AI Operations Manager: Role or Toolset?

Search for “AI operations manager” and you get two different things wearing the same name. Half the results are job postings and career guides. The other half are software products promising to run your operations for you. Both use the phrase with confidence, and they mean opposite things by it.

I have spent twelve years running delivery operations for digital marketing agencies, and I sell both halves of that phrase. I fill the operations seat as a fractional hire, and I build AI automation into agency delivery. So I have an obvious interest in the answer, and you should read what follows with that in mind. Here is my verdict, stated up front so you can stop reading if you already disagree: for an agency of five to fifty people, an AI operations manager is a toolset bolted onto an operations seat. It is not a replacement for the person sitting in it.

Is this an article about becoming an AI operations manager?

No. If you searched the phrase because you want that title on your CV, this is not the article for you. I have no career path to sell and no advice worth your time on how to get hired into one. This is written for the person on the other side of that decision: the agency owner, COO or operations director trying to work out whether the phrase describes something to buy, something to recruit, or a repackaging of both.

Why does one phrase mean two different things?

Because two markets arrived at it from opposite directions. Recruiters needed a title for operations people expected to be fluent with AI tools, which is a fair thing to want. Software companies needed a category name that carried more weight than workflow automation, and a job title makes a better one than a feature list. So the phrase now names a person in one context and a product in another, and neither side has any reason to clear it up.

The confusion is not harmless, because it changes what agencies buy. Treat the category as a product and you start evaluating vendors. Treat it as a role and you start writing a job description. Those are different budgets, different timelines and different failure modes, and the thing that should decide between them is what is actually broken in your delivery, not which half of the search results you landed on.

What can the software layer genuinely own?

A fair amount, and I want to give it credit because I build with it. Inside agency delivery, the steps where AI earns its keep all share the same shape: a defined input, a defined output, and a human who checks the result before a client sees it. Within that shape it does real work.

• Drafting status updates and meeting notes. The input is structured, the output is reviewed, and the hours saved are consistent week to week.
• First-pass QA against a checklist you wrote. The machine flags what fails. A person reviews the flags. The obvious misses get caught without spending senior time on them.
• Routing and categorising inbound requests. The routing logic is yours. AI applies it faster than a person working through an inbox.
• Assembling reporting commentary. The numbers come from your tools. AI writes the first draft of the narrative around them. Someone edits for tone and accuracy before it goes out.
• Research and briefing prep. Background and competitor context assembled before a kickoff, then sharpened by a human who knows the account.

None of those is the job. They are tasks inside the job, and that distinction is the whole argument. I have written up where I think AI belongs in agency delivery, and where it does not, on my AI automation services page.

What can it not own?

Consider what running delivery for an agency actually consists of in a normal week. Someone chairs the standup. Someone checks capacity and decides what moves when two clients need the same person on Wednesday. Someone chases the unclear brief before three days get spent building the wrong thing. Someone decides a project will miss its date and that the client hears it from us first, with a revised timeline attached. Someone owns the Friday status so that Friday is a send and not a Thursday-night scramble. I have set that week out day by day on my fractional operations manager page.

Now go back through that list and mark which items are information problems and which are decisions with a consequence attached. The information problems are automatable, and should be. The decisions are not, and the reason is not that the technology is immature. Four constraints hold no matter how good the models get:

• Accountability. If output goes to a client under your name and a wrong answer costs you the account, a person owns that decision. Software cannot be held responsible for anything.
• Client judgement. When a client is unhappy, when scope is ambiguous, when a deliverable is technically correct and strategically wrong, the call depends on understanding a relationship rather than a dataset.
• Unverifiable output. If you cannot cheaply check whether an answer is right, you have not saved time. You have moved risk somewhere you cannot see it.
• Undefined process. AI cannot run a process nobody has defined. It will produce something plausible and wrong, and you will find out when a client finds out.

A tool can tell you that a project is slipping. Deciding what to sacrifice, telling the client before they notice, and carrying the consequence if that call was wrong is the seat, and the seat is a person.

What happens when an agency buys the tool instead of filling the seat?

I have watched AI implementations fail inside agencies, and the patterns repeat. Four of them account for most of it.

Quality drifts. The output is good in week one. By week six nobody is checking it the way they did at the start, because it still looks fine at a glance.

The review burden moves rather than disappearing. You automated the drafting, so now someone reviews every draft. If the review takes as long as the original task, you added a step instead of removing one.

The tools change underneath you. Models update and features shift, so an automation that behaved one way in March behaves differently in June. With nobody maintaining it, it breaks quietly.

The team abandons anything that costs more time than it saves. This is the most common ending, and the quietest. The tool works. Nobody uses it. Nobody tells you.

Every one of those has the same cause. Nobody owned it. The tool was bought as a substitute for the seat, so no one was assigned to keep it honest, and an automation without an owner degrades on a schedule you cannot predict.

So does a five to fifty person agency need an AI operations manager?

Yes, in the sense the phrase should mean, and no in the sense most of the marketing means it. You need an operations owner who is fluent in AI. You do not need, and cannot buy, a product that replaces that owner. The order matters more than the label: process first, automation second, AI only where it beats a rule. Rules are predictable, auditable and do not drift when a model updates, and a surprising share of what gets sold as AI is work a plain rule handles better and cheaper.

In practice that means one of two hires, sometimes both in sequence. If the seat is empty, fill the seat, and give whoever sits in it AI as leverage on the parts of the job that have a defined input and a checkable output. If the seat is filled but the systems around it are manual, that is a build, and it is AI implementation work rather than a management hire. What you should not do is buy a product on the strength of a demo and hope it does the job of a person who does not exist yet.

My basis for that read, so you can weigh it: twelve years running delivery operations for marketing agencies and 92 completed engagements, working daily inside ClickUp, Asana, Monday.com, Airtable and Notion with n8n connecting them. Ask anyone selling you either half of this category to show you the equivalent before you believe them.

Where should you start?

The prior question, before you evaluate a single vendor or write a single job description, is which parts of your delivery week are information problems and which are decisions. That is what the Agency Ops Audit answers. Two weeks, $1,500 fixed, a written 90-day operations roadmap and one 60-minute readout call. You come out of it knowing which steps are worth automating, in what order, and which ones need a person to own them.

If the roadmap says the answer is a seat, you will know what to hire for. If it says the answer is a build, you will know what to build. Either way you will have stopped shopping for a job title that is currently being sold as software.