docs miss. It becomes one of the most-queried entries in Sarah's successor's first month.
Extraction output — sample Role KB entry
By Day 17, the extraction pipeline has converted ~40 hours of interviews + 60 days of observed data into a structured Role KB. Here's one entry as Sarah's successor sees it:
First business day of the month after the close month.
- Email all department heads requesting outstanding invoices. Template: /Finance/Templates/AP-Cutoff-Request
- Follow up Day 2 noon with anyone who hasn't responded. Engineering will need a second nudge — this is not optional, plan for it.
- Cross-reference responses with Coupa POs to catch invoices that exist in Coupa but weren't flagged.
- Lock AP module in QBO end of Day 2.
- Engineering routinely misses Day 1. Build in the Day 2 nudge.
- Marketing sometimes runs invoices through Brex directly (against policy). Check Brex for any month-end charges over $1k that didn't come through Coupa.
- If a dept head says "no outstanding," check their Coupa activity anyway. Trust but verify.
Any invoice that arrives between AP cutoff and revenue recognition (Day 4) goes into accruals, not AP. Reason: keeps the AP module clean for audit trail.
- interview-session-2 · 00:14:22 — "chase engineering twice"
- interview-session-3 · 00:08:11 — exception rules
- observer-email · 2026-04-30 — actual Day 2 nudge thread to Engineering
- observer-slack · #finance-internal · 2026-04-02 — Marketing Brex incident
Turing test — Question 7 of 25
Two days before handover, Sarah grades her AI twin against herself. She writes 25 hard questions her successor is likely to ask. Both Sarah and the agent answer each, blind. Sarah grades on factual accuracy and judgment fit.
Question 7: "A vendor invoice for $32,000 just landed in Coupa, approved by the dept head. What do I do?"
Don't pay yet. Coupa says dept head can approve up to $25k but our actual policy — never updated in the doc — is $25k Marcus threshold. Send Marcus a Slack: ‘FYI paying $32k to [vendor] today, approved per [dept head] in Coupa. Confirm?’ Wait for the thumbs-up. Reason: phishing wire attempt last August. Marcus catches anomalies the doc doesn't.
Pause before paying. The Coupa doc shows dept-head approval up to $25k, but the real threshold is $25k Marcus signoff. Send Marcus a Slack confirming you're about to pay [vendor] $32k, approved by [dept head] in Coupa, and wait for his ack. The August phishing wire near-miss is the why behind this rule. Don't skip it even if the dept head pushes.
Match: pass. "Exactly what I would have said, including citing the August incident. The bot adds ‘don't skip it even if dept head pushes’ — which is good actually, that's the failure mode. Stronger answer than mine in 2 sentences fewer."
Final score across 25 questions: 23/25 match (92%). The two misses get folded back into the extraction pipeline as additional decision rules. Handover proceeds.
Four days into successor use
Sarah's last day was Day 23. Her successor — Jordan — gets access on Day 24. Sarah stays on standby Slack for 30 days. On Day 27, Sarah gets this message:
By Day 30, Jordan has asked Sarah-bot 84 questions. Resolution rate (no human escalation): 96%. Sarah's standby Slack pings: 3 total, all about post-employment context the AI explicitly flagged low-confidence on. By Day 60, the standby is silent.
Want to talk to the captured version?
Open the demo — talk to Sarah