Your People Are Already Managing AI Agents. Nobody Gave Them the Job.

The AI manager has arrived without a title, a mandate or a pay grade. California just wrote the first job description.

By John Margerison · Mon, 12 Oct 2026 · 1017 words

I have spent a year telling boards that the AI manager is the next role on the org chart. This month I want to correct myself. The role has already arrived. It arrived without a title, a mandate or a pay grade, and that is the problem.

Start with the agents, because they are the thing being managed. Salesforce's Agentic Enterprise Index, published on 7 August, reports that the organisations it tracked nearly tripled their activated agents between February 2025 and April 2026, that the average agent went from two skills to six, and that the average employee engaged with an agent 300 per cent more often each week. KPMG's third-quarter pulse survey of 314 US leaders at companies with a billion dollars or more in revenue found multi-agent deployment up from 6 per cent to 25 per cent in a single quarter.

Now look at who is running them. The same KPMG survey found that 44 per cent of organisations report significant workforce adoption, up from 10 per cent a year ago, and that 71 per cent of leaders say the people who use AI well are outperforming everyone else. One in four expect to rewrite promotion criteria around AI literacy and agent delegation. Shimon Hason, who runs Incredibuild, put it plainly in CTech on 4 October: every employee is, in a sense, a manager, and job interviews will soon ask how a candidate manages agents, not only people.

So the management work exists. What about the job?

Here the evidence is thin, and I would rather say so than dress it up. No official statistics break out AI manager or AI operations as a title. The best sample I can find comes from a recruiting firm, Axial Search, which counts about 93 new US AI operations postings a week at a median advertised salary of 156,000 dollars, and reports that only 5 per cent sit at director level. Indeed Hiring Lab found in July that AI had reached 822 distinct US job titles, about one in twelve, with 63 per cent of them outside tech. Named postings are rarer still. Disney advertised a Director of AI Enablement and Legal Engineering on 24 September, a role that redesigns legal workflows, sets testing practice and reports the return on AI spend. That is an AI manager by any sensible definition. It is also one posting.

Put those two pictures together. Agents are multiplying at the rate of a new team every quarter. The people supervising them are the existing workforce, graded on it now and promoted on it soon. The formal role barely exists. The gap between those two facts is where the risk lives, and two surveys from the last month measure it.

Dataiku's Harris Poll of 685 CIOs, published 24 September, found that 84 per cent say employees are creating agents faster than IT can govern them, 94 per cent in the US, and that 81 per cent lack complete oversight of agents built outside approved channels. Asked who is accountable when an agent goes wrong, the answers split four ways, none above 23 per cent. A Morning Consult survey for Guild.ai, published two days earlier, found 96.4 per cent of IT decision-makers confident their agent inventory is complete, while 66.7 per cent had suffered an agent-related operational failure in the past year and only 31 per cent could stop a misbehaving agent automatically.

Read those together. Nearly everyone is confident they know what is running. Two thirds have already been hurt by something that was running. That is not a tooling gap. It is an ownership gap. The work of managing agents has been pushed onto employees who have no authority to switch anything off, no log that proves they looked, and no line in their job description that says the outcome is theirs.

Which brings me to the part that moves this from an HR question to a liability one. On 30 September Governor Newsom signed SB 947. From 1 July 2027, a California employer that primarily relies on an automated system to discipline or fire someone must direct a human reviewer to corroborate the decision from personnel files, work product or witness interviews, and must hand the employee a written notice saying a human did so. If the reviewer cannot corroborate the output, the employer cannot use it. A companion bill, SB 951, adds AI disclosures to mass layoff notices.

Strip the legal language away and that is a job description. A named person, accountable for an AI output, who checked it against independent evidence and signed. California has written the AI manager into statute before most companies have written it into an org chart. Other states copy California, and your insurer reads the same bills you do.

I wrote The AI Manager because I believed this role would be the most valuable hire of the next three years. I did not expect a legislature to publish the first job spec.

So here is the one thing I would do next week, and it costs an afternoon.

List every AI agent that can act on a customer, an employee or money. If you cannot produce that list, that is your finding, and you are in the same room as the 81 per cent of CIOs who told Harris they lack full oversight. Then write one human name against each agent. Not a team, not a committee, a person. Give that person three things: the log of what the agent did, the authority to stop it, and a line in their role that says the outcome is theirs. Then put it on their next review, since you are about to grade them on AI use anyway.

That is the AI manager. You can hire for it, promote into it, or wait for a regulator to assign it. The hiring wave everyone is waiting for is not arriving as a flood of new titles. It is arriving as a duty attached to people you already employ. Name them before someone else does.

The AI Manager is free to read at book.theaimanager.ai. More at johnmargerison.com.