# What Does an AI Automation Agency Actually Do? > Source: > Published: 2026-09-20 13:00:10+00:00 TL;DR: the label covers four very different businesses. The one you want ships to production and can show you a system running right now. The single most useful question you can ask is what happens when the automation breaks at 2am, because only one of the four kinds has an answer. The term did not exist a few years ago and now there are thousands of them. A large share are one-person operations that learned Make from tutorials and put 'agency' on a landing page. A smaller share are engineering teams building systems that run a business. Both use the same words on their websites and the output is not comparable. I run one of these, which you should factor in. What follows is the version I would want if I were buying, including the parts that are inconvenient for me to write. Before evaluating anyone, work out which of these you are talking to. Each is legitimate for some jobs and wrong for others, and the confusion is where most bad engagements start. Configures Zapier or Make flows between standard apps. Fast, cheap, genuinely useful for simple work. Cannot help you when the requirement outgrows what the platform does natively, and typically has no answer for error handling beyond what the platform provides. Right choice for a two-app flow. Wrong choice for anything your operations depend on. Sells one product, usually a support bot, configured to your content. Can be excellent at that one thing. Becomes a problem when the actual bottleneck is elsewhere and the answer is still a chatbot, because that is what they sell. Produces strategy, process maps, a roadmap, and a recommendation to engage someone for the build. Valuable in a large organisation with real change management to do. In a company of twelve people it usually converts budget into a document. Designs and builds systems that run in production, including the unglamorous parts: retries, error handling, alerting, logging, and a handover. Costs more than the connector shop and delivers something you can rely on with money or with customers. This is the category worth paying for when the process matters. A real engagement has stages, and the ones that get skipped are always the same ones. Mapping what you actually do, not what the documentation says you do. This is where the surprises live: the step somebody added two years ago, the spreadsheet nobody mentioned, the customer type that gets handled differently for historical reasons. Half the value of a good agency is delivered here, before anything is built. Deciding which platform, which model, which integrations, and crucially what is deliberately left manual. Good design includes the decision not to automate something. If a vendor proposes automating 100% of a process on the first pass, they have not thought about the exception cases yet. The automations, the prompts, the integrations, and the error handling. The error handling is what distinguishes the price bands. A flow that works on the happy path can be built in a day. A flow that behaves correctly when the third-party API returns a 500 halfway through takes considerably longer and is the reason you are hiring someone. Not testing that the demo works. Testing with your actual malformed inputs, your duplicate records, your supplier who sends invoices as photographs. Anyone can pass a test they wrote themselves with data they chose. Putting it into production with a way to see what happened and an alert when something did not. A system that fails silently is worse than no system, because you will trust it for the three weeks before you find out. Documentation, access, and a named person who can change it later. This is the stage most commonly skipped and the most expensive to skip. Ask these of anyone you are evaluating. What matters is not whether they have an answer, it is whether the answer is specific. Can I see a system you built that is running in production right now, and can I speak to the client running it? What happens when an integration fails at 2am? Walk me through it. How do you find out something broke: does the system tell you, or does the client tell you? What does the handover include, and who can maintain this if you disappear? Who owns the accounts, the code and the data when this ends? What did you get wrong on your last project and what did it cost to fix? That last one is the most informative question on the list. Everyone who has shipped real systems has a story. Anyone who says nothing has gone wrong has either not shipped much or is not being straight with you. **Red flags:** Demos only in test environments, no live systems to point at Promises to automate complex judgment-heavy work end to end No documentation or handover plan in the proposal Hourly billing with no fixed scope and no estimated total Cannot explain what happens when the automation breaks A maintenance retainer with no list of what it covers Case studies with no client names, no numbers, and no way to verify them The shape is consistent across the good ones, whatever their size. It starts with a process audit: a conversation where they map what you do, say what is automatable, and give a rough return estimate. This should cost nothing and take 30-60 minutes. If it costs money and produces only a document, you are talking to the consultancy. Then a fixed-scope proposal: what gets built, what it will do, what it costs, when it lands. Variable scope with hourly billing can run indefinitely without a clear outcome, and the incentives under hourly billing point the wrong way for you. On timelines: a focused automation for a single process should take two to four weeks to build and deploy. A system touching several departments takes four to eight. Anything quoted materially longer without a specific reason deserves a question, and anything quoted materially shorter usually means the error handling is not in scope. This is the part clients most often discover too late, and it is worth putting in writing before anyone starts. The accounts. The Make, n8n, OpenAI and Claude accounts should be yours, billed to you, not sitting inside the agency's workspace. The code and the flows, exported and in your possession. The data, in a format you can read without their tooling. Documentation good enough for a competent third party to take over. The prompts. They are part of the system and they are often where the real work is. An agency that resists this is selling you a dependency rather than a system. There are reasonable middle grounds, for instance them holding operational access while you hold ownership, but the default should be that you can walk away and the thing keeps running. Roughly: a single flow $500-1,500, a real business process $1,000-3,000, a reliable pipeline with an AI layer and error handling $3,000-8,000, a custom system replacing a back office $8,000-25,000. Running costs $50-200/month for most small businesses. The full breakdown, including the costs that do not make it into proposals, is in our pricing article. Sometimes the right advice is that this is not a purchase. If the job is one flow between two apps you already pay for, and someone on your team is comfortable with a bit of configuration, build it yourself in an afternoon. Make and n8n are genuinely approachable for that tier of work, and paying someone four figures to do it is paying for confidence rather than capability. The line worth paying across is reliability. When the process touches money, customers, or compliance, and a silent failure would hurt, that is the point where the unglamorous engineering is the product and it is worth buying. Our case studies are public with real clients, real stacks and real numbers. If you want to work out what your business actually needs, the audit at 2pizza.team/audit takes two minutes and no call. *Originally published at [2pizza.team](https://2pizza.team/blog/what-is-ai-automation-agency). We build AI and automation systems for small teams - fixed price, two to six weeks. [See the work](https://2pizza.team/work).*