To choose an AI automation agency, ask for a recorded run of something they built, find out who writes and maintains the code, and get the project tied to one number you can check, such as hours saved or leads answered inside a minute. Those three answers sort agencies faster than any pitch. The rest of this post covers the questions behind them, what we would put in the scope, and the mistakes we see buyers make.
We are an AI automation agency ourselves, so weigh this accordingly. We have tried to write it as if you were interviewing us and two other shops next week.
What an AI automation agency actually does
Most of the work is plumbing. A form fill needs to land in the CRM, a reply needs to be read and sorted, a paid deposit needs to start an onboarding plan. An AI step goes in where a person used to read and judge something, and plain code handles everything that has to be exact.
That split matters when you compare agencies. Our AI agent development services page says the same thing: some builds are agents that decide what to do next, but most are AI steps inside a workflow, which is usually the more reliable choice. An agency that sells you an "autonomous agent" for a job a three-node workflow would do is selling you maintenance.
Five questions that sort agencies quickly
1. Can you show me a run, not a screenshot?
A logo grid and a hero screenshot prove nothing. A recorded run shows the trigger, each step, what the AI was asked, and where it stopped. Ours are public. One of them, the Stripe deposit to project kickoff automation, verifies a signed webhook, drops duplicates, and halts with a stated reason when a test payment arrives with no matching proposal. That stop-and-say-why behavior is the thing to look for in anyone's demo.
Those runs were recorded on an invented business, and the page says so. If an agency shows client results, ask whether the client agreed to be named. If it will not name anyone, ask why the case study has a number in it.
2. Who writes the code, and who maintains it?
Some agencies resell a low-code platform and configure it. Some write custom code with an AI coding assistant and review it. Both can be fine. What goes wrong is when nobody can say which one is happening, or when the person who built it is a contractor the agency will not name.
We build with n8n, Make, GoHighLevel and Claude Code, depending on the job, and we say which on every scope. Ask the same of whoever you are talking to: which tools, which accounts they live in, and who has the logins on day 91.
3. What number will this be judged on?
"Automate our lead follow-up" is not a scope. "Every inbound reply is classified and a person is alerted inside a minute" is. A good agency pushes you toward a measurable line before it quotes, because it wants a fair acceptance test as much as you do.
If you cannot name a number, that is useful information too. It usually means the first project should be a short audit of where the hours go, not a build.
4. Where does a person stay in the loop?
Language models are wrong sometimes and sound sure when they are. The question is what happens then. In our reply-handling builds the AI classifies and a person decides on anything ambiguous. In our verified lead list build, the run stops for the owner's go before bulk credits get spent, and catch-all emails never reach the sending tool.
Ask any agency for its three most recent human-handoff points. If the answer is "there aren't any," budget for cleaning up a mistake in public.
5. What happens when it breaks?
It will break. A connected tool changes a field, an API key expires, a model update shifts how a prompt behaves. Ask what the agency monitors, how you find out, and whether fixes after handover are a retainer, an hourly rate or a new project. There is no right answer, but there needs to be an answer.
Agency, freelancer or in-house
| Option | Fits when | Watch for |
|---|---|---|
| Freelancer | One clear workflow you can review yourself | Single point of failure, credentials in their accounts |
| Small agency or studio | Several systems, ongoing monitoring, a handover | Reseller configuration passed off as engineering |
| In-house hire | Automation is a daily need, not a project | Salary before you know what to build |
We are a small engineering shop, not a large bench of developers, and we say so plainly. That is a good fit for a small business with a handful of systems and a bad fit for a company that needs forty workflows migrated by quarter end. Pick the size that matches your job.
What we would put in the scope
Here is a worked example of the structure we use, written as a scope for a reply-handling build. The figures are placeholders for you to replace, not quotes.
- Trigger: a reply lands in the sending tool's inbox.
- AI step: classify as interested, not now, wrong person or automatic reply.
- Plain code step: create or update the CRM record and drop duplicates.
- Human step: interested replies alert a named person in chat.
- Acceptance test: ten sample replies, run end to end, with the expected class for each written down before the build starts.
- Handover: workflow export, a one-page runbook, and the logins in your accounts.
The acceptance test is the part buyers skip and later regret. Writing the expected answers down first means nobody argues about whether the build works.
Mistakes we would avoid
Buying on tool names. "We are n8n experts" says nothing about whether the workflow handles a duplicate webhook. Ask about the failure case instead.
Skipping the handover. We have inherited builds where the previous builder held the only admin login. Put account ownership in the contract on day one.
Starting with the most complex job. The first project should be small enough to finish in a few weeks, so you learn how the agency works before the big one.
Treating AI as the whole product. In our experience the AI step is a small part of the build and the rest is error handling, deduplication and logging. If a proposal spends three pages on the model and one line on error handling, ask about that line.
Pricing and tool costs, as of November 10, 2026
We have not quoted any vendor plan prices here, because they change and you should read them at the source. Before you compare quotes, check the current pricing pages for each platform in the proposal (n8n, Make, GoHighLevel, and whichever AI model provider is used) and ask the agency which plan tier the build assumes. A fair proposal separates the project fee from those monthly tool costs, which you pay the vendors directly. If a quote bundles them with no breakdown, ask for one. Our n8n pricing post covers how that tool bills.
What we would do in your position
Shortlist two or three agencies and give each the same one-paragraph problem. Ask for a scope with a number and an acceptance test, not a price. The agency that pushes back on your problem statement, asks which system holds the source of truth, and tells you what it would not automate is usually the one that has built this before.
Then run a small paid first milestone. Keep the logins in your own accounts from the start. You will know within a few weeks whether to continue. If you want us in that comparison, our AI agents page lists what we build and what we do not, and you can see working examples under automations.
Frequently asked questions
- How do I choose an AI automation agency?
- Ask each agency to show a recorded run of something they built, ask who writes and maintains the code, and ask what the number is that the project will be judged against. Agencies that answer all three with specifics are worth a scoping call. Agencies that answer with a tool list or a slide deck usually have not shipped much.
- How much does an AI automation agency cost?
- It depends on the scope, and any agency quoting a flat figure before it has seen your systems is guessing. Most small-business work is priced per project with a fixed first milestone, plus the usage costs of the tools themselves, which you pay the vendors directly. Ask for the project fee and the monthly tool costs as two separate numbers.
- Should I hire an AI automation agency or a freelancer?
- Hire a freelancer when the job is one well-defined workflow and you can review the result yourself. Hire an agency or small studio when the work touches several systems, needs monitoring afterward, or cannot stop if one person is unavailable. Either way, the questions are the same: who owns the credentials, who fixes it when it breaks, and what do you get at handover.
- What should be in an AI automation agency contract?
- The contract should name the deliverables, the acceptance test for each milestone, who owns the workflows and accounts at the end, and what happens when a connected tool changes its API. It should also say where your data is processed and which AI model providers see it. If any of those are missing, ask for them before signing.