What an AI automation agency does, and 7 times you shouldn't hire one
Most AI automation agency pitches are a demo of a model call, which is the easiest hour of the project. What you are actually paying for sits underneath it. A schema that can hold the output. A job that survives being run twice. Systems you do not control, rate-limiting you at the worst moment. A screen someone trusts enough to approve things on. We run seven live products out of a small New Zealand studio, and almost none of that work is prompting.
So this is written from the build side. We put these systems in and then live with them afterwards, which is the part that shapes everything below.
What an AI automation agency actually builds
An AI automation agency that only shows you the prompt is showing you about a tenth of the build. The rest is the same stack as any other software project, and it gets touched at every tier.
What that looks like in practice, from work running in this studio.
The data model, before anything else
Automation writes into a schema, and if that schema was not designed to hold what the automation produces, you get plausible garbage that nobody notices until an audit. MyScrapbook Studio is at 55 migrations across a marketplace that pays artists real money, so every automated write has to land somewhere that reconciles: who earned what, against which sale, in which payout run.
That means provenance columns most briefs never mention. Which run produced this row, from which input, reviewed by whom, and can it be reversed. Retrofitting that after the fact is a migration against live data, which is the expensive way to learn it.
Backend, and the part that gets skipped
A cron job calling an API is a demo. Production automation is a queue with retries, backoff, and idempotency, because the interesting question is what happens when the same job runs twice.
MSS carries 82 services behind subscriptions, marketplace payouts and asset processing. A payout job that runs twice is not a bug report, it is money out the door. Making a job safe to re-run is usually more work than making it run, and it is the first thing dropped when someone is quoting on the build alone.
Middleware: the systems you do not control
Most of the real engineering in an AI automation agency project is integration. Stripe, App Store Connect, Play billing, Supabase, the model APIs themselves. Each brings its own auth model, its own rate limits, its own pagination, and a sandbox that behaves differently from production in ways you find out about on a Friday.
This layer is where estimates go wrong, because it is the one part you cannot inspect before you start. A quote that treats “connect to their existing system” as a line item has not read the API docs.
The client, where the complexity actually surfaces
Frontend gets treated as decoration on automation projects, and then it eats the schedule. Every ambiguity in the system has to end up visible on a screen. What state is this in. What changed. What failed overnight, and who needs to decide something about it this morning.
Most of that is interaction design, settled before anyone writes a component. The states nobody puts in a brief are where the time goes: empty, loading, half-finished, stale, failed. A tool somebody sits in all day needs keyboard paths. By the twentieth screen it needs a design system, or screen twenty looks like a different product.
Accessibility and responsive behaviour belong in that same phase. Both change how components are structured, so bolting them on later is a rebuild rather than a pass.
Performance depends on what you are rendering, and the problems are not interchangeable. Layerpaint paints 3D models in the browser through Three.js, so the work is draw calls, texture memory, and keeping the interface responsive while the GPU is busy on hardware we do not get to choose. MyScrapbook Studio is a design editor: layers, drag and drop, undo history, autosave, and export at print quality instead of screen quality. The Flutter apps are offline first, so the client owns a local database and has to resolve conflicts once the network comes back.
For automation work the review screen carries the whole risk. If checking the output takes longer than doing the job by hand, people quietly stop using it inside a month, and the model takes the blame for what was a design problem. So the screen has to make each decision small, and it has to leave a record of who made it.
Then the operational layer
This is the least glamorous tier and the reason things keep working. Alerting when a job fails, logs you can actually read six months later, and a scheduled check that the output still looks sane. Our sitemap submission runs daily through IndexNow and the Search Console API because both engines retired anonymous pings, and the job verifies the file still serves rather than trusting that it does.
A person checks the output before it counts. That sentence is the difference between automation that survives a year and automation that gets switched off in week three, and any AI automation agency that skips past it is selling you something that will quietly rot.
The part that gets underestimated
Almost nobody budgets for the boring middle. Credentials expire, APIs deprecate endpoints on their own schedule, and a job that has run silently for four months will fail on a day when the output looked plausible enough that nobody noticed.
Anthropic’s own documentation is blunt about the need for evaluation and error handling around model calls, and that guidance holds whether you build it yourself or buy it. The build is maybe a third of the cost. The rest is watching it.
When you don’t need an AI automation agency
Here are seven situations where hiring an AI automation agency will cost you money and return nothing. We have hit most of these ourselves.
1. The task runs less than weekly. Automating something that happens monthly usually costs more in build and maintenance than the person spends doing it. Sit down and multiply the minutes by the frequency before anyone quotes you.
2. Nobody has written the process down. If the steps live in a single staffer’s head and change depending on the day, there is nothing to automate yet. Write the runbook first, follow it manually for a month, then talk to an AI automation agency about the version that stabilised.
3. The step should be deleted, not automated. A surprising share of recurring work exists because of a decision made three years ago that nobody has revisited. Automating it makes that decision permanent.
This is the cheapest win available to most teams, and no agency will find it for you. Deleting the work removes their invoice.
4. Something off the shelf already does it. Zapier and its competitors cover an enormous amount of ordinary connective work for a monthly fee smaller than a single consulting day. Try the boring tool before commissioning a custom one.
An AI automation agency earns its fee when your case genuinely falls outside what those tools reach, which is rarer than the sales page implies.
5. The input is on paper or in a phone call. Automation needs a machine-readable trigger. If the process starts with a scribbled form or a voicemail, your first project is digitising the intake, and that is a different job at a different price.
6. You cannot describe what a good result looks like. You need a check that a machine can run, or a person who will look at every output and say yes or no. Without one of those you have built a system that produces confident output with no way to tell whether it is right.
7. Nobody will own it after handover. Every automation needs someone whose job includes noticing when it stops. Where that person does not exist, the work degrades into a broken job that everyone assumes is still running.
If three or more of those describe your situation, the answer is to fix the process, and no AI automation agency can do that part for you.
What AI automation work in NZ actually costs to keep running
Search “ai automation nz” and you find a thin list of providers, most of them new, most of them pricing on the build. That pricing shape hides the real number. A job that calls an external API is a standing commitment to that API’s release schedule.
Our sitemap jobs are a fair example. They were quick to write and they have needed attention twice since, both times because the platform changed underneath them rather than because our code was wrong.
So when you scope work with an AI automation agency, ask what happens in month seven. Ask who gets the alert, who has the credentials, and what the fix costs. An AI automation agency that only quotes the build is quoting a third of the project.
What we would ask for if we were buying
Ask for the smallest possible first project, priced fixed, with a working result inside two weeks. Ask to see the failure path before the happy path. Ask what the manual fallback is on the day the automation is down, because there will be such a day.
Anyone confident in the work will answer those quickly. An AI automation agency that gets vague around the failure path is telling you something useful.
How we decide, in practice
The rule we use is simple enough to apply in a meeting. If the task recurs at least weekly, has a written procedure, and produces output a person can verify in under a minute, it is worth automating. Miss any one of those and it goes back on the shelf.
That rule is why our automation list is short. Sitemaps, Search Console, App Store reviews, and the content gate all clear the bar comfortably. Plenty of other candidates did not, and the studio is better for the ones we declined to build.
Seven live products come out of this: Layerpaint, MyScrapbook Studio, and five iOS apps built in Flutter. The fastest of those went from first commit to live on the App Store in six days, GastroLog.
None of that speed came from a clever model call. It came from clearing out the weekly admin that would otherwise eat the days a senior developer needs for the product itself. The automation is support work, and treating it that way is why it stays useful.
The same thinking works on the buying side. An AI automation agency is a way to buy back specific hours, and you should be able to name those hours before you sign anything.
If you want to see what else this studio has shipped and how it was built, the studio homepage has the current product list and Received Wisdom has the rest of the arguments. If you are weighing up an AI automation agency for your own work, start by counting the minutes. That count usually answers the question before anyone quotes you.