How AI Automation Agencies Work (and What to Ask One)

An AI automation agency is a services firm that designs, builds and often maintains automations that use AI models inside your business: workflows in tools like Zapier, Make or n8n, customer chatbots, AI agents that act on your systems, and the integrations that connect them. You are buying a working system and, ideally, the ability to keep running it without them. The useful questions are less about which model they use and more about scope, how they test, who holds the accounts and keys, and what you own when the engagement ends.

This is a guide for hiring one. If you would rather build the automations yourself, start with workflow automation tools, which compares Zapier, Make and n8n for small teams.

What an AI automation agency builds

Most engagements fall into four kinds of work, and many projects mix them:

• Workflow automations: a trigger and a series of steps, with a model doing the one step that needs reading or judgment, such as classifying an inbound email or summarizing a call. Built in Zapier, Make, n8n or custom code.
• Chatbots and assistants: a website or help desk bot that answers from your documents, books appointments or hands off to a person. The hard part is the hand-off and what the bot is allowed to promise.
• AI agents: a model with tools that takes actions, such as updating a CRM, drafting replies or filing tickets. AI agent use cases lists forty by team.
• Integrations and data plumbing: connecting your CRM, accounting, help desk and spreadsheets so the automations have something reliable to read. This is often most of the hours, and it is the part that keeps working if you change models.

A good agency will also tell you when a step should stay a plain rule. Moving a paid invoice into accounting does not need a model; deciding which of forty support emails is urgent might.

How engagements and pricing models work

Prices vary too much by region, scope and firm to quote usefully, but the shapes are consistent. Ask which one you are being offered, and what happens at the end of it.

• Paid discovery or audit: a fixed fee to map your processes and recommend what to automate. Worth it if the output is a written document you keep, whoever builds.
• Fixed-scope build: a set price for a defined set of automations, with acceptance criteria. The scope document is the contract that matters.
• Time and materials: billed by the hour or day. Flexible, and only as predictable as the reporting.
• Monthly retainer: maintenance, monitoring and small changes. Most AI automations need some, because models, prompts and connected apps change.
• Per automation or per agent: a price for each workflow or bot, sometimes with a monthly charge to keep it running.
• Usage pass-through: model and platform usage billed to you, either at cost or with a markup. Ask which, and ask for the bills to be in your name.
• Outcome or revenue share: payment tied to a result, such as leads booked. Rare, hard to measure, and the agreement needs to define the measurement precisely.

What it costs to run an agent after it is built, in tokens, platform fees and people’s time, is covered in what AI agents really cost to run.

How to evaluate an AI agency

1. Ask for a live demo of something they built for a similar business, running on real data, not slides or a screen recording.
2. Ask how they test. A serious answer names a set of example inputs with expected outputs, run again whenever a prompt or model changes.
3. Ask what happens when the model is wrong. Look for a person who approves risky actions, a log of what the automation did, and a way to turn it off.
4. Ask who will do the work. Many agencies subcontract; that is fine if you know and the contract covers it.
5. Ask for two references you can call, ideally a client whose project ended so you hear about the hand-over.
6. Check whether they can explain the work without the jargon. If you cannot follow the scope document, you cannot accept the work against it.

Red flags the FTC has already acted on

The FTC has been clear that AI claims are held to the same standard as any other marketing. Announcing its Operation AI Comply↗ (opens in a new tab) sweep in September 2024, the agency said “there is no AI exemption from the laws on the books.” That sweep included schemes that promised AI-powered online stores and passive income that, according to the complaints, never materialized.

• Guaranteed returns. Specific promises of revenue, savings or payback in a set number of days, without evidence you can check.
• Claims that the AI replaces a professional. The FTC’s case against DoNotPay, the “robot lawyer”, turned on claims of that kind made without testing to back them.
• Refund or buy-back guarantees that are vague about conditions.
• Pressure to remove or change negative reviews, which the FTC alleged in the Ascend Ecom complaint.
• No written scope, or a scope that describes features but no acceptance criteria.
• They want your passwords rather than an invitation to your accounts.

One pattern deserves its own warning. If what you are being sold is a way to start your own AI agency, with coaching, white-label software or a license to resell, you are probably buying a business opportunity, not a service. The FTC’s Business Opportunity Rule↗ (opens in a new tab) requires sellers to give prospective buyers specific information to help them evaluate one. In March 2026 the FTC announced that Air AI and its owners will be banned from marketing business opportunities↗ (opens in a new tab) under a settlement over deceptive earnings and refund claims, including failing to provide the required disclosure documents.

Who owns the code, data and accounts afterward

Settle this before work starts, in writing. Copyright in code normally belongs to whoever wrote it unless the work is “made for hire” or rights are transferred. The U.S. Copyright Office’s circular on works made for hire↗ (opens in a new tab) explains that a commissioned work only qualifies if it falls into one of nine listed categories and both parties sign a written agreement saying so. Computer programs are not on that list, so many contracts with outside developers rely on a written assignment of rights instead. Have a lawyer review the clause; this is not legal advice.

• Code and workflow exports: delivered to a repository or workspace you own, with the prompts, schemas and test examples included. Prompts are part of the system.
• Accounts: the automation platform workspace, the model provider account and the API keys are created in your company’s name, with the agency invited as a member you can remove.
• Data: what the agency may keep, for how long, and whether any of it trains a model. Ask which model providers will see your customers’ data.
• Reusable components: agencies often keep their own libraries. Fine, as long as you get a license to use them that survives the end of the contract.
• Hand-over: a written runbook, an hour of walkthrough, and a date after which they no longer hold credentials.

Questions to ask an AI automation agency

1. Which three automations would you build first for us, and what would each save or change?
2. What will you need access to, and can we grant it by invitation with the least permission that works?
3. How will we know an automation is working a month from now? What gets logged?
4. Which actions will need a person to approve them before they run?
5. If a model provider changes or retires a model, who updates the automation, and how is that billed?
6. What do we own at the end, and what do we need to keep running it without you?

Keeping the engagement on one board

An agency project produces a stream of small decisions and requests: this workflow, that exception, the bot should never quote a delivery date. fenbs can hold that list. A team board per client lets the agency’s people work the four lanes (To Do, Next Up, In Progress and Completed) while the client joins with the suggested Client role to see and comment, or Reporter to add tasks too. If the agency connects its AI assistants over MCP, each change they make is recorded in History by name, so the client can tell what a person did. Standing instructions, such as what the bot may promise, go on the Decisions and rules page as rules, decided by a person, and every connected assistant reads the rules first.

It is a task board, not agency software. fenbs has no time tracking, invoicing, due dates or settable assignee, so billing and deadlines live elsewhere. The agency setup is described in fenbs for agencies, and what a client should be able to see is in client project tracker.