“We need AI for the business” is too broad a development brief. Describe a situation instead: who receives the request, what they look up, where they transfer information and what counts as completion. That description can become a custom workflow whose usefulness is testable.
Bring an example of everyday work
Prepare several anonymised inputs, an example of a good output and frequent exceptions. Explain current systems and access restrictions. A case involving repeated cross-checking or questions to colleagues is especially useful.
You do not need to select a model or station vendor first. Establish functions and quality criteria, then validate an appropriate hardware class. A standalone, memory-rich AI mini PC is one hosting option rather than the project's objective.
Example: a bespoke purchasing approval
A company uses a particular approval sequence. The assistant prepares a comparison, sends it to the project lead for review and, after approval, passes it to purchasing. Define states, owners and transition evidence. The model helps read and draft; business rules control the route.
If an approver is absent, the process must not invent a replacement. Document delegation rules separately. Exceptions like these often determine project scope more than the number of screens.
Record the pilot agreement
- Inputs and authorised sources. - Outputs, mandatory fields and acceptable error boundaries. - Reading operations, drafts and actions requiring approval. - Evaluation examples, workload and a manual fallback. - Process ownership and support after launch.
Expand from accepted results
After accepting a bounded workflow, add adjacent roles for email, documents, assignments or metrics. Each expansion changes data access or authority and needs its own evaluation. AI Office is ready to work through that process with a customer: understand operations, propose a scope and develop a solution for the existing infrastructure. Start with one clear assignment the team genuinely performs regularly.