🚨 AI agents work. Managing them is the second business you never signed up for

The City of Raleigh, North Carolina, runs a help desk on ServiceNow's AI platform. By the public numbers, the rollout has been a clean win:

All from the public ServiceNow case study. Nice numbers.

But buried in the rollout was a detail nobody puts on the slide deck. When the help desk first deployed the agent, the supervisor's workload didn't drop. It spiked. The supervisor still managed humans — but now also had to train the agent, monitor its behaviour, and quality-check its responses.

The CIO, Mark Wittenburg, summed it up in one sentence: "That's been a transition for the supervisor."

Now hold that picture in your head. A city. A salaried supervisor. A team to back them up. One agent. Workload up.

Now ask the question that should be making every agency owner uncomfortable: if one agent did that to a supervisor with a team, what does a ten-agent stack do to a solo coach with no team?

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Every "AI replaces contractors" pitch you've seen — including, honestly, the one in my own three-tier stack doc — leads with the same eye-catching stat.

$500 a month in AI tools can replace $50,000 a month in contractors.

The math is real. The agents work. The savings are not imaginary.

But the pitch quietly assumes something nobody priced in: that someone, somewhere, will manage the replacement. In a 500-person company, that “someone” is a middle manager getting stretched thinner. In a solo agency, that someone is you — between client calls, at 11 pm, on a Saturday.

The technical name for this role is "AI governance." The plain-English version: the agent manager. It's the job that doesn't exist on your org chart yet, doesn't have a salary, doesn't have a line item — and you're already doing it.

Badly, probably. Most owners are.

Here are the four invisible jobs that landed on you the moment you bought the stack.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Think of every new contractor you've onboarded. You spend the first week explaining how you work, what "done" looks like, where the files live, which clients are sensitive, and what tone to use in emails. The contractor learns. After a month, the explaining stops.

An AI agent doesn't learn the way a contractor does. Most agent setups today use what researchers call "naive memory accumulation" — meaning the agent forgets context between sessions unless you build the memory layer yourself. Every time the model gets updated, the training partly resets. Every time you onboard a new client, you re-explain.

For one well-tuned agent, the trainer role is roughly a few hours a week, in my own build. Multiply that across a five-agent stack. That's a part-time job you didn't realise you'd taken.

An AI agent is like an enthusiastic new intern who is occasionally, confidently wrong. The technical name is "hallucination." The plain version: the agent makes things up — and says it like it's true.

How often does this happen in production?

Somebody has to read every output before it leaves the system. For an agency, that's the difference between a delighted client and a client who fires you in week three because the agent quoted a fake stat in a deliverable.

That somebody is you.

An AI agent is like a contractor who quietly bills by the minute for every step of their thinking — and every retry, and every sub-task, and every time they re-read their own notes. You set up the agents. The agents work. The bill grows. Nobody warned you it would.

Three real incidents — all sourced — that should change how you think about pricing your retainers:

The reason this happens: an AI agent doing real work consumes 5 to 30 times more compute than a simple chatbot answering a question. The agent thinks step by step, calls tools, re-checks itself, retries when something fails — and the bill is for every step.

Gartner now forecasts that 40% of AI agent projects will be cancelled by 2027 due to cost overruns alone. Not a technical failure. Just bills that arrived faster than the savings.

Somebody has to watch that bill before it watches you. That somebody is you.

An AI agent is like a contractor you handed a spare key to your office, your CRM, and every client file you own. If the contractor is well-behaved, you'll never notice. If something goes wrong — a mistake, a stolen laptop, a bad instruction — they can reach everything.

The numbers from the past 12 months:

For a solo agency, the most realistic version of this isn't a $27M heist. It's a client's customer list leaking through an agent that had broader access than it needed — and you not noticing for three months.

The safety guard's job is to check what every agent can see, what it can do, and whether it should still be running at all. That's also you.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Across my own RFA build and the public reports above, the pattern is consistent enough to name:

🟢 Week 1: euphoria. The agents work. Tasks that took hours now take minutes. The math is real. You tell your spouse you might have just bought your evenings back.

🟡 Week 2: the first surprises. The first made-up answer — caught by you, before the client saw it. The first bill that's bigger than you expected. The first "wait, who's supposed to be checking this?" moment at 10:30 pm on a Tuesday.

🔴 Week 3: the math stops working. You're spending three to four hours a day managing the agents. The same hours the contractors used to absorb. The $500-versus-$50,000 stat is still technically true on the invoice side — but your time, the most expensive resource you have, is now fully consumed by agent management.

This isn't speculation. Imperial College London and Microsoft published research last month showing that AI adoption may paradoxically increase workplace burdens — because workers find themselves "babysitting multiple AI agents and correcting their errors." Gartner's lead researcher Helen Poitevin put it more bluntly: "Workforce reductions may create budget room, but they do not create return."

Same pattern. Different scale.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Here's where the question turns sideways.

If you're doing the four jobs badly for yourself — and you plan to sell agent builds to clients — what happens when those clients try the same thing?

They'll do it worse. Less context, less technical fluency, less patience. The build will work for two weeks. Then the bills, the made-up answers, the access leaks, and the silent retry loops will start. They'll either cancel the retainer or call you to fix it.

That call is the retainer.

What enterprises with whole teams call "AI governance" — what ServiceNow has built an entire control-tower platform around — is the same four jobs, just with budget and headcount attached. For your clients, you can be the headcount.

This is the part the "$500 replaces $50,000" pitch leaves out: the agent stack doesn't replace contractors. It replaces the contractors and creates four new jobs that all land on you. The retainer is what pays a real person — you — to do those four jobs for the client.

Price the build. Then price the management layer separately. Recurring. Honest. Defensible in any DM.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Of the four invisible jobs — trainer, quality checker, bill watcher, safety guard — which one ate the most of your time last month?

I genuinely want to know whether the pattern I'm watching holds outside my own build. Hit reply, or come tell me in the RFA Skool community.

The conversation is sharper over there because everyone running these stacks knows the 11pm version of the problem. Free to join. The next post in your inbox will go deeper on what an "agent manager" line item actually looks like on a retainer invoice.

— Bibhash