Your first AI agent failed. Write the post-mortem six months early.
Copy a 10-section failure review for a Business Central agent and fill it in before go-live. Each blank you cannot answer is a gap to close while it is still cheap: an owner nobody named, an approval queue nobody reads, a baseline nobody recorded.
No sign-up. No email. Nothing you type leaves your browser.
How to use it
- Date the failure. Pick a date six months after go-live and write in the past tense. "It is 14 May. The agent failed." Past tense forces a specific story.
- Fill every section. Write "unknown" where you do not know. Each "unknown" is a launch blocker for its owner.
- Get three names on it. The decision owner, the approver and a finance reviewer. Review it again on the date you wrote at the top.
The template
Plain text. Paste it into a document, a ticket or a wiki page. Square brackets mark what you replace.
PRE-LAUNCH POST-MORTEM Agent: [name and the one decision it makes] Written on: [today's date] Failure date (6 months after go-live): [date] Decision owner: [one named person] Approver: [name] Finance reviewer: [name] 1. HEADLINE It is [failure date]. The agent [what it did wrong] and [who or what it hit]. In one sentence: 2. WHAT THE AGENT DID Action taken: [e.g. changed a reorder quantity, released a purchase order, posted an invoice] System and record changed: [Business Central table, document or field] Did a human approve it before it ran? [yes / no / approved without reading] 3. WHO OWNED THE DECISION Named owner on launch day: [name, or "nobody"] Backup when the owner is out: [name, or "none"] Who was told, and how: [Teams card, email, dashboard, nobody] 4. THE APPROVAL QUEUE Approvals per approver per day at launch: [number] The number one approver reads properly in a day: [your budget] What reached the approver first: [the highest-value item, or whatever arrived first] 5. MONEY Cost of this failure: [amount and how counted] Baseline recorded before launch: [yes, where / no] Monthly AI spend, including credits and review time: [amount] Monthly rework: [AI decisions overridden or reversed x cost to undo] 6. DETECTION How we found out: [who, how, when] Time from failure to detection: [hours or days] Could we reconstruct why it acted, from inputs to approver? [yes / no / partly] 7. THE STOP RULE What would have made us stop the agent earlier: [metric and threshold] Who has the authority to stop it today: [name] 8. UNDO How the action is reversed: [steps] Time to reverse it: [hours or days] What cannot be reversed: [e.g. a payment released] 9. WHAT WE CHANGE BEFORE LAUNCH 1. [change] Owner: [name] Due: [date] 2. [change] Owner: [name] Due: [date] 3. [change] Owner: [name] Due: [date] 10. SIGN-OFF Decision owner: [name, date] Finance reviewer: [name, date] Next review: [failure date]
Which section catches which failure
Agent failures cluster at the five stages of a decision: Signal, Route, Approve, Execute and Audit. The template maps to them, plus one cost check that sits outside the model.
| Failure | Template section | Question it forces | Tool |
|---|---|---|---|
| Alert fired and nobody owned it | 3. Who owned the decision | Who was named on launch day, and who covered when they were out? | Decision Latency Diagnostic |
| Approvals clicked, not read | 4. The approval queue | How many approvals reached each approver, against how many they read? | Approval Budget Calculator |
| Agent wrote the wrong record | 8. Undo | What reverses it, how long does that take and what cannot be reversed? | AI-to-Operations Checklist |
| Nobody could say why it acted | 6. Detection | Does the log tie inputs to an approver, or only record that the task ran? | Decision reasoning capture |
| Cost passed the saving | 5. Money and 7. The stop rule | Was the baseline recorded before launch, and what threshold stops the pilot? | AI ROI Calculator |
Common questions
Questions about the pre-launch post-mortem
What is an AI agent post-mortem written before launch?
A review of a failure that has not happened yet. You write it as if the agent already failed six months after go-live, then fill in what broke, who owned it, what it cost and how you reversed it. Every blank you cannot fill is a gap to close before launch.
Who should fill in the template?
The named owner of the agent's decision writes it, with the approver and a finance reviewer. If nobody can be named as owner, that is the first finding. Do not let the vendor or the implementation partner fill it in alone, because they will not carry the cost of the failure.
Does the template contain benchmarks or statistics?
No. It is a blank structure. Every number in it is yours: your baseline, your spend, your approval volume. Where a number needs a method, the template points to a calculator that runs in your browser.
Do you store the template or ask for an email?
No. The page is ungated. The copy button puts the template on your own clipboard and nothing is sent to IntelliConnectQ.
A filled template shows the gaps. The Diagnostic scores them.
Sections 3, 4 and 6 describe the Route, Approve and Audit stages of your own decision path. The Decision Latency Diagnostic scores all five in 12 questions and shows which stage holds the longest wait.
Named owners, approval thresholds and an audit trail on every decision are decision infrastructure. For the pilot-level stop rule behind section 7, read AI pilot stop criteria.
Run the Decision Latency Diagnostic