"AI automation agency" barely existed as a phrase two years ago. Now there are thousands, and a significant share of them are one person, a workflow builder and a course they bought about how to sell workflow builders.
That's not a reason to avoid the category — the underlying work is real and the returns are unusually good. It's a reason to know what you're buying, because the gap between the best and worst version of this service is larger than in almost any other kind of software work, and it isn't visible in the pitch.
Here's what the work actually consists of, where it goes wrong, and what we'd ask if we were the ones hiring.
What the work actually is
Strip away the branding and an AI automation project is nearly always one of four things.
Moving data between systems that should already talk. The unglamorous majority of it. An order lands in one system, someone retypes it into a second, exports a spreadsheet, and emails it to a third. No AI required for most of this — it's plumbing, and it's usually where the largest single block of wasted hours is hiding.
Turning unstructured input into structured data. This is where language models genuinely changed what's possible. Invoices, CVs, tender documents, support emails, contracts, call transcripts — anything that arrives as messy human text and has to become fields in a database. Before 2023 you needed a template per document type and it broke constantly. Now it mostly works, and this is where the most value is created.
Drafting and triage. Classifying incoming requests, routing them, drafting a first response, summarising a long thread into a decision. The model doesn't finish the job; it removes the blank page and the sorting.
Monitoring and alerting. Watching a source — a portal, a set of listings, competitor pricing, a regulatory feed — and telling the right person when something changed and why it matters. Cheap to build, and often the thing clients end up valuing most, because it replaces a task nobody was reliably doing at all.
Notice what isn't on the list: an autonomous agent that runs your business. That gets sold a lot. It doesn't survive contact with real operations, and building the four things above properly is worth considerably more than a demo that impresses for a week.
Where these projects actually fail
Not in the model. Almost never in the model.
Nobody mapped the process first. The most common failure is automating a process that shouldn't exist. Half the steps in a typical workflow are there because of a system limitation from six years ago or a person who left. Automating those steps preserves them forever, and now they're in code and harder to remove. The first useful deliverable in any of this work is an honest map of what actually happens — not what the process document says.
Nobody planned for the 10% the model gets wrong. A workflow that's 90% accurate is either a large win or a liability, depending entirely on what happens to the other 10%. If wrong answers land silently in a system of record, you've built something that quietly poisons your data. If they route to a human with the model's uncertainty attached, you've built something useful. This distinction is the single biggest quality difference between practitioners, and it's invisible in a demo, because demos only show the happy path.
No evaluation, so nobody knows if it's working. Ask how a change to a prompt gets tested. If the answer is "we try it and look at the output," the system will drift silently the first time a model version changes or an input format shifts. Real evaluation — a fixed set of cases with known correct answers, run on every change — is the boring difference between a system that survives a year and one that quietly stops being trusted.
It was built somewhere the client can't reach. A workflow living in a contractor's personal account, with API keys nobody documented, is a hostage situation waiting to happen. It should run in your infrastructure, with your credentials, and you should be able to fire the agency and keep the system.
The last mile never happened. The automation works, and people carry on doing it manually, because nobody was trained, nobody owns it, and the old way still exists as an option. This is a change management problem wearing a technology costume, and it kills more projects than any technical issue.
What to ask before hiring anyone
"Show me something you built that's still running." Not a demo — something in production with a client who's had it for a year. Anyone can build an impressive workflow in a week. Building one that still works after the client's CRM changed, the model was upgraded and the person who requested it left is a different skill.
"What happens when the model is wrong?" The best answer is specific and slightly boring: confidence thresholds, a human review queue, what gets logged, how errors are noticed. A vague answer here means you'll find out in production.
"Who owns the accounts and the code?" You should. Get it in writing.
"What does this cost to run per month, at our volume?" Model and platform costs are real and scale with usage. An agency that hasn't calculated this hasn't thought about your case.
"What would you tell us not to automate?" The most useful answer in the whole conversation. Anyone who says everything is a candidate is selling hours. Low-volume, high-judgement, high-consequence work is usually better left alone — and a partner who says so is telling you they're optimising for your outcome rather than the invoice.
"How do we hand this over?" There should be a documented answer that doesn't depend on the agency existing.
On pricing
The hourly model is a bad fit for this work and everyone knows it. The whole point is that the thing runs without people, so billing by the hour puts the client and the agency on opposite sides of every efficiency question.
The alternatives worth discussing: a fixed price per workflow with a defined scope, which is honest and puts delivery risk where it belongs; a monthly retainer for running and maintaining a set of systems, which is what most of the ongoing value actually is; or, increasingly, pricing tied to volume of work handled — per document processed, per ticket triaged. That last one is where the category is heading, because it's the only structure where the client pays in proportion to what they got.
Be suspicious of anything priced like enterprise software for a workflow that took four days to build. And be equally suspicious of the bottom of the market, where the price makes sense right up until you need someone to fix it.
The honest summary
The good version of this work is not exotic. It's mapping a process properly, automating the parts that deserve to exist, designing carefully for the cases where the model is wrong, testing it so it doesn't drift, and handing it over in a state where you don't need the agency any more.
That's less exciting than most of what gets pitched. It's also the version that's still saving someone ten hours a week two years later, which is the only measure that matters.
If you want to talk about a specific process rather than a demo, that's the conversation we prefer to have anyway.