Short answer
An AI project brief is a one- to two-page document telling a builder, vendor or internal team what an AI system must do: the workflow, inputs, outputs, systems, human review points, success measure, privacy constraints and what is out of scope. This template has 14 numbered fields with guidance for each, and a completed example you can copy.
- Describe the job and the result, not the technology.
- Attach 10 to 20 real examples of inputs and good outputs.
- Say exactly where a person reviews output before it takes effect.
- List what is out of scope; it heads off scope disputes later.
Use this brief when you are about to ask a vendor, a development partner or your own team to build or configure an AI system. It sits between the AI business case (why it is worth doing) and the build itself. If you have not mapped the workflow yet, do that first with the AI workflow mapping template.
The AI project brief template
Copy the table into a document and replace the guidance column with your answers.
| # | Field | What to write |
|---|---|---|
| 1 | Project name, owner, date | One name, one accountable owner, the version date |
| 2 | Problem | One paragraph: what goes wrong today and what it costs, with numbers |
| 3 | Users and affected people | Who uses the system, and who is affected by its outputs (customers, staff, suppliers) |
| 4 | Current workflow and volume | Link to the workflow map; tasks per month, peaks, and how volume varies |
| 5 | What the AI should do | Verbs: read, extract, classify, draft, check, route, summarise. One line each |
| 6 | What the AI must not do | Actions it may never take, such as sending to customers, approving payments or deleting records |
| 7 | Inputs | Types and formats (email, PDF, scanned form, spreadsheet), where they arrive, and 10 to 20 attached examples |
| 8 | Outputs | Format, destination system, and an example of a good output for each attached input |
| 9 | Systems and access | Each system it reads from or writes to, the account it uses, and the permission level |
| 10 | Human review | Who checks what, before what happens, and how fast they must respond |
| 11 | Success measure and baseline | The number to move, its value today, the target and the date |
| 12 | Data and privacy | Personal information involved, where it may be processed, retention, and who approved it |
| 13 | Constraints | Budget range, deadline, tools already licensed, response time, hours of operation |
| 14 | Out of scope and open questions | What is not included in this phase, and what you do not know yet |
How to fill it in
- Start with fields 2, 5 and 11. If you cannot state the problem, the job and the measure in a sentence each, the project is not ready to brief.
- Collect the examples before writing field 7. Pull 10 to 20 real inputs, including the messy ones. Remove personal information unless the builder is already approved to handle it.
- Write a good output for each example. This is the single most useful thing you can give a builder; it doubles as the first test set.
- Be specific about review. “Staff will check” is not enough. Name the role, the moment and the turnaround. Our guide to planning human review for AI output helps.
- Fill field 12 with your privacy officer. Under the Privacy Act 2020, you stay responsible for personal information a provider processes on your behalf.
- Have the workflow owner sign it. Then send the same brief to every vendor so quotes are comparable.
Completed example
Illustrative example: a 60-person commercial builder in Christchurch wants help pulling key requirements out of tender documents. See our workflow page on tender document extraction for the general pattern.
| # | Field | Example answer |
|---|---|---|
| 1 | Project, owner | Tender requirements extraction; Estimating Manager; v1.0, 20 October |
| 2 | Problem | Estimators spend about 6 hours per tender finding scope, dates, insurance, retentions and special conditions across 200 to 600 pages. Two clauses were missed last year |
| 3 | Users, affected | Four estimators use it; clients and subcontractors are affected by the bid |
| 4 | Workflow, volume | 12 to 18 tenders a month, peaking in February and August |
| 5 | Should do | Read tender PDFs; extract 22 listed fields with page references; flag unusual clauses; draft a one-page summary |
| 6 | Must not do | Price anything; contact clients; change files in the document system |
| 7 | Inputs | PDF and Word tender packs from email and the client portal; 15 past tenders attached |
| 8 | Outputs | Summary in the bid folder plus a spreadsheet of the 22 fields, each with page numbers; examples attached |
| 9 | Systems | Read-only access to the tenders folder in SharePoint; write access to one output folder |
| 10 | Review | Lead estimator checks every field against the cited page before the bid meeting |
| 11 | Measure | Time to first summary: 6 hours today, target under 90 minutes including review, by end of March |
| 12 | Data, privacy | Little personal information; client confidentiality applies; provider must not train on documents |
| 13 | Constraints | Already licensed for Microsoft 365; summary needed within 2 hours of receipt |
| 14 | Out of scope | Quantity take-off from drawings; open question: how to handle addenda issued late |
Common mistakes
- Briefing a solution. “Build us a chatbot” hides the job. Describe the work and the result, and let builders propose how.
- No examples, or only easy ones. Builders quote on what they see. Hidden exceptions become variations later.
- Vague success. “Improve efficiency” cannot be tested. Use a number with a baseline and a date.
- Missing the “must not” list. Field 6 is where you set the limits that keep an AI system safe, especially for AI agents that can take actions.
- Leaving out who reviews. MBIE’s quickstart for businesses says a human in the loop is needed where important decisions or actions are made or informed by AI (MBIE). The brief is where that becomes a job, not a hope.
Next step
Agree how the finished system will be tested before you sign: the AI agent acceptance checklist turns fields 5, 6, 10 and 11 into pass or fail tests. If you want a partner to scope and build it, our custom AI development work starts from a brief like this one, and writing a workflow specification goes deeper on field 4.
Sources
- Privacy Act 2020, New Zealand Legislation, version as at 1 October 2026
- Quickstart: Businesses using AI tools, Ministry of Business, Innovation and Employment, accessed 19 October 2026
Written by the Levy Street team in Wellington. Last updated 19 October 2026.