Data and AI projects: what belongs in the brief?

Start with the decision you want to improve

What should users be able to do after the assignment that is difficult, unreliable or impossible today?

“We want to make better use of our data” gives a specialist little to work with. “The procurement team wants to identify orders needing review before its daily meeting” provides a user, a moment and an action. Explain the current process and its difficulties: files to reconcile, late information or inconsistent rules.

Describe which errors matter. An unnecessary alert and a missed at-risk order have different consequences. Who will check the output? When should users return to the existing process? These answers shape evaluation as well as solution design.

Example: two departments want to forecast sales but treat cancelled orders differently. The first piece of work may be to reconcile their definitions and sources. The brief then has a clear objective: establish a shared baseline before testing a forecast.

Check what the team can actually access

For each source, record its content, history, update frequency and the people who can explain its rules and authorise access.

An available file may still be unsuitable for the task. Changing identifiers, entry dates confused with event dates or discontinued categories can distort analysis. Prepare a sample you can share that includes routine cases and known exceptions. Flag what still needs investigation.

List practical dependencies: extracts to request, access to arrange, a working environment to provide and business contacts to involve. A one-off export may support exploration; a service used every day needs a maintained supply of data.

An available file may still be unsuitable for the task.

Bring in the relevant owners when personal data is involved. CNIL’s AI guidance addresses purpose, roles and data protection by design. If access rights or architecture need reviewing, include cybersecurity expertise.

Match the specialist to the work

The main task is more informative than a tool list. Roles overlap, so ask each professional which activities they have performed and which they would take on.

A data engineer works on data flows and preparation. Building those flows and testing a predictive hypothesis are separate responsibilities, even if one person has experience of both.

For a document assistant, specify whether the work involves integrating an existing model, building document retrieval or training a model. Add the languages, users, document permissions and host application. “Generative AI” alone leaves these choices open.

Still deciding on the technical priority? Our guide to tech trends and the skills your project needs places data and AI work alongside other workstreams before you complete the brief.

• Data analyst: interpret metrics, explore data and produce analysis business teams can use.
• Data engineer: build the flows, transformations and checks that supply the data.
• Data scientist: test a modelling hypothesis, compare approaches and analyse errors.
• Machine learning / MLOps engineer: prepare testing, deployment and monitoring for a model in operation.
• Generative AI engineer: integrate models and documents into an application, then evaluate responses in the intended use case.

Agree what you will review at acceptance

Replace “deliver a prototype” with a set of outputs that business and technical teams can assess together.

For data flows, ask for transformation rules, chosen checks and instructions for handling failures. For a model, agree evaluation data kept separate from development data and a comparison with the current method or a simple rule. Observed errors belong in the findings.

Moving into operation adds integration, deployment and monitoring work. Google Cloud’s MLOps documentation explains these areas. Specify what your platform already provides and what the consultant will need to prepare.

Observed errors belong in the findings.

Involve the receiving team when defining deliverables: code and versions, documentation, known limitations, access and handover instructions. If exploration shows that the data cannot support further work, its findings and recommendation may be the agreed outcome.

Use the brief to lead the conversation

Give the professionals you meet the same scenario and ask what they would check first. Their reasoning helps you compare experience, approach and how they would work with your team.

Three useful questions are: “Which unknown could change your proposal?”, “What do you need from our team?” and “What evidence will help us decide whether to continue?”. Then ask about a comparable contribution, distinguishing personal work from the team’s overall result.

For an assistant being opened to more departments, the conversation may reveal that document permissions and response evaluation need most attention, even though the original request focused on model selection.

Share these details with our Talent Agents. WheelOfWork helps you find and meet Data & AI professionals. Agree the technical scope, approvals and handover with the consultant and the people responsible for your project.