Stork Delivery
Give your business assistant a clear job
A business assistant is easier to build and evaluate when it has a defined audience, a specific task, and a clear boundary.
The problem: an assistant that promises everything has no testable job
“Ask anything about the business” is an appealing description. It is also a difficult starting requirement.
Different questions need different sources, permissions, definitions, and levels of certainty. A question about a policy document is not the same task as a question about today's unshipped orders.
Give the assistant one clear job before expanding its reach.
A practical approach
Write a task statement
A useful statement identifies the user, the question, and the expected output. For example: help account managers locate the current service terms for a customer and open the supporting agreement.
That statement tells the team which records matter and who should judge the result. It also limits what the assistant is expected to do.
Compare it with “improve productivity.” That may be the business objective, but it does not give the engineering team a task it can test.
Choose the right information path
Document questions may benefit from retrieval with citations. Questions about current structured records may need a controlled lookup. Calculations may belong in ordinary application code with a model explaining the verified result.
Avoid forcing every task through a collection of document excerpts. A current order count should come from a defined query with appropriate access controls, not from the model's general knowledge or an old export it happened to retrieve.
Show the relevant scope and freshness alongside the answer when those facts affect its interpretation.
Make uncertainty useful
Define what happens when a customer name is ambiguous, a source is unavailable, or the question crosses the assistant's boundary.
A clarification can be a successful outcome. So can directing the user to the person who owns the decision. The application should not fill a gap in company information with an unsupported guess.
Let users inspect the records behind the answer and report a problem without reconstructing the whole interaction for support.
Expand from evidence
Observe which questions succeed, which require manual correction, and which users keep asking despite being outside the initial scope. Use that evidence to choose the next capability.
Recheck permissions, evaluation, and operating cost when expanding. A new data source or action can change the application's responsibilities substantially.
The practical promise of AI connected to business data is that it helps someone complete a useful task with the right context. A clear first job makes that promise easier to deliver and gives the team a sound basis for deciding what comes next.
The result: one useful question answered with the right evidence
Define a first task for a named audience, then choose examples covering a valid answer, an ambiguous customer, missing information, and denied access. Have the business owner review the answer and the records supporting it. The initial capability is ready to expand when users can complete that task and recognize its limits. Keep broader requests as evidence for prioritization, rather than silently widening authority.