30-Day Executive AI Sprint Checklist

No live model writes until this pipe and the fail-closed schema (2.2) pass. List CRM / billing / ERP endpoints for the target workflow. Prefer native API/webhook over Zapier on the critical path. Prove one E2E message with retry+idempotency on non-prod/shadow. Log drop/dupe behavior.

Failure mode: Data sticks in rate limits or drops silently: ops fixes it with another CSV.

Version a JSON schema (required/types/enums). Reject malformed payloads BEFORE the model with a reason code. Store schema with the integration. Add one regression each for missing field / wrong type / extra money field.

Failure mode: Invalid shape corrupts records; parsers break on the next model bump.

Name the operator queue. Target ~30s approve/reject with one click. Schema-pass+uncertain: queue; schema-fail stays fail closed. Time one dry run this week.

Failure mode: Edge cases stall or return to spreadsheets: the pilot works until the first weird invoice.

Configure exponential backoff and automatic failover queues: if model API latency exceeds 8 seconds or returns HTTP 429/5xx, immediately route payloads to the human fallback backlog without dropping state or silently retrying indefinitely.

Failure mode: API outage freezes mid-market billing run; silent drop causes 400 invoice disputes; ops resorts to manual midnight CSV exports.

Run the agent in shadow on the next batch of real transactions (target: 50 consecutive side by side with the human path). Log every mismatch: field, amount, timing, who would have caught it. No customer facing or GL write from the agent this week: compare only. Park repeats that fail the same way; fix schema/queue before adding volume.

Failure mode: Premature live deploy: errors land in customer or auditor sight; team reverts to FINAL_v12 and calls AI done.

Against the shadow log, score match rate and time to clear vs the human baseline from sprint_1_3. Keep the existing >= 98% accuracy gate. List variance types that must stay in the exception queue. Written go/no-go: live autonomy only if gate met AND queue clears exceptions without Excel rebuild.

Failure mode: Silent variance drifts into the production ledger while everyone watches dashboard green.

One page exception runbook (what lands in queue, who clears, approve/reject in one click, when to fail closed). Live dry run with named operators on real flagged items. COO / Head of Ops (or close owner) signs the SOP. Confirm exceptions clear faster than rebuilding the sheet. If not, stop widening scope.

Failure mode: Team abandons the tool because exceptions take longer than the old CSV path: close pack snaps back to FINAL_v12.

Map sprint work to the three stages already on the page: Preliminary (OpEx) · Application (CapEx window) · Post implementation (OpEx). Retain Never Automate sign off, versioned schema, shadow/verification log, operator runbook. CFO/controller confirms how YOUR auditors expect internal-use software documented. Golden Door does not invent a tax opinion. Separate API/token run cost (usually OpEx) from build documentation. Honesty: ASC 350-40 eligibility depends on facts and auditor judgment: this check is documentation discipline, not a promised CapEx outcome.

Failure mode: Custom agentic build expensed as pure SaaS OpEx with no stage file, or CapEx claimed without evidence.

Instrument continuous cost per transaction telemetry: track inference API usage and hosting OpEx against the baseline operator cycle time cost captured in sprint_1_3. Confirm marginal cost per transaction is under 5% of legacy human operator expense.

Failure mode: Unmonitored recursive prompt loops inflate monthly LLM cloud bills beyond the cost of human operators; initiative is cancelled as cost prohibitive.