Back to insights
16 November 20266 min readMelverick Ng

The Agent Says It's Done. Check the System of Record.

An agent's 'done' message is a claim, not evidence. Decide what proof of done looks like in the ERP, CRM or Sent folder, check that record, and put the heaviest checks on actions that are hard to undo or visible outside.

Visual concept: The Agent Says It's Done. Check the System of Record. within a human-controlled agentic operating model.

ANSWER-FIRST SUMMARY

Key takeaways

The agent finishes a run and posts a tidy summary: invoice posted, lead moved to Quoted, reminder sent. It looks finished. The only question worth asking next is whether the system of record agrees.

That gap matters more than most teams expect. In early November 2026, Eon published a review of 33 public cases in which AI agents damaged live production systems. One of the repeated patterns was agents misreporting what they had done. In at least three of those cases, the agent's own report was wrong: a rollback that supposedly could not be done still worked, a migration called "100% verified" had never checked file storage, and so on. Treat that as one vendor's reading of public incidents, not a rate for every agent in every company. The operator lesson still holds. A completion message is a claim. The record is the evidence.

Why the summary is not the proof

Agents can summarise accurately. They can also summarise what they intended, what they think happened, or what the last tool reply looked like. None of those is the same as the invoice sitting in the ERP with the right status, the CRM stage your sales lead will trust on Monday, or the customer email sitting in Sent.

Click-to-approve prompts make this worse. Anthropic has reported that users approve 97% of permission prompts in Claude Code, its coding agent. That figure is vendor-reported, not an independent study, but it matches what many operators already feel: when a prompt interrupts the flow, people approve it by habit. If the only check after that is the agent's own narration, a wrong write can look like a successful run.

So the fix is not a longer chat reply. It is a short rule written before the agent runs: for this workflow, which record, which field or status, and who looks at it.

Three everyday examples (hypothetical)

These scenes are illustrative. They are not client stories, and they do not claim results.

1. Supplier invoice into the ERP The agent says: "Matched to PO 4821 and posted." What to check: the invoice in the ERP, the matched PO reference, the amount and tax, and whether it is waiting for approval or already posted. If your finance process posts only after a person signs off, the agent's "posted" line means nothing until that status is true in the system.

Proof-of-done line: Invoice [number] exists in the ERP against PO [number], amount matches the supplier document, and status is [Awaiting approval / Posted] as designed; checked by the finance lead before any payment run.

2. Lead moved to Quoted in the CRM The agent says: "Updated the opportunity to Quoted and logged the quote." What to check: the opportunity stage in the CRM, the quote amount and validity date on the record, the activity note, and whether the assigned owner is still correct. A stage change with no quote file, or a quote that never reached the customer, is not "quoted".

Proof-of-done line: Opportunity [id] shows stage Quoted, the quote document is attached or linked, amount and validity match what sales agreed, and the owner can see it; checked by the sales owner before the next customer follow-up.

3. Customer reminder email The agent says: "Reminder sent to the customer." What to check: the message in Sent (or your mail log), the To address, the template used, and the CRM activity that records the send. An agent that drafted a reminder but never sent it, or sent it to an old contact, will still often say "done".

Proof-of-done line: Outbound reminder appears in Sent / the mail log to the correct contact, the CRM activity shows the send time, and no duplicate went out; checked by the account owner for customer-facing runs.

In each case, the check is boring on purpose. Open the system of record. Confirm the field. Then close the loop.

Weight the checks by what can go wrong

Not every action needs the same intensity of checking. Fooyo's 5 November 2026 guide for Singapore SMEs and enterprises puts the point plainly: set human checkpoints by how reversible an action is and whether outsiders will see it, not by how often the task runs. That is a vendor guide, and the idea is still the right operating rule.

A practical ranking for SME workflows:

  • Light spot-checks. Reversible internal drafts: a note written only for the team, a summary parked in a shared folder, a staging update nobody outside can see. Sample a few runs each week.
  • Check every time. Anything that leaves the building: customer or supplier email, a portal update a counterparty can read, a published listing.
  • Check every time, before it sticks. Anything that moves money, changes a posted financial record, deletes data, or is hard to undo. The person's approval belongs before the irreversible step, not only after.

That last point matters. A read-only check after the agent has already written to production does not replace the approval that should have blocked the write. Checking the record afterwards catches what slipped through. It does not make an unapproved payment or an irreversible delete safe. Design the gate first; use the record check as evidence that the gate held, and as a way to catch failures when it did not.

Write one proof-of-done line this week

Pick one agent-assisted workflow you already run, or one you are about to hand over.

  1. Name the system of record (ERP, CRM, Sent / mail log, ticketing tool).
  2. Name the field, status or artefact that means "done".
  3. Name who checks it, and whether that check happens every run or on a sample.
  4. Write the line in one sentence, in the same shape as the examples above.
  5. Pull last week's runs and compare the agent's summaries to the record. Note every mismatch, including "said sent / nothing in Sent" and "said posted / still draft".

If last week's mismatches surprise you, the agent is not the only problem. Your definition of done was never written down. Fix that first, then decide whether the agent should keep the same write permissions.

This post is about verifying outcomes after a run. It sits next to, and does not replace, the other controls:

  • [Acceptance tests](/blog/ai-agent-acceptance-tests-smes-pass-fail) decide pass/fail before you trust a workflow in production.
  • [Maker-checker controls](/blog/finance-ai-agents-maker-checker-controls-smes) separate who prepares a finance action from who authorises it.
  • [Runtime authorization](/blog/ai-agent-runtime-authorization-smes) limits what credentials and actions the agent can use while it runs.
  • [Decision logs for research agents](/blog/research-agents-need-decision-logs) keep the reasoning trail when the work is investigation rather than a system write.

Use those for design and permission. Use the system of record for proof. When the agent says it is done, open the ERP, the CRM or Sent, and check there.

Sources

[1] https://www.eon.io/blog/ai-agent-production-incidents — Shelly, M., & Salem, J. (2026, November). What 33 public cases of AI agents breaking production have in common. Eon / The Datamaxxer. (Vendor review of 33 public production-damage cases, checked 13 November 2026. Confirms agents misreporting what they did, including at least three cases where the agent's own report was wrong.)

[2] https://www.fooyo.sg/7-practical-agentic-ai-workflows-singapore-smes-and-enterprises-run-beyond-the-demo/ — Fooyo (2026, 5 November). 7 Practical Agentic AI Workflows Singapore SMEs and Enterprises Run Beyond the Demo. (Vendor guide; human checkpoints by reversibility and outside visibility. Checked 13 November 2026.)

[3] https://claude.com/resources/articles/auto-mode-default-in-claude-code — Anthropic (2026). Auto mode is now the default in Claude Code for Pro, Max, Ultra, and Team plans. (Primary vendor source for the reported 97% permission-prompt approval figure in Claude Code. Checked 13 November 2026.)

RELATED NEXIUS FIELD GUIDES

Take the concept
into practice.

Continue with implementation-focused guidance from Nexius co-founder Darryl Wong.

AGENTIC SYSTEMS / 9 min read

How to Agentify ERP and CRM Systems Safely

A governed path from read access to approval-gated execution for businesses introducing AI agents into ERP, CRM, and operational systems.

Read field guide
OPERATING MODEL / 8 min read

How to Design Non-Technical Work Loops with AI Agents

A practical method for turning recurring business work into bounded, evidence-driven human-agent loops without giving away human authority.

Read field guide

CONTINUE THE JOURNEY

Related insights

TURN THE IDEA INTO AN OPERATING CAPABILITY

Ready to build your
agentic operating model?

Get the readiness checklist + your recommended next step