15 Agents Today, 15,000 by 2028: The Role Nobody's Hiring For Yet

Then there's the part nobody's costing for. Every second brief I take now has the word "agent" in it somewhere. When I ask who's accountable if one of them does something nobody sanctioned, the answer is usually a pause.

Read that again. The hard part isn't whether the agents work. It's that nobody owns them.

What that gap actually tells you

Businesses that ran a pilot in 2024 are now trying to make agents part of how work gets done. That isn't a bigger version of the same technical problem — it's an operating model problem, and it lands on people rather than IT.

Which is why the hiring response so far has been slightly off. Organisations buy more capability and never buy control. They add another engineer to a problem that needed an owner.

What an agent manager actually does

The title came out of Harvard Business Review earlier this year and it's already showing up on job boards. The work is defining what an agent is permitted to do, who signs off when it goes off-script, how its output gets checked, and who carries the consequence when it's wrong.

So it isn't a second machine learning engineer. It's an accountability role with a technical surface, not a technical role with an accountability footnote — brief it as an engineering hire and you'll get engineering candidates and still have no owner. I've mapped it against the rest of the emerging set in AI job roles 2026.

It's also not an AI Governance Specialist. They write the framework; the agent manager runs agents inside it.

Why it lands differently here

The KPMG and University of Melbourne study of 48,000 people found roughly half of Australian employees admit to using AI in ways that breach their employer's rules. Only 30 per cent thought the benefits outweighed the risks — the lowest of the 47 countries surveyed.

And we have no AI Act. The National AI Plan is a roadmap, with legislation not expected until 2027. So for the next eighteen months your operating model is your governance. At executive level that's increasingly a Chief AI Transformation Officer. The agent manager is the layer beneath, and the one more businesses are missing.

If you're hiring into this

Process ownership under ambiguity. Have they been accountable for a complex system working when the rules weren't written yet? Ask for the instance, not the philosophy.

Failure-mode literacy. Can they describe where their tools break without being prompted? The strong ones volunteer the limitations.

Domain depth over code depth. An agent managing claims needs a manager who understands claims. Technical fluency builds on that far faster than the reverse.

Build, borrow or buy

One of the rare roles where I'd put building ahead of buying, and I don't say that often. The market is thin because the role is barely eighteen months old — anyone claiming five years of it is telling you something useful. Your operations and risk leads already hold the harder half of the capability.

If you'd rather have the model set and handed over, that's a borrow: fractional AI leadership runs roughly $1,500–$3,000 a day for one to three days a week. Our Bot, Build, Borrow, Buy framework maps where each option fits.

Where this is going

Agents are being deployed faster than they're being governed, and governance here is a person, not a policy document. Name that person early and you spend the next two years scaling. Leave it and you spend them untangling.

If you're working out what that looks like in your business, that's the conversation I'm having most weeks. Reach out — it's worth thinking through before the agent count gets away from you.

‍