An AI assistant that can search across everything an employee has access to is genuinely useful. It is also an unusually effective way to discover that your permissions were never as tidy as you assumed.
This is the central thing to understand before adopting one: the assistant does not create new access. It surfaces what a person could already have found, far more efficiently than they ever would have by searching manually.
Readiness is mostly about information, not AI
Know what is over-shared
Review broad sharing before deployment: documents shared with everyone, links available organization-wide, legacy sites with permissions nobody has examined in years.
Label what is sensitive
Sensitivity labelling is what lets you treat confidential information differently. Without it, everything is equally available to summarise.
Decide what to do with stale content
Old drafts and superseded policies are a problem once an assistant can cite them with equal confidence to the current version.
Expectations worth setting
Adoption disappoints most often because of unmanaged expectations rather than capability. Two things are worth saying plainly to everyone:
- It produces drafts, not finished work. The person remains responsible for what they send.
- It can be confidently wrong. Verification is not optional, particularly for figures and citations.
Starting narrow
Broad rollouts tend to produce a burst of experimentation followed by quiet abandonment. Starting with a specific group and a specific recurring task gives you something to learn from.
- Pick one team with a well-defined, repetitive writing or summarising task.
- Fix the information governance for the areas that team touches.
- License that group and work alongside them for several weeks.
- Collect what worked and what did not, in their words.
- Use that to decide whether and how to widen access.
The cost question
Per-seat AI licensing is a significant recurring commitment. It is worth being able to name the specific work it replaces or improves for each person licensed — a general sense that it is useful tends not to survive a renewal conversation.