An agent is a permission problem, not a prompt.
From A$7,500 — ex GST, usage at cost. An agent does work: it reads the state of things, decides the next step and acts. Which means the interesting question is never how clever it is. It is which actions it may take on its own, which need a person, and what it does when it runs out of either budget or ideas.
Sort every action into reversible and irreversible before you build anything. Drafting an email, labelling a ticket, queuing a job, writing a note: reversible, and an agent can own them. Sending the email, issuing the refund, deleting the record, publishing the page: irreversible, and those wait for a named person. That list is most of the design, it is specific to your business, and writing it with you is the first thing we do. An agent without the list is either useless, because it can do nothing, or dangerous, because it can do anything.
How it is built.
What is in the build.
The permission list, written with you
Every action it can take, sorted into reversible and irreversible, with the named approver for each. It is a document before it is code.
One step at a time, with a reason
The model decides the next step and records why. A run you can read afterwards is a run you can trust and correct.
A budget and a step limit
Per run and per month, enforced in code. When either is reached the run stops and says so rather than looping expensively.
An audit trail
Every step, its reason, what it touched, who approved it. If it does something odd, you can see exactly where it decided to.
A dry run
The whole thing with every action stubbed, so you can watch what it would have done before letting it do anything.
An off switch that works
One toggle stops it, and nothing queued runs afterwards. We hold ourselves to this: our own tools have a kill switch and the admin says which tools it can actually reach.
Three situations where we will talk you out of it.
| If this is you | What we will say instead |
|---|---|
| The task is deterministic | If the rules can be written down, write them down. A cron job and forty lines of code is cheaper, faster and cannot surprise you, and we will tell you when that is the answer. |
| Nobody will approve anything | The irreversible lane needs a person who will actually look. If there is nobody, the honest build is one that only ever drafts — which is often plenty. |
| It needs to be right every time | An agent is a very good assistant and a poor guarantee. Where being wrong is unacceptable, the agent prepares and a person commits. |
What it costs, from.
| What | From | Covers |
|---|---|---|
| AI agent or automation | A$7,500 / US$4,900 | one workflow, end to end |
AUD ex GST for Australian clients, USD everywhere else. Model usage is passed through at cost under a ceiling you set, and hourly work is A$140 / US$95. The cost guide prices a larger build with every hour shown, and its form carries the estimate into the message.
Straight answers.
What is an agent, as opposed to an automation?
An automation runs steps you specified. An agent decides what the next step should be. That is the whole difference, and it is why an agent needs a permission list and an automation does not. If your process is the same every time, you want the automation, and it is cheaper.
What stops it running up a bill?
A budget per run and per month, checked before each call, priced from the usage the API reports. When it is reached the run stops and records why. We run the same pattern on our own tools, including a free one that refuses with a 429 and tells the visitor which limit was hit.
Can it work inside my existing systems?
If they have an API, yes, and the permission list is written per system: read-only in the accounting package, draft-only in the inbox, full access in the thing it was built to manage. Different lanes in different systems is normal and expected.
Do you have one running?
Yes — Bilby is this shape. Every night it decides what to do next on a store’s page: which element is the add-to-cart, whether a popup is in the way, whether the step worked, what is worth reporting. Its irreversible lane is empty by design: it is not allowed to buy anything, which is why it stops at the payment step.
How do I know what it did?
The audit trail, and a report you can read without us. If an agent's actions can only be explained by the people who built it, it is not finished.
How long?
Four to eight weeks for a first agent doing real work with approvals in place, and you will see a dry run well before the end — which is usually when the permission list gets its last and most useful round of edits.
Where to go from here
- AI chatbots — when the job is answering questions from content you already have
- AI development — the architecture of a system of ours that runs every night
- Bilby — an agent of ours you can watch work