Notes

How to Write an AI Project Brief a Builder Can Actually Quote From

What should an AI project brief include?

A good AI project brief covers five things on a single page: the job to be done, the inputs it works from, what done looks like, the exceptions, and who approves the output. Think of it as the business problem, stated so a stranger could check it. No technology choices, no model names, no feature wish list. If you can write those five sections in plain English, any competent builder can judge feasibility, tell you whether the work is worth doing, how long it will take and what it will involve.

Most briefs fail in the other direction. They describe a department, a vision or a tool, and leave the builder to guess the job. Guessing is where scope creep is born.

Why does the brief matter so much?

Because the brief sets the scope, and the scope sets everything else: the timeline, the fixed price, and what you will be judged against at the end.

A vague brief does not save you time. It moves the hard thinking to the middle of the build, where it is slower and more awkward, and where every clarification becomes a change. A clear brief also protects you. When the builder comes back with a proposal, you can hold it up against the page and see whether they answered your question or a more convenient one.

You do not need a technical background to write a brief. You need to know the work, which you do better than anyone you will hire.

How do you describe the job?

Describe one task, in one sentence of plain language, without using the word AI anywhere in it.

"Read every supplier invoice that arrives in the accounts inbox, check it against the purchase order, and prepare it for approval in Xero." That is a job. "Use AI to improve our finance function" is not, because nobody could tell you on any given Friday whether it had been done.

Pick a task, not a department. Broad briefs get broad, slow, unfinished answers. If you have several candidates, write a sentence for each and choose the one that is frequent, repetitive and painful, a choice we explore in what to automate with AI first. Then add two or three lines on why it matters: who does it now, roughly how often, and what goes wrong when it slips.

What do you need to say about inputs and data?

List where the work starts and where the information lives, as plainly as you would explain it to a new hire on their first morning.

Useful questions to answer in the brief:

Be honest about the mess. If half the job lives in someone's head or a spreadsheet only one person understands, say so. That is not a reason to stop, it is the single most useful thing a builder can know at the start. Preparing your data for AI covers what good enough looks like.

How do you define what done looks like?

Write one measurable sentence that would be plainly true or false thirty days after go-live. These are your success criteria, and one is usually enough.

"Every invoice that matches its purchase order is ready for approval the same day it arrives, and every mismatch is flagged to a person with the reason." Someone can check that. "Faster, more accurate invoice processing" cannot be checked, so it can never fail, so it can never succeed either.

Alongside the done sentence, note today's numbers if you have them, even rough ones: how long the job takes, how often it goes wrong. That becomes the baseline, and it is the thing most projects wish they had captured later. We go further into measures in how to measure ROI on AI.

Why do exceptions belong in the brief?

Because the exceptions are where the real work is, and they are invisible to anyone who has not done the job.

The normal case is easy to describe and usually easy to build. The awkward cases decide whether the finished thing is trusted. Write down the ones you know: the supplier who always sends two invoices in one PDF, the customer who emails from three different addresses, the job that needs a site visit before anything else happens. Then say what should happen to anything that does not fit, which is nearly always the same answer: stop and send it to a named person.

A brief with a short list of exceptions is worth more than a long list of features.

Who should approve the output?

Name a person, not a role, and say which outputs they approve before anything leaves the building. Other stakeholders can be consulted, but one person decides.

Approval gates are a design decision, and they belong at the start. Decide which actions the system may take on its own, such as filing, tagging or drafting, and which need a person to say yes, such as sending to a customer, paying a supplier or changing a record. You can relax gates later once the work has earned trust. You cannot easily add them after something has gone out wrong. The same person should be the one who signs off at the end, so the build is judged by someone who knows the job.

What should you leave out?

Leave out the solution. You are briefing the problem, and a good builder will propose the approach.

Skip the model names, the tools you read about and the feature list from a competitor's website. Do list the deliverables you expect to keep at handover, such as the documentation and access to your own accounts, and note any integration that has to work on day one. Do include real constraints: systems that cannot change, data that must stay in Australia, privacy obligations, busy periods when nothing should go live. The OAIC is the place to check privacy obligations in general terms, and anything specific to your business is worth confirming with your own adviser. For a broader view on oversight and accountability, the Australian Government's AI Ethics Principles are short and readable.

Bring the page, we will bring the questions

A one-page brief with a clear job, honest inputs, a done sentence, the known exceptions and a named approver is enough to start a real conversation. It does not need to be polished or follow a template. It needs to be true. The builder's proposal should then turn it into milestones, usually starting with a proof of concept on your own data.

If you have a draft, or just the first sentence, send it to us. We will read it, ask the questions a builder should ask, and come back with a fixed scope in writing. You can see how that conversation runs on the method page.

Keep reading