Our agent sent a note that no person had reviewed.
On 5 September 2026, one of our agents sent an end-of-day note. Its footer said it was an automated note from an AI. It even cautioned against drawing much from its small sample. There was nothing hidden about how the text had been produced. What the footer could not tell the recipient was whether another person had checked it. Nobody had.[S1]
The note went to the person responsible for that agent, so the recipient could make that judgment themselves. Now consider the same handover to a client. This is the practical question raised by the case, not a second reported incident: who has checked the figures, understood the recommendation and decided that the document is ready to use?
That is where “AI-generated” stops helping. It identifies the source of the draft. The team still needs to make a separate decision about who stands behind the work and who can put it right.
The document should name the person who actually checked it.
Our published proposal adds two fields to a document: who prepared it, and who reviewed it. The second names a person, their role and the date. Until the check happens, that field stays open. The graphic above shows these two states as a template, not as evidence of a completed client review.[S1]
For an agent that drafts proposals, decide first who owns the outcome. That person arranges review and handles exceptions. A colleague with the right expertise can perform the check, but the team should know whether that colleague can also approve the document for use. Keep those responsibilities explicit instead of assuming that everyone understands the same thing by a signature.
The record below separates four questions that a single “done” status can hide. It is our suggested template; adapt the approval step to the document and its consequences.
| Question | What to record |
|---|---|
| Who owns this work? | The human owner and their deputy, agreed before the agent starts. |
| Which draft are we discussing? | The preparing agent, document version and preparation date. |
| Who actually checked it? | The reviewer, role, reviewed version, date and any unresolved issues. |
| May this version be used? | The approval decision and person authorised to make it, where approval is required. |
A review of version one must not silently become a review of version two. If the agent changes a figure afterwards, the record should make clear whether that change has also been checked. This is a practical addition to the two-line label, not a claim about a system we have already deployed.
A reviewer needs the means to challenge the draft.
Imagine asking someone to check a proposal while withholding the meeting notes it was based on. They can read the prose, but they cannot assess whether it reflects the brief. The same problem appears when they can spot a mistake but cannot stop the document being sent.
The Spanish data protection authority, AEPD, describes meaningful human intervention through authority, relevant knowledge, independence and diligence, and the practical means to intervene. Its guidance concerns automated decisions; we use those criteria here as a prompt for reviewing documents, not as a legal compliance test.[S2]
Before assigning the reviewer, ask whether they can change or stop the result, understand the subject and the system’s limits, and question it without pressure to approve. Then check the practical conditions: access to the relevant evidence, usable tools and enough time for this particular document. Time is one part of those means. A name in a field cannot supply any of them.
An empty reviewer field needs a decision from the owner.
Suppose a proposal is due to go out and the review is still open. Do not have the agent fill in its owner’s name to make the document look complete. Ask the owner to arrange a qualified reviewer, change the timing or narrow what can be sent under the agreed process. Decide the release conditions in advance for work where an error would have serious consequences.
This is a recommendation for designing the handover. It is not how every document at MING already works. In the source published on 23 September, the two-line label was in use for The Hybrid Company publication, but had not been rolled out across client deliverables. In the accompanying LinkedIn discussion, the author also explains that an empty field was not yet a general shipping block.[S1][S3]
That distinction matters when you try the method. Count the documents with missing review separately from those that were reviewed. Investigate why the field is empty before introducing a target that rewards people simply for filling it in.
After a mistake, use the record to organise the correction.
When a wrong figure reaches a recipient, the human owner should have a route to the affected version, its reviewer and the action taken with it. Our proposed first step is to limit further use where appropriate, arrange the correction and involve the person who can change the underlying workflow. The point of naming people is to make that response possible, not to assume the reviewer alone caused the failure.
Connect the document record to the agent’s audit trail. That answers the next question: what did the system actually do with the work? A review and a successful delivery are separate events. Both need their own evidence.
Test the template on one document type for a week.
Choose a document your agent already prepares and a person already checks. Agree who owns the outcome, who can review and approve it, and what happens when they are unavailable. Add the preparation and review fields to that template. After a week, inspect the underlying records and measure the time the checks actually took. A completed field is the beginning of that inquiry, not its conclusion.
If the review load is too high, the answer may be a smaller task or a different workflow. Our guide to an agent’s running costs shows how to include that human work.
When that person changes, use our agent handover guide to check the continuing task, knowledge and connections.
The proposed review method has not been shown to reduce errors.
Evidence and limits
- The opening comes from a published MING account of one agent note. Its underlying records are private.
- The two-field label is a proposed working practice with reported adoption in one publication as of 23 September 2026. It is not a verified firm-wide rollout.
- The source explicitly says the effect of a named reviewer on error detection has not been tested. We claim no reduction in errors.
- Version, approval and incident-handling suggestions on this page are our practical extensions. This guide does not allocate legal liability or establish regulatory compliance.
