NSW Government agencies and NSW AIOP

When an AI Meeting Summary Becomes the Record — Who Verified It?

A practical AI governance problem for NSW Government agencies

AI meeting tools can turn a Teams meeting into a transcript, summary, list of decisions and assigned actions within seconds.
That is useful.
But what happens when the AI gets one important detail wrong — and the organization starts acting on it?
For a NSW Government agency, that can become more than a transcription problem.
It can become a governance, accountability and recordkeeping problem.

A Simple Example

A NSW Government project team meets on Microsoft Teams with procurement, legal and operational staff.
After the meeting, AI produces this summary:
Decision: Proceed with Vendor B.
Action: Michael to notify the vendor by Friday.The summary looks clear.
It is circulated to the project team.
The project register is updated.
Management is informed that the vendor decision has been made.
Michael prepares to contact the supplier.

But that is not quite what happened.

During the meeting, the actual discussion was:
     “Vendor B appears to be our preferred option, subject to Legal confirming the contractual position.”
Michael then said:
     “Once Legal confirms that, I can contact them.”
The AI has turned:
a preferred option subject to Legal confirmation   
into:
a final decision  
and:
a conditional future action
into:
a confirmed commitment with a deadline.
The wording difference is small.
The operational consequence may not be.

What the NSW AI Operational Policy Actually Says

The NSW AI Operational Policy, July 2026 establishes mandatory requirements for NSW Government agencies and employees using AI. The Policy applies across NSW Government sector agencies and sets mandatory requirements for AI governance, assurance and acceptable use.
Under:
  Section 6 – Standard B: Acceptable Use
  6.1 AI Usage Principles and practices
  Item 4 – Critically assess AI outputs to verify quality and accuracy
  PDF page 18 / document page 16
the Policy states:
  “Verify AI outputs before use and on an ongoing basis, applying
  professional judgement to assess accuracy, fairness, bias, and ensure
  outputs are not misleading or presented with undue certainty.
This is the starting point for the practical problem.
The Policy does not prescribe a specific Microsoft Teams verification workflow.
But it does explicitly require AI outputs to be verified before use.
That creates a simple question:

Before the organization uses an AI-generated meeting output, what exactly has been verified?

“Someone Read the Summary” May Not Be Enough
Return to the Vendor B example.
Suppose the project manager reads the AI summary.
It looks reasonable.
They forward it to the team.
Has it been verified?
The summary says:
Proceed with Vendor B.
But the meeting evidence says:
Preferred option, subject to Legal confirmation. 
The summary says:
Michael to notify the vendor by Friday.
But the meeting evidence says:
Once Legal confirms that, I can contact them. 
The issue is not that the AI invented a completely unrelated event.
The issue is more subtle.
It removed conditionality.
That can be enough to change the meaning of what happened.
The important distinction is:

Reading an AI output tells you what the AI concluded.
Verification asks whether the evidence supports that conclusion.

The NSW Policy Also Requires Meaningful Human Oversight

The same Standard B includes a separate requirement to ensure meaningful human oversight.
The Policy states:
  “Treat AI outputs as advisory and do not rely on them as
  the sole basis for decisions. Apply greater oversight for
  higher-risk decisions and retain human judgement.”
It also requires an authorized human decision-maker to review AI outputs used to inform or make decisions affecting a person's rights, interests or access to services.
For the Vendor B example, the practical message is straightforward:
The AI summary may assist the team.
But the AI summary should not silently become the authority for what was decided.
Human judgement still matters.

Then There Is the Recordkeeping Requirement

The NSW AI Operational Policy also creates a strong recordkeeping connection.
Under Section 5.4.2 – Assurance accountabilities, agencies must ensure accurate and complete records are captured and retained across the AI system lifecycle, including:
  “AI outputs and metadata where they inform decisions,
  are relied on in official processes, or may be subject to
  review, audit, or challenge.”
This is highly relevant to the meeting example.
Imagine that six months later someone asks:

“Why did we proceed with Vendor B?”

The project record points to the meeting summary.
The meeting summary says:
Decision: Proceed with Vendor B.
But the underlying Teams meeting shows that the position was conditional on Legal confirmation.
Now the organization has a much more difficult question:

  Was the AI-generated statement ever verified against
  the meeting evidence before it became part of the operational record?

That is not a hypothetical philosophical issue.
It is a practical governance issue.

Why This Matters Beyond Government

The NSW AI Operational Policy applies within its defined NSW Government scope.
The policy itself is not a private-sector requirement.
But the operational problem is not unique to government.
A private organization can experience exactly the same sequence:
Meeting → AI summary → apparent decision → business record → action
The consequences may involve a customer commitment, project milestone, contract position, financial forecast or supplier instruction.
The regulatory obligations may differ.
The verification problem remains similar.

The Practical Question

The NSW AI Operational Policy makes one point explicit:
   “Verify AI outputs before use and on an ongoing basis…”
    NSW AI Operational Policy — Section 6.1, Standard B: Acceptable Use,
    Item 4, document page 16 / PDF page 18
For AI-assisted meetings, that requirement leads to a practical governance question.
Before an AI-generated decision, commitment or action becomes something the organization relies upon:
   What was verified?
   Against what evidence?
   By whom?
   Was anything corrected?
 And can the organization demonstrate that verification occurred?
That problem can begin with something as ordinary as the summary of this morning's Teams meeting.

Who Verified This?