How to choose an AI automation agency: the questions to ask

The AI automation space exploded over the last two years, and so did the number of people calling themselves an agency. The automation-as-a-service market is heading from roughly 10 billion dollars in 2025 to around 36 billion by 2030, and a lot of that growth is firms built by people who learned the tools last year and slapped up a website. Some are excellent. Plenty will sell you something that looks brilliant in a demo and falls over the moment real work hits it.

This is the buyer’s side of the table: the questions that tell a real builder from a deck-seller, in the order worth asking them. The one to hold above all the others is ownership. A good agency builds you something you could run without them. A bad one builds a dependency and quietly rents it back to you. Get a pen, here’s exactly what to ask.

1. First, work out if you even need an agency

Before you hire anyone, check whether the job is something you could do yourself for the price of a subscription, because in 2026 a surprising amount of it is. No agency writing a guide like this will tell you that, for an obvious reason. We build this stuff and we’d still rather you didn’t pay us for work you don’t need.

The honest line is about complexity. A single, self-contained automation, “when a form comes in, add the row and send the welcome email”, is now well within reach of a non-technical operator using the AI you already pay for. Claude with connectors or ChatGPT can walk you through it, and an off-the-shelf tool like Zapier or Make will run it for the cost of the plan. If that’s your whole problem, you don’t need an agency, you need an afternoon.

You need an agency when the work spans several systems that have to stay in sync, needs custom logic that no off-the-shelf tool quite fits, touches sensitive data, or has to be reliable enough that you genuinely can’t babysit it. That’s real engineering, and it’s worth paying for. So the first question is the one you ask yourself: is this a job, or is it a system? Pay for the system. Do the job yourself.

The trap. “We could automate that for you” is true of almost anything, including the thing you could’ve set up before lunch. An agency that takes a one-step automation and scopes it into a five-figure project isn’t building you a system, it’s padding a bill. Know roughly where the do-it-yourself line sits before you take the call, so you can tell.

2. Ask them to show you something they’ve actually built

Lead with this, because it’s the fastest filter there is: “Show me a specific system you built, the problem it solved, and the number it moved.” Not a logo wall, not a testimonial, not “solutions we’ve built for businesses like yours”. A named client or at least a real, described one, a specific workflow, and a before-and-after with actual figures.

The difference is everything. “We helped a retail company improve efficiency” means nothing. “We built invoice processing for a logistics firm that cut manual handling from four hours a day to twenty minutes” is a real claim you can probe. NVZN, an agency that ranks for this exact question, puts it plainly in their own buyer’s guide: “we saved this client 15 hours per week” beats “we helped optimise their workflow” every time.

Watch for the demo dodge. A prototype that works in a controlled demo is not a system running in production with real, messy inputs, and the two are worlds apart. If all they can show is a slick demo or a proof of concept, they have demo experience, not production experience. Ask whether you can speak to a client whose system is live and working right now. A confident builder shares that readily. A seller starts talking about NDAs.

3. Ask what they need to understand before they quote

A good agency can’t price your job on the first call, and a good one knows it. So ask directly: “What do you need to understand about my business before you can scope this?” The answer tells you whether they’re thinking about your problem or reaching for a template they apply to everyone.

You want them asking the questions back at you. What’s the current process, step by step? How clean is your data? What systems does it need to plug into? What does done actually look like, and how will we know it worked? What are your compliance constraints? A proper discovery process surfaces the awkward stuff, the data that’s a mess, the integration that’s harder than it looks, the edge case you forgot, and that’s exactly the stuff that changes the scope and the price. They’re doing you a favour by finding it early.

The trap. A polished proposal landing in your inbox within a day of the first call is not a sign of efficiency. It’s a template with your name pasted in. Nobody can scope an automation accurately without understanding your data and your systems, so a fast quote means they’ve either skipped that or guessed it. You’ll pay for the guess later, when the build hits the reality they never asked about.

4. Ask who owns the code, the prompts and the configs

This is the question that matters most, and the one buyers forget to ask until it’s too late. Put it flatly: “When this is finished and paid for, who owns it? All of it?” The answer you want is you. The code, the prompts, the workflow configurations, the documentation, and a clear path to your own API keys, transferred to you on final payment and written into the contract in black and white.

There are three ways this goes, and Automely lays them out honestly in their own guide. Full ownership transfer: everything custom is handed to you, you control it, low risk. Managed service: the agency runs the system on their infrastructure and you reach it through their dashboard, which means you depend on them for every change, every price rise and every minute of uptime. Hybrid licence: your custom parts are yours, but a platform layer underneath is licensed from them, so you need to know exactly which is which. Ownership transfer is the one to insist on for anything custom you’ve paid to have built.

The reason this is the spine of the whole decision: ownership is what separates a partner from a landlord. If you own it, you can run it, change it, move it, or hand it to a different team. If you don’t, you’re not a customer who bought something, you’re a tenant who’ll keep paying rent for as long as you want the lights on. Vagueness here, “it runs on our platform”, “we handle all that for you”, is not convenience. It’s the dependency being built on purpose.

5. Ask what it’s built on, and whether you could run it without them

Ownership on paper means little if the thing is built so you could never actually use it. So go one level deeper: “What’s this built on, and if we parted ways tomorrow, could I or another developer pick it up and run it?” You’re testing for portability, and the build choices give it away.

Portable looks like open, exportable tools and your own accounts underneath. n8n, for instance, is open-source and can be self-hosted, so the whole workflow can live somewhere you control with no per-task pricing and no platform holding it hostage. Make scenarios can be exported. Your automations run on your own API keys, not the agency’s. The code sits in a repository you can access. None of this requires you to be technical. The point isn’t that you’ll personally self-host anything, it’s that you could hand the lot to any competent developer and they’d be up and running, because nothing essential is locked away.

Locked-in looks like the opposite: the core logic lives inside the agency’s own proprietary platform, the keys are theirs, and “exporting it” means rebuilding it from scratch somewhere else. That’s the rent trap dressed up as a tech stack.

The maturity call. Be honest with yourself about what you’ll actually do here. Most operators will never log into n8n or read a line of the code, and that’s completely fine. This test isn’t about you running it. It’s about whether you could hand it to someone who would. A build you can move is a build you own. A build only the people who made it can touch is one you’re renting, whatever the invoice says.

6. Ask what happens when it breaks

It will break. Not might, will. APIs change, tools update, a supplier sends an invoice in a format nothing’s seen before, and the workflow that ran perfectly for six months hits something it can’t handle. Any agency that implies their system won’t fail has either never run one in production or isn’t being straight with you. The real question is what happens next.

So ask: “How do you know when something’s gone wrong, and what’s the process to fix it?” A serious builder has real answers. Monitoring and alerts that flag a failure before you find out from an angry customer. A defined way for the system to flag low-confidence output for a human to check rather than acting on a guess. A response time for critical breaks. Documentation good enough that someone other than the original builder can fix it. This matters most for anything touching money, customer messages or compliance, where a quiet wrong answer is expensive.

The answer that should worry you is the reassuring one: “it’s very accurate, it won’t really make mistakes.” That tells you they’re thinking about the demo, not the year after it. There’s more on where AI systems break, and the human checkpoint you keep on anything that spends money or goes public, in our explainer on agentic AI for business owners.

7. Ask how they price it, and what you’re paying for

Pricing in this space is all over the map, which is exactly why you need to pin it down. The same words, “AI automation”, cover a 500-dollar Zapier setup and a six-figure build. NVZN’s rough public guide is a fair anchor for small business: a simple automation connecting a few apps runs about 500 to 2,000 dollars, a custom multi-step workflow 2,000 to 10,000, a full system build 10,000 to 25,000 or more, and ongoing support somewhere between 500 and 5,000 a month. Other agencies quote several times that for the same labels, which tells you the label is meaningless. Only the scope sets the price.

Simple automation $500–2k

Multi-step workflow $2k–10k

Full system build $10k–25k+

What you’re really judging is the structure, not the number. A healthy shape is a fixed price for a defined build, paid against milestones, with an optional and clearly described monthly retainer for support afterwards. That’s honest: you can see what you’re buying and what it costs to keep it running. The shapes to avoid are full payment upfront, which removes their reason to deliver on time, a fixed quote for a scope nobody’s defined yet, which is a guess you’re underwriting, and an open-ended retainer for work you can’t see or measure, which is just a subscription to your own software. If the retainer is bigger than the build and you can’t point to what it buys each month, that’s the rent trap again.

Red flags: walk away if you hear these

Most of the warning signs are the inverse of the questions above, and any one of them on its own is a reason to slow down. A proposal within a day, with no real questions about your business first. Promises of “10x your revenue” or “replace your whole team”, which are sales, not engineering. No verifiable case study with real numbers, only demos and “trust us”. Vague or evasive answers about who owns the finished work. No plan for what happens when it breaks. Full payment demanded upfront instead of staged against milestones. Jargon used as a wall, where they can’t explain in plain words how the thing will actually work.

And the expensive quiet one, because it rarely shows up as a single dramatic moment: a build you can’t run, see or move without them. It feels fine on day one, when everything works and they’re handling it all. It bites in month eighteen, when you want to change something, or change agencies, and discover you can’t, because you never owned the thing you paid for.

Do it yourself, buy off the shelf, or hire someone

Do it yourself when it’s a single, simple automation. The AI you already pay for plus a tool like Zapier or Make will handle it for the price of a subscription, and you’ll understand your own systems better for having built it. Don’t hire out an afternoon’s work.

Buy off the shelf when a finished product already does your job. If standard software covers what you need, a custom build is a worse deal, more expensive and more to maintain, for no extra benefit. Reach for a build only when your process is genuinely yours and the ready-made tools fight it.

Hire an agency when the job is a real system: several tools that have to stay in sync, custom logic, sensitive data, reliability you can’t personally supervise. When you do, run every agency through the seven questions above, and weight ownership heaviest. The right one builds you something you could run without them and is happy to hand it over, because their pitch is the quality of the work, not control of the off switch. If you also want the capability to live in your own team afterwards, our take on agentic AI for business owners and the back-office automation map are the places to start, and they’ll tell you a fair bit about what to expect a good agency to do for you, too.