How is it different from a classic specification?

You describe screens less and behaviour more. Examples of real cases with the expected answer replace a large part of the specification, and the agent’s limits are written in black and white.

Should you specify which AI model to use?

No, unless there is a real constraint such as data location. Describe the expected result and let the provider propose the model, asking that it can be replaced.

How long should it be?

One page, plus the examples in an appendix. A longer document is not read, and ages from the first trials.

GuidesProject brief

AI project brief: the one‑page template

By the Banzai Labs teamUpdated on November 11, 20262 min read

Why an AI brief is different

In a classic project, you specify what the software does at each click. With an agent, you cannot list every case: handling the unexpected is exactly its job.

So you specify something else: the task, the material, the limits, and examples.

The template in seven sections

1The task
2Who uses it, and when
3The data
4The examples
5The limits
6The success measure
7The constraints

1. The task

One sentence, as you would explain it to a new colleague. “Every Monday, produce a sales report for the week for each store, with anomalies.”

2. Who uses it, and when

The person, the moment, the tool where the result lands: an email, a chat tool, an existing screen.

3. The data

Where it is, and how to access it.
Who is responsible for it.
What is sensitive and must not leave.

It is the section providers read first. It accounts for most of the budget.

4. The examples

Ten to twenty real cases, each with what is expected. Include easy cases, hard cases and cases where the right answer is “I don’t know”.

These examples will serve as the test set throughout the project.

5. The limits

What the agent may do alone, what it must propose, what it must never do.

Level	Example
Alone	Read orders, write the report
With approval	Change a price, send an email to a customer
Never	Refund, delete data

6. The success measure

A number, and how to record it: time saved per week, share of cases handled without correction, response time.

7. The constraints

Build budget, monthly running budget, date, hosting, language, compliance.

What not to put in it

The name of a model or a tool, unless there is a real constraint.
A list of features “just in case”.
Vague requirements: “smart”, “reliable”, “fast”. Replace them with an example or a number.

The questions a good provider will ask you

What happens today when the task is done badly?
Who will review what the agent produces?
How many times a day does the task come up?
Who will be responsible for the agent once it is in service?

If nobody asks them, it is a bad sign.

A filled-in example

Task. Change prices, stock, product pages and translations across several Shopify stores from a conversation.

User. The merchant, in their usual AI assistant.

Data. The stores, through their API. No customer data is needed.

Limits. Free reading. Every change is displayed before being applied.

Success. A change across five stores takes the time of one.

That is the brief for our project Comptoir, summarised.

And then

Send the page to two or three providers and compare their questions as much as their prices. See AI agency, freelancer or in-house team and How much an AI project costs.

Frequently asked questions

How is it different from a classic specification?: You describe screens less and behaviour more. Examples of real cases with the expected answer replace a large part of the specification, and the agent’s limits are written in black and white.
Should you specify which AI model to use?: No, unless there is a real constraint such as data location. Describe the expected result and let the provider propose the model, asking that it can be replaced.
How long should it be?: One page, plus the examples in an appendix. A longer document is not read, and ages from the first trials.