A useful first AI project has a clear job to do, a person responsible for it and a practical way to tell whether it helped.
That might be helping staff locate approved product information, preparing a draft for review or extracting fields from a familiar document. The scope should be narrow enough to test against real examples, including the difficult ones.
Start with the work, not the label
Some tasks suit predictable rules. If a form is complete, assign it to a team and send an acknowledgement. AI may add little to that process. Other tasks involve varied language or unstructured information, where AI may help, provided its output can be checked.
Our recommendation is to choose a repeated task with a visible cost in time or effort. Avoid beginning with a broad instruction to “put AI everywhere”. It makes success harder to define and creates too many moving parts at once.
Prepare the knowledge it will depend on
A knowledge assistant needs useful source material. Check which documents are current, who owns them, what is missing and who should be allowed to access them. Contradictory policies and old product information do not become reliable because an AI system can read them.
- Choose approved sources and assign an owner.
- Define access permissions for different users.
- Agree how updates and withdrawn documents are handled.
- Decide when the system should say it does not know.
For an illustrative product-information assistant, success might mean finding the relevant specification and linking to the approved source. It should not invent an answer when the required detail is absent.
Design for uncertainty from the beginning
NIST’s Generative AI Profile addresses risks including confabulation, data privacy and human interaction with AI, and provides actions for managing those risks.[1] We translate that into a practical project principle: test the failure cases alongside the intended use.
Include ambiguous questions, missing information, outdated documents and requests beyond the user’s permissions. Check whether the assistant can resist instructions embedded in retrieved material that conflict with its intended task. Define how staff report problems and who decides whether the system needs to change.
Measure the whole task
Fast output can still be expensive if someone has to rewrite or verify most of it. Compare the existing workflow with the assisted workflow, including review and correction.
| Measure | What to check |
|---|---|
| Time | Time to complete the job, including review. |
| Quality | Accuracy and usefulness against agreed examples. |
| Control | Permission handling, escalation and visibility of errors. |
| Adoption | Whether the intended team actually uses it. |
| Cost | Usage, support, maintenance and human effort. |
A pilot should have a review point and a decision: expand, improve, pause or stop. It is useful to learn that an approach is not suitable before it becomes a major commitment.
Build on what works
We can help identify the first use case, prepare the information, connect suitable AI technology and evaluate the result. We focus on practical business solutions rather than developing foundational AI models.
You bring the knowledge of your business and the people responsible for decisions. We bring process understanding, development and integration capability. Together, we can build a bounded solution, learn from everyday use and decide what should come next.
Research & further reading
1. NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1, 2024). The assistant example is illustrative, not a client case study.