AI agents now do most of our agency admin. Here's what stuck.

We run a Melbourne web agency on a set of AI agents that do the admin: the reporting, the triage, the research, and the syncing between systems that nobody wants to do by hand. Eight of our twelve recurring admin jobs now happen without anyone starting them. This is what we built, what it costs to keep alive, and the work we deliberately left to people.

Search for "AI automation agency" and look at what comes back. Most of it is people selling a course on how to start one. The most-upvoted discussion we found on the subject is somebody asking, reasonably, whether the whole category is "a profitable business or just an AI gimmick."

Fair question, and here is our answer from the only position we can honestly speak from. We build AI automation, and we run it on ourselves first. Not as a demo. As the actual operating system of the business, on the work nobody enjoys doing twice.

The short answer: automate the gathering, keep the deciding. If a task's answer already exists in writing, an agent can do it, and will do it every week without being asked. If it ends in a judgement somebody is accountable for, the agent gathers and the person decides. Then build the check that tells you when the automation has stopped, because it will, and nothing crashes when it does.

Jump to:

• What is worth automating
• Check a task of your own
• What we run on our own business
• Common questions

What is worth automating first?

The work that already has a written answer. The same test decides what a support bot is allowed to say, and it holds across every kind of business process. We covered the support case in its own article.

If the answer to a task already exists in writing somewhere, it is a candidate. If it depends on judgement, an exception, or something only one person knows, it is not.

That single rule kills most of what people try to automate first. Chasing a client about an overdue invoice looks repetitive, but the decision about how hard to chase this particular client this particular week is judgement, and it is exactly the sort of thing that damages a relationship when a machine gets it wrong.

Compiling the list of who is overdue, however, is not judgement. It is a query.

Almost every process in a business splits along that line, and the split is usually obvious once you look for it.

What do we run on our own business?

Twelve recurring admin jobs, and this is the state of every one of them. Eight run without anyone starting them. Three run up to the point where a person has to make a call. One we have deliberately left alone.

Three shapes account for most of the value, and none of them replaced a person. They removed the friction that stopped a useful thing happening consistently.

Reporting that arrives whether or not anyone remembers. Time worked, proposals outstanding, what changed this week. None of it is analytically clever. The value is entirely that it turns up without anyone deciding to produce it, which means it also turns up in the weeks when everyone is too busy, which are precisely the weeks you most need to know. A report you have to remember to run is a report you stop running in month three.

Message triage. Enquiries and messages arriving across several channels, read, classified and routed to whoever should see them, rather than everything landing in one inbox and being sorted by whoever opens it first. This is the automation that most changes the texture of a week, because it removes a category of small decisions rather than a category of work.

Research and audit prep. Before a sales conversation, pulling together the technical state of a prospect's site: what is slow, what is missing, what is broken. That used to be an hour of unbillable work that often did not happen, so the conversation started vaguer than it should have. Now it happens every time, because the cost of doing it fell to nearly nothing.

Underneath all of it is one piece of plumbing. Our assistant reads the business directly through a gateway: time entries, the pipeline, invoices, analytics, Search Console, the tag manager, and a registry of which system owns which fact. Each is a small connector holding its own credential, and the assistant asks the connector rather than a person. That is what "agents running the admin" means in practice. It is not a robot. It is a set of questions the business can answer without interrupting anybody.

What does running your own AI tooling actually cost?

An operations function you did not previously have. This is the part missing from every AI automation pitch we have read.

If you run your own AI tooling, you are now operating infrastructure. For us that means a set of connectors, each with its own revocable credential, reachable only through a gateway. Servers need patching. Security updates need a maintenance window. Credentials expire on a schedule somebody has to know about. Services that talk to each other occasionally stop. All of it unglamorous, none of it optional.

That is why the weekly checks exist, and why they are built to go red when they cannot confirm something rather than staying quiet. A check that passes when it has verified nothing is worse than no check, because now you trust it.

For us that trade is worth it, because the return is real and we have the skills in-house. For many businesses it is not, and the honest recommendation is often to use somebody else's hosted product and accept the constraints. An automation that saves two hours a week and costs three hours a month to keep alive is not a win, and you will not discover that until month four.

Anyone quoting you a build cost without describing the running cost is quoting half the project. It is the first thing to ask an agency, including us.

How does an automation fail without telling you?

Quietly, while reporting success. The dangerous failure is not the agent that crashes. That one announces itself.

There are three shapes this takes. We design against all three, because they are the ones that survive a demo and turn up in month seven.

The sync that looks full. One record in a batch carries a field the database rejects, the loop catches the error and moves on, and the copy ends up looking complete while missing a whole category of record. Every sync we build reconciles its counts at the end and reports both numbers, because "completed" hides this and a copied count beside a skipped count does not.

The job that runs twice. A scheduled task overlaps itself, or somebody re-runs it by hand, and the second pass does not know about the first. Totals double, and a doubled total looks like a good month rather than a fault. Anything that writes gets a key that makes a repeat write harmless.

The alert that counts the wrong thing. A notification fires on an event that is not quite the event you care about. Automated traffic reads as interest, a retry reads as a second visit, and the number climbs for reasons nobody checks. Whatever a number will be used for, filter it at the point it is recorded rather than the point it is read.

Apply that to any automation you have come to rely on. The weekly report stops including a section. The triage stops catching one category of message. Nobody notices, because the thing you built the automation for was not having to pay attention.

So build the check in from the start:

• Make silence loud. If the report did not arrive, that should be visible. An automation that only speaks when it succeeds cannot tell you it stopped. Ours are built so that a run which verifies nothing fails.
• Log what it did, not just that it ran. "Completed" is not information. How many records it copied, and how many it skipped, is.
• Check the total before the detail. A number that is too good, too low or too round is the fault announcing itself in the only voice it has.
• Look at the output monthly, deliberately, against reality. Not because you expect a problem, but because that is the only way you will find one.

What have we deliberately not automated?

The list is more useful than the enthusiasm.

• Anything that speaks to a client in our voice without a person reading it first. Drafting is fine. Sending is not.
• Anything touching money. Quotes, invoices, and what a project should cost. The gathering is automated. Every judgement is not.
• Anything where being wrong is expensive and being slow is cheap. That is most of the interesting work, and it is a genuinely good filter.

The same filter runs the other way in our client work. The automations we build for other businesses sit on the gathering side of the line, and the ones we talk people out of are the ones that would have put a machine in a conversation. Our case studies are mostly websites, because for most of the businesses we work with the website is the automation: it answers the same questions, every day, without anyone being asked.

How do you start?

Not with a platform. With one thing that annoys you every single week, and three questions about it.

1. Does the answer already exist in writing?

If not, write it down first. That is usually the actual project, and it is the one step every vendor skips because it cannot be sold. A process nobody has written down is a process that lives in one person's head, and an agent cannot read a head.

2. What happens if this runs and produces nothing?

If nobody would notice, you are building something you cannot trust. Decide now what an empty result looks like and who sees it. That decision is cheaper on day one than on the day you find out.

3. Who maintains it in six months?

If the honest answer is nobody, use a hosted product instead of building. A hosted product's constraints are annoying. An orphaned automation is a liability that looks like an asset.

Get one thing working. Live with it for a month. The second is far easier than the first, and the third teaches you nothing new.

Common questions about AI automation

Will it still be running in year two?

Build the automation your team can still trust in year two, once the person who set it up has moved on and nobody remembers what the empty result is supposed to look like.

Every automation demo shows the same thing: a trigger, a result, a saved hour. None of them show the Tuesday in month seven when a credential expires, the report arrives with a section missing, and everyone reads it anyway because that is what they do on Tuesdays.

That is the day worth planning for. Automate the gathering, keep the deciding, and build the check that makes silence loud. Do that and the rest is plumbing, which is the part we like.