Skip to article
LOOKUP

Where should an AI assistant stop?

Conceptual illustration of a pause for human approval. This is not a product demonstration.

By Peter Kantarelis | The AI IT Guy | Approximately five minutes

An AI assistant can prepare a useful follow-up and still need someone to decide whether it should be sent.

I'd want that decision made clear before connecting it to the business.

In my recent articles, I've written about understanding a job and giving an assistant the right information to help with it.

The next question is what happens when its preparation is ready to become an action.

Sending an email or changing a customer record can affect someone beyond the person using the assistant.

We need to decide who can authorise that step and what, exactly, they are approving.

A request isn't an agreed commitment

Take an illustrative example of a team following up after a customer meeting.

The customer asks whether the work could be completed by Friday.

The person in the meeting says they'll check with delivery.

An assistant prepares the follow-up and writes, "We'll have that completed by Friday."

The sentence is clear, but it changes an unresolved request into a commitment.

I'd want the preparation to preserve the difference.

The meeting record could say the customer requested Friday and delivery hasn't confirmed it.

The draft email could explain that the team is checking the date.

The proposed next action could be to ask the person responsible for delivery to confirm what is possible.

Those are useful things to prepare without deciding the answer on that person's behalf.

Approve the action someone will actually take

"Looks good" can mean several things.

Someone may be satisfied with the summary while still needing to check the proposed date.

They might approve the wording for one recipient without agreeing to send it to everyone copied on the earlier conversation.

I'd make the important approval specific enough to inspect.

For this example, the reviewer needs to see the exact email, who will receive it and the commitment it contains.

The person approving the date also needs the authority to make that commitment.

An assistant shouldn't have to guess whether approval of a summary also permits it to send the email or update the customer record.

If the wording changes after approval, I'd want the changed version reviewed before it leaves.

The approval should belong to the action we examined, rather than whatever the assistant happens to prepare next.

Make the software match the decision

Writing "ask me first" in an instruction helps explain the intended behaviour.

I'd also check what the connected software actually allows.

If the assistant only needs to prepare a draft, why give that connection permission to send?

If we've decided a person must approve a change, where is that approval checked before the change happens?

These are questions for the actual system, including the account used by each connection.

Microsoft's Copilot Studio documentation, for example, describes tools using either the user's authentication or credentials supplied by the agent's author.[1]

That choice is part of what needs examining when an assistant connects to another service.

Microsoft says Copilot operates within existing permissions and access controls.[2]

Those existing permissions still need to match the access appropriate to this particular job.

A clear business rule needs a suitable implementation, followed by testing.

Test what happens when approval is missing

I'd include an incomplete example in the test, not only one where everything goes smoothly.

In the Friday follow-up, the delivery manager hasn't confirmed the date.

The useful result is an accurate draft with the date still unresolved.

If someone tries to send or commit a change without the required approval, does the process stop at the point we agreed?

What happens if the reviewer is unavailable?

What if they approve a different date, or the customer changes the request while the draft is waiting?

We should be able to see what is waiting, who needs to decide and which version they are reviewing.

That doesn't mean every job needs a complicated approval chain.

A person checking and sending a draft themselves may be enough for a simple process.

More connections create more things to examine, so I'd add them for a clear reason.

Check the result after the action

An approved action still needs to complete properly.

For the follow-up, I'd want to know whether the email was sent to the agreed recipients and whether the customer record reflects what was actually confirmed.

If a connection fails, the process needs to show that failure.

It shouldn't describe the job as complete just because an attempt was made.

I'd also decide how someone investigates or corrects the result if it differs from what was approved.

An email can't always be taken back, which is another reason to check the commitment before it leaves.

Keep the help proportionate to the job

The aim is to make the work easier to prepare and review.

If the review takes as long as doing the job ourselves, that belongs in the assessment of whether the assistant helped.

So does the effort spent maintaining its instructions and connections.

An ordinary workflow may handle an approval more predictably than asking AI to interpret it each time.

The assistant can still help with the part that needs interpretation, such as preparing a draft from the meeting and relevant correspondence.

I'd start with one job, make its stopping point visible and inspect the complete result before extending it.

Which step in your business would you want a person to approve before an assistant could act?

Reading notes

[1] Microsoft: Configure user authentication for tools.

[2] Microsoft: Security for Microsoft Copilot.

Follow The AI IT Guy on LinkedIn

Subscribe to Peter Kantarelis's weekly LinkedIn newsletter for practical ideas on AI and better business. LinkedIn can notify you when each edition is published, according to your notification settings.

Subscribe on LinkedIn

Back to Insights