An AI UGC agency is a production line, not a roster.

What breaks when UGC production goes from one brand to eight, which parts are worth standardising first, and why the connection to the generator matters more than the generator.

roster.json

{
  "clients": [
    {
      "name": "sleep-tea",
      "creator": "f-35-kitchen",
      "aspect": "9:16",
      "length": 20,
      "deliver": "drive://clients/sleep-tea/creative"
    },
    {
      "name": "dog-food",
      "creator": "m-40-garden",
      "aspect": "9:16",
      "length": 15,
      "deliver": "drive://clients/dog-food/creative"
    }
  ],
  "weekly": { "hooks": 5, "refresh": "monday 07:00" }
}

One row per account. The Monday batch is this file plus a single instruction.

What changes at volume

Nothing about producing one UGC ad is hard. The difficulty is entirely in the second dimension of the problem, which is how many accounts are running at once.

Take a small roster. Eight clients, two products each, five hooks per product per month. That is eighty briefs and eighty finished assets a month, before any of them get recut for a second placement. With a shoot based process that number is a hiring plan. With generation it is a Tuesday, provided the work around the generation has somewhere to live.

The trap is assuming the generator is the constraint. It rarely is. The constraint is that eighty assets means eighty briefs collected, eighty files named, eighty approvals chased and eighty rows in a report that a client will read in a meeting. Agencies that grow this line and do not automate the coordination end up selling hours they cannot bill, which is the same problem the shoot based version had.

The other thing that changes is what the client is buying. When one variant costs a day, the deliverable is the video. When it costs minutes, the deliverable is the testing programme: how many angles you tried, what you learned, and which one is now carrying the account. That is a better thing to sell, and it is a harder thing to fake.

The six parts that break first

In roughly the order they show up as the roster grows. None of them are creative problems, which is why they get ignored until they are urgent.

Brief intake

Every client sends the brief in a different shape: a call, a doc, a message, a spreadsheet. The translation into something producible is unpaid work that grows with the roster.

Consistency per account

Ad forty for a client should look like ad one. Without a written creator and setting spec, whoever produces it that week quietly restarts the brand from scratch.

Approvals

Five assets per client per week is fine. Five approval threads per client per week is not, and it is where the calendar time actually goes.

Delivery and naming

Files arrive named after the tool that made them. The buyer then cannot tell which ad won, so the reporting conversation starts with archaeology.

Reporting back

Clients want to know what was tested and what happened. That answer is easy when the brief, the run and the file name share a convention, and painful when they do not.

Margin per asset

Generation cost is small. The hours around it are not. An agency that does not automate the coordination sells hours it cannot bill.

Building the line

Three things to write down per client. Once they exist as files, the weekly batch stops being a project and becomes a run.

Standardise the brief
One shape for every client: product, audience, format, creator, hook, close, length, frame. Anything the client sends gets translated into that shape once, and the shape is what production consumes.
Freeze the creator per account
Write the creator, the setting and the read into a per client spec and reuse it. Continuity across a campaign is worth more than variety, and it removes a decision from every single job.
Deliver into their system
Finished files go to the folder or the ad account the client already uses, named to the convention, with the brief attached. Delivery that requires a person to remember something is delivery that eventually fails.

.mcp.json

{
  "mcpServers": {
    "wireflow": {
      "type": "http",
      "url": "https://www.wireflow.ai/api/mcp"
    }
  }
}

One endpoint covers the whole roster. Access is approved once, on the first call.

Being straight about it with clients

This is a commercial question more than an ethical one, and the commercially safer answer happens to also be the honest one.

Tell the client the performers are generated, in writing, before the first delivery. Finding out later from their own legal team is the version that costs you the account.
Label synthetic media where the platform asks for it. The major ad platforms have AI content disclosure settings, and using them is cheaper than an appeal.
Never present a generated person as a real customer, a real founder or a named expert. That crosses from production method into a claim about who is speaking.
Keep the brief and the run log per asset. When a client asks why an ad said something, the answer should be a file rather than a memory.
Sell the programme, not the shoot that did not happen. Twenty tested angles a month is a real deliverable, and it does not require anyone to pretend.

What MCP changes for an agency

Every UGC tool on the market is a dashboard a person sits in front of. An agency does not need a better dashboard, it needs the tool to be callable by whatever already runs the week.

Intake, generation and delivery become one script. The brief is read from wherever it landed, the run happens, the file is named and dropped in the client folder, and nobody opened a tab.
The roster lives in a file, so a new client is a new row rather than a new process. The Monday batch scales with the file, not with the producer.
Budget is visible before the spend. The credit balance is a tool call, so an automation can check first and stop instead of failing halfway through a client batch.
Everything is text, which means the whole month is auditable. What was briefed, what ran, what came back and what got delivered are all in the log.
It runs unattended. A weekly refresh does not need a person awake, which is the difference between a service you sell and a task you do.

Wireflow is our own product and the one we recommend here, because its generation workflows sit behind a hosted MCP endpoint that any client can call. The general version of this pattern is UGC automation, and the setup from the beginning is on the AI UGC generator overview.

monday batch

You

Run the weekly batch for every client in roster.json. Five hooks each, deliver to their folders.

get_credit_balance{}

↳ balance sufficient for the queued jobs

Agent

Two accounts in the roster, five hooks each. I will keep the frozen creator per account so the new ads match what is already running, and name the files to the convention.

run_workflow{ "workflowId": "ugc-ad-v4", "inputs": { "client": "sleep-tea", "length": 20, ... } }

↳ executionId exe_77c10, nine more queued

get_execution{ "executionId": "exe_77c10" }

↳ status completed, output sleep-tea_hook01_f35_20s_9x16.mp4

Agent

Ten files delivered to the two client folders and logged in the tracker. One dog food hook came back four seconds short because that script is thin, so it is worth a rewrite before the review call.

get_credit_balance, run_workflow and get_execution are real tools on the server. Only run_workflow spends credits.

Where the rest of this lives

The three pages a producer usually wants next, in the order the work happens.

Writing the scripts the line runs on: AI UGC ads.
Freezing a creator per account: AI UGC creator.
Placement specs for the delivery step: UGC video ads.

Questions agencies ask about AI UGC

What is an AI UGC agency?

It is an agency that produces creator style video ads using generation instead of, or alongside, booked creators. The client relationship is the same as any performance creative retainer: briefs in, tested assets out, a report on what worked. What changes is the cost and the turnaround of a single variant, which changes how many variants the retainer can afford to test.

Should we tell clients the creators are AI generated?

Yes, in writing, before the first delivery. It is a disclosure question on the platform side too, since the major ad platforms expect synthetic people to be labelled. Agencies that treat it as a normal production method have an easier conversation than agencies that hope nobody asks.

Can we white label AI UGC production?

You can present the output as your own creative service, the same way you would with any production partner. What you should not do is describe generated performers as real customers, or imply a shoot happened. Sell the outcome, which is a tested set of ads per week, rather than a story about how it was filmed.

How should an agency price AI UGC?

Per asset pricing stops making sense once an asset takes minutes. Most of the value moves to the testing programme: a monthly retainer for a set number of tested angles per client, with the reporting attached. That prices the judgement, which is the part that is actually scarce.

What does MCP mean for an agency?

MCP is the Model Context Protocol, an open standard that lets an assistant call an outside service as a tool. For an agency it means the generator can be driven by the same automation that already handles intake, naming and delivery, rather than being a website someone has to sit in front of. One connection covers every client account.

How many clients can one producer handle this way?

More than with a shoot based process, though the honest limit is not generation capacity. It is how many creative conversations one person can hold at once. Automating intake, naming and delivery buys back the hours that were never the interesting part of the job.

Connect it once and run the first client batch

Add the endpoint to whatever already runs your week, point it at one account, and see what a five hook batch costs in time. The read only tools are free to call.