Who Approved That? The rules AI agents need before we hand over control

A recent story caught my attention.

A content creator in Toronto reportedly asked Meta's AI agent to help sell a keyboard on Facebook Marketplace. Somewhere along the way, the agent shared his home address with a prospective buyer, who subsequently turned up late at night.

According to reports, Meta's explanation was that the agent had acted within the permissions the user had granted. The user, however, believed the agent would check with him before sharing that information.

Both may have understood the permissions differently.

And therein lies the problem.

Just because an AI agent has permission to do something, does that mean it should?

It's a deceptively simple question. And one I suspect businesses will increasingly have to confront.

We are moving from AI that advises to AI that acts

For the last few years, most of our conversations about AI have revolved around productivity. How much faster can we work? How much content can we generate? How many processes can we automate?

But something more fundamental is changing.

AI is moving from being a tool that helps us make decisions to a system that can make and execute decisions on our behalf.

An AI assistant drafting an email is one thing. An agent sending that email, negotiating commercial terms or committing a company's money is quite another.

The difference isn't technological. It's about authority.

And I'm not convinced we've spent enough time defining where that authority begins and ends.

We've seen this movie before

I don't think we should overreact to early incidents.

Every meaningful technology has had its growing pains. The early internet had serious privacy challenges. Programmatic advertising promised efficiency and scale, but it took years for the industry to confront issues around transparency, verification, fraud and accountability.

We eventually built frameworks to address many of those problems.

The question is whether we want to repeat that journey with AI agents.

Because the technology is moving quickly, and the commercial incentives to deploy it are significant.

The temptation will be to prioritise speed, efficiency and scale, and deal with governance when something goes wrong.

That would be a mistake.

Advertising offers an early glimpse of what's coming

This is particularly interesting in the world of media and advertising.

We're beginning to see agent-to-agent media buying, where AI systems representing advertisers and publishers can negotiate and facilitate transactions.

It's already happening. This month, Fox Advertising extended its agentic platform to live and linear TV, with agency groups including WPP, Omnicom, Horizon Media and Stagwell executing agentic transactions on live campaigns. A day earlier, LG Ad Solutions and Magnite opened LG's Home Screen inventory to agentic buying, with the first campaign running for a group of Toyota dealers.

Think about what that potentially means.

An advertiser's agent identifies an audience, evaluates inventory, negotiates pricing and executes a campaign. A publisher's agent responds with available inventory and commercial terms.

All with considerably less human intervention.

Having spent much of my career in advertising and technology, I find this genuinely exciting. There is some inefficiency in how our industry operates, and agents could eliminate a fair amount of it.

But there's an important distinction between automating a process and delegating commercial judgment.

Earlier in this series, I wrote that judgment is becoming the scarcest skill in a sales organisation. Agents raise the harder follow-up question: what happens when we start delegating it?

Who decides whether the price is acceptable? Who determines whether an inventory source is appropriate for a brand? Who takes responsibility when an agent optimises for efficiency but compromises a broader commercial objective?

Most of the announcements I've read lean heavily on speed, with agents replacing manual execution with automated workflows. Fox is an encouraging exception. It describes a governance layer that monitors agent activity throughout the workflow, and says human oversight remains essential.

That's sensible. It's also the right instinct to have at this stage.

But we should recognise that human approval may eventually become the very friction these systems are designed to eliminate.

And that's where the real challenge begins.

Four questions every business should ask

I don't believe we need to slow innovation. But we do need to be more deliberate about the rules within which agents operate.

If I were introducing autonomous agents into a commercial organisation today, I'd start with four questions.

1. What information can the agent access?

An agent doesn't need access to everything simply because doing so makes it more effective.

Customer records, commercial agreements, pricing, internal communications and personal information all carry different levels of sensitivity.

Access should be determined by necessity, not convenience.

2. What decisions can it make independently?

There's a difference between recommending a discount and offering one. Between identifying a sales opportunity and making a commitment to a customer. Between optimising a campaign and increasing its budget.

Businesses need to define which actions agents can execute, which require approval and which should remain entirely human decisions.

3. When a human approves, is it real approval?

We frequently hear the phrase "human in the loop."

It sounds reassuring. But is it always meaningful?

I recently came across an interesting piece by Madhusudhan Konda on LinkedIn, titled "Is Human-in-the-Loop Becoming the Problem? When Human Oversight Becomes Rubber Stamping."

He makes an important distinction between having humans "in the loop," approving individual actions, and humans "on the loop," defining boundaries, supervising the system and intervening when necessary.

It's a distinction worth thinking about.

Having led commercial teams for much of my career, I see a parallel. We don't expect managers to approve every customer interaction or commercial decision. We establish authority levels, escalation points and accountability. People operate independently within those boundaries.

Why should AI agents be any different?

If an agent is making hundreds of decisions every hour, expecting a human to approve each one isn't necessarily governance. It could simply become an administrative exercise.

The answer isn't more approvals. It's designing better oversight, where human judgment is applied to decisions that genuinely matter.

And importantly, ensuring someone remains accountable for the decisions being made.

4. Who is accountable when something goes wrong?

This, to me, is the most important question.

If an agent makes a decision that damages a customer relationship, violates a commercial agreement or exposes sensitive information, who owns that decision?

And can the organisation reconstruct what happened?

What information did the agent use? What authority did it have? What action did it take?

If we cannot answer those questions, we have delegated authority without establishing accountability.

That's a dangerous combination.

There's another stakeholder we shouldn't forget

Much of the conversation about AI agents focuses on the relationship between the user and the technology.

But agents rarely operate in isolation.

A sales agent accessing a CRM is interacting with information about customers. An agent reading an executive's inbox may be processing confidential correspondence from colleagues and partners. A marketing agent communicating with consumers is representing a brand.

These people may never have consented to the agent's deployment or decided what permissions it should have.

Yet they are affected by its actions.

Which means agent governance cannot simply be a conversation about what the person deploying the technology is comfortable with.

It must also consider the people on the receiving end.

The time to establish the rules is now

Here's why I think this conversation matters today.

We are still relatively early in the deployment of autonomous agents. Many implementations remain limited in scope, and humans are still involved in consequential decisions.

That's an opportunity.

Because once agents are deeply embedded in commercial workflows, processing thousands of transactions and interacting with millions of customers, changing the rules becomes considerably harder.

We've seen this with other technologies.

Governance introduced after scale is usually more expensive, more complicated and more reactive than governance designed from the beginning.

And the commercial pressure to remove human intervention will only increase as agents become more capable.

We should be preparing for that future now, rather than waiting for an incident to force the conversation.

Where I come out

I'm optimistic about AI agents.

I believe they'll fundamentally change how businesses operate, how commercial decisions are made and how organisations serve their customers.

But the freedom to act must come with clearly defined boundaries, meaningful oversight and accountability.

Trust shouldn't mean giving an agent unlimited freedom. It should mean knowing exactly what freedom we've given it.

Because whether it's sharing someone's home address or committing millions in advertising spend, the question remains the same.

Who approved that?

Disclaimer: The views expressed here are solely those of the author and do not in any way represent the views of exchange4media.com.