The number one question every ops leader asks: what if the agent gets it wrong?Fair question. Here's exactly how exceptions work — the threshold, the escalation UX, the override workflow, the weekly review, and the correction feedback loop. Nothing hidden.
| Agent | Process | Steady-state | Week 1 |
|---|---|---|---|
| Bookkeeping agent | Invoice categorization | 1–2% | 8–12% |
| Recruiter agent | Outreach approval | 3–5% | 20–30% |
| Inbox agent | Reply approval | 2–4% | 15–25% |
| Voice agent | Call classification | 0.5–1.5% | 5–8% |
Week 1 rates are elevated by design — the agent surfaces borderline cases to calibrate against your corrections. By week 3–4 it settles to the steady-state range as the RAG index and threshold tuning converge.
Claude returns a structured confidence value on every classification, extraction, or drafted output. Below your configured threshold (default 80–90% depending on agent) the action does NOT execute — it becomes an exception.
You get a message on your phone: agent name, action proposed, confidence, key context (vendor history, prior threads, candidate profile). Reply 1 to approve, 2 to edit, 3 to skip. Latency: sub-second.
The override — approve, edit, or skip — writes to the audit log with your identity, the original agent proposal, and your final decision. Edits feed back into the agent's context so it gets closer next time.
Every Monday you get a report: exception count, override rate, top exception patterns, week-over-week trend. We use it as the input to weekly agent tuning — adjust thresholds, add examples, refine prompts.
🤖 Bookkeeping agent — needs review Invoice from: Acme Corp Amount: $12,450.00 Proposed GL: 5240 (Software) Confidence: 76% (below 90% threshold) Reason: Amount is 34% higher than typical Acme invoice ($9,300 avg). No matching PO found. Reply: 1 = approve as proposed 2 = edit (I'll ask for corrections) 3 = skip — don't post 4 = show me more context
Every escalation carries: the proposed action, the confidence score, the reason it's uncertain, and 4 numbered options. Median time to decide: 8 seconds.
Your reply is captured in the hash-chained audit log with your identity, the timestamp, and the original proposal. Editors feed back into the agent's RAG index — the next similar invoice from Acme Corp will use your corrected GL coding.
If you don't reply within your configured SLA (default 4 hours), the action is auto-skipped and re-queued for next-day review. Nothing silently proceeds.
We can send a demo exception via WhatsApp during the audit call. You'll see the exact UX before you commit to a single dollar of outcome billing.
Book the audit →