How To Run An Agency Mostly On AI (Without Blowing Up Client Work)

You run an agency mostly on AI by moving one function at a time, writing the operating rules before the agent touches anything, running the agent alongside the human for a few cycles, and keeping a human sign-off on everything that leaves the building. You do not announce a transformation, switch five things over in a month, and hope. The agencies that blow up client work are the ones that treated this as a tooling change instead of an operating change.

Start with the rule, not the agent

Every function you want an agent to run needs a written operation first. Not a prompt. An operation: what the inputs are, when it runs, what thresholds trigger what actions, what the output looks like, and what "done" means. If you cannot write that down for a function, you are not ready to hand it to an agent, and the honest truth is that you were probably running it on tribal knowledge with your humans too.

This is the step owners skip because it feels like paperwork. It is the whole game. An agent executes the rule you gave it. A vague rule gets you confident, fast, wrong output. A precise rule gets you something you can actually review against.

Write the operation as if you were briefing a very literal new hire who will do exactly what the page says and nothing else. Because that is what you are doing.

The order that protects client work

Sequence matters more than speed. Here is the order I would use, and why.

1. Internal reporting and monitoring

Nothing the client sees directly. Daily account checks, spend monitoring against caps, pulling the numbers into the same format every week. If the agent gets it wrong for a few days, you catch it internally and nobody outside knows. This is where you learn how to write rules the agent can follow and how to review its output, with no client exposure.

2. First-draft production

Ad variants, email drafts, landing page copy, briefs, outlines. The agent produces to a written brief; a human edits and approves. Your production people stop being the bottleneck on volume and become the taste layer. Quality often goes up here, because the human is now reviewing twenty options instead of making three while tired.

3. Rule-bound account actions

Pausing ad sets above a threshold, scaling within a cap, launching campaigns from an approved template. This is live money, so it comes after you have proven the agent follows your rules on the two previous functions. Caps go in at the platform level as well as in the operating instructions. We walk through this specific function in AI agents vs media buyers.

4. Client-facing output

The weekly report the client reads, the scheduled email that goes out under their brand, the content calendar. Agent produces, human reads and sends. This step never loses the human read. Not after three months, not after a year. It is the cheapest insurance you will ever buy.

5. What stays human

Strategy, offer design, creative direction, the review call, the renewal conversation, and the decision about what to do when the numbers do something the rules did not anticipate. These are the product. Agents support them with research and drafts; they do not own them.

Run in parallel before you switch

For each function, there is a window where both the human and the agent do the job. Same inputs, same week. You compare. When the agent's output passes the same review the human's did, consistently, you retire the human step on that function and redeploy the person. When it does not, you fix the rule, not the agent, and run another cycle.

Two things this gives you. First, no cliff edge for the client: the work never stops being done by someone who knows how. Second, your people see the comparison with their own eyes, which turns a scary change into a visible one. Most good people are relieved to stop doing the clerical half of their job.

Headcount: let attrition do it

The question underneath all of this is "do I have to fire people?" and for most agencies the answer is no. Agencies turn over production staff constantly. Stop backfilling production roles as they open, move the function to an agent with the departing person's help in their last weeks if you can, and redeploy the people you keep onto strategy and review. Within a reasonable stretch you have a smaller team doing higher-value work, and nobody was marched out of the building.

What this does to the P&L is covered in what full-time agency staff actually cost you beyond payroll. The short version: each production role you do not backfill removes not just a salary but a management slice, a software seat, a recruiting cycle and a slice of your own attention.

Tell the clients

Put the operating model in your contract language and say it out loud at the next review: here is how the work is produced, here is the named person accountable for your account, here is the review step before anything ships. Clients are buying outcomes and accountability. When the weekly report is clearer and arrives on time every week, the method stops being a concern. When you hide it and they find out, you have a trust problem that has nothing to do with AI.

The mistakes that actually blow things up

• Switching a function and removing the review in the same week. Pick one. Do the other later.
• Agents with no spend cap. If an agent can act on a budget, the cap is written down and set at the platform, before it runs once.
• Letting an agent change its own rules. Agents escalate. Humans change rules. Keep that line hard.
• Treating the chat window as the agent. If a person has to prompt each step, you have a faster employee and the same bottleneck. The functions above run on schedules and triggers.
• Going five functions wide before one is solid. Depth first. One function all the way through parallel-run and switchover teaches you more than five half-done.

What it looks like when it works

The owner reads reports instead of writing them. The strategist spends the week on client thinking instead of in the ads manager. Production volume goes up, review gets stricter rather than looser, and the next client does not require the next hire. That is the model described in what an agency without employees is, reached by sequence rather than by announcement. It is how I run my own company today, and if you want to see how it maps onto your book of clients, the application on the home page is the way to start that conversation.

FAQ

Which agency function should move to agents first?

The one that is most repetitive, most rule-shaped, and least visible to the client if it wobbles for a week. For most agencies that is internal reporting or first-draft production, not client strategy or anything that ships under the client's name without review.

Should I tell clients I am moving work to AI agents?

Yes. Put it in the contract language and say it plainly in the next review. Frame it as how the work is produced, with a named human still accountable and signing off. Hiding it is a bigger risk than disclosing it.

How do I avoid quality dropping during the transition?

Run the agent and the human in parallel on the same function for a few cycles, compare outputs, and only retire the human step when the agent's output passes the same review the human's did. Never switch a function over and remove the review in the same week.

Do I have to lay people off to do this?

No. Most agencies have enough natural turnover that simply not backfilling production roles gets them most of the way. Redeploy the people you keep onto client strategy and review, where their judgment is the product.