Before the model, get the foundations right.
Why permissions, data quality, and evaluation determine whether enterprise AI is useful beyond a prototype.
A useful pilot needs a useful question
A fluent answer can make an AI prototype look more mature than it is. The harder questions arrive when the system touches real business information, serves different users, or becomes part of an operational workflow.
Before selecting a model, define a task narrowly enough to evaluate. “Help people find the current policy and identify the relevant passage” is a stronger starting point than “build an enterprise assistant.” The first statement gives you users, source material, and a result you can inspect.
Treat access as part of retrieval
Search results are not automatically safe to put into a prompt. A document index may contain information from systems with different permissions, retention rules, and ownership. The AI application needs to respect those boundaries at the point of retrieval.
Map identity, source permissions, and document ownership before building the happy path. Test what happens when a user changes role, a document becomes restricted, or a source is removed. A retrieval result that is relevant but unauthorized is still an incorrect result.
Make evaluation representative
Build a small set of realistic tasks with people who understand the work. Include ambiguous questions, missing information, conflicting sources, and requests that the system should decline or escalate.
Compare the result against the current process. Look at factual support, completeness, time saved, and the effort required to review an answer. Model quality matters, but the unit of value is the completed task.
Design the uncertain path
A useful workflow makes uncertainty visible. It can show supporting sources, separate a draft from an approved record, and ask for review before a consequential action.
Define who is responsible when the system cannot produce a supported answer. Decide how an issue is reported, how outputs are reviewed, and how feedback reaches the people maintaining the application. These paths deserve as much design attention as a successful response.
Earn the next increment
Start with a bounded workflow and a clear baseline. Expand only when the evidence supports it. A stronger data foundation or a better integration may create more value than moving to a larger model.
The goal is a dependable capability inside the business. Model choice is one of the decisions that supports it.
A QUESTION TO TAKE BACK
What needs to be true for this AI workflow to be useful, safe to operate, and worth maintaining?