Where should your first AI agent project start?

Choose a bounded task

Look for a repeatable information task: finding an answer in approved documents, extracting fields for review, preparing a draft, or routing a request. Give the first version a clear input and expected output.

If the process includes payments, changes to important records or other consequential actions, identify where a person must review and approve the work.

Understand the information it can use

An assistant needs authorised access to the right information. Confirm which documents and tools are relevant, who can access them and how updates will reach the system.

An integration is also a permission decision. Define what the assistant can read, propose and change. Keep those abilities matched to its task.

• Approved data sources and owners
• User roles and tool permissions
• Information freshness
• Sensitive data and excluded records
• Fallback and escalation routes

Test examples from your own workflow

Build an evaluation set from representative cases: ordinary requests, incomplete inputs, ambiguous questions and examples that should be refused or escalated. Ask the team to show you the outputs and failure behaviour.

Measure usefulness against the process you intend to improve. A demonstration with carefully selected examples does not establish reliable behaviour across all your work.

Separate build cost from operating cost

Model usage, document processing, hosting, integrations and support can create ongoing costs. A proposal should explain the intended workload and which charges vary with usage.

We recommend beginning with a scoped workflow, reviewing its performance and cost, then deciding whether to expand. Bring us a process and representative, authorised examples to start the conversation.