Most internal AI projects begin with a broad promise: give everyone an assistant for everything.
That promise is hard to evaluate and even harder to trust. A general assistant may sound impressive in a demo, then leave people unsure what it knows, where its answers came from, or when they should use it instead of asking a colleague.
The agents that get used tend to begin with a much narrower job.
Start with a recurring question that already creates work. A support leader may need a weekly readout on ticket volume and recurring themes. A sales manager may want to know which late-stage deals have gone quiet. An operations lead may need a morning view of exceptions that need attention.
The narrower starting point is not a limitation. It gives the agent a purpose people can recognize.
Pick a job with a clear owner
Choose a question that has three qualities.
It should happen often enough that people remember the pain. It should draw on sources your team already trusts. And someone should care whether the answer is useful.
“Help us understand the business” is too broad. “Give the support manager a weekly view of new ticket themes and the accounts most affected” is a real job. It has a user, a cadence, and a way to tell if the output helped.
Connect only the context the agent needs
More data is not always better. An agent with access to every source can create confusion when the same term means different things in different systems.
Give it the sources needed for its job, then describe the important context. Which table holds the metric? What does an active customer mean? Which fields should be treated carefully? Who is allowed to see the result?
Jovis organizes approved connections within a workspace, and agents use the sources available to that workspace. This makes it easier to give an agent a defined role instead of turning it into an opaque search box for the whole company.
Design the first question
Do not leave the first interaction to chance. Put a useful question in front of people.
For a revenue agent, that might be: “Which deals expected to close this month have had no customer activity in the last 14 days?” For a product agent: “What issues were opened after the latest release, and are they concentrated in one area?”
Good starter questions teach people what the agent can do. They also reveal gaps quickly. If the answer is vague, the source may be missing context. If the answer is correct but not useful, the question may not match the real decision.
Treat the first month as research
Watch what people ask after the first answer. Those follow-ups are often more valuable than the original prompt. They show the language your team uses and the decisions they are trying to make.
Keep a short list of questions the agent could not answer. Some will need a better source. Some need a clearer definition. Some should remain a human conversation. That is fine. Trust grows when an agent is clear about its role.
An internal AI agent succeeds when it becomes part of an existing habit. Build for one real habit first. The broader use cases will be easier to see once people have something reliable to return to.