Rolling out an AI coworker without losing your users’ trust
What I took away from Databricks’ Genie One rollout playbook, and why I think the sequencing matters more than the model.
I’ve watched a few AI assistant launches go the same way. The demo lands well, leadership gets excited, access goes out to half the company, and within a month usage drops off. When you ask people why, the answer is rarely about the tool being slow or hard to use. Someone got a wrong number in front of their manager, told their team, and that was that.
So I was glad to see Databricks publish a step-by-step playbook for rolling out Genie One that treats trust as the thing you’re actually building. The example they open with will be familiar to anyone in a data team: a sales director needs to understand a soft regional pipeline before a meeting, messages the one analyst who knows where the good data lives, and gets her answer two days after she needed it.
Genie One is meant to close that gap. You ask in plain language and get an answer grounded in governed data, with sources cited. The playbook is about how you get from that promise to finance, ops and sales using it every day.
Why these rollouts stall
The failure pattern they describe is doing too much at once. A dozen data domains in the first week. Four teams with four definitions of “active customer”. Nobody owning the semantic layer, and no way of checking whether answers are right. Users lose confidence fast after one bad answer, and they don’t come back easily.
Their answer is sequencing. Start with a narrow set of questions and a few early adopters, fix what they find, and only then widen the scope.
The four building blocks
Before the phases, it helps to know what you’re actually deploying. There are four pieces, and their weight shifts as adoption grows.
Phase 0: get the foundations right
This is the part most teams want to skip, and the part I’d spend the most time on. Pick one team with a recurring, painful question set. Their example is sales ops and pipeline health. Then build the semantic layer in order:
- Domains first, to group assets by business purpose. Pages have to sit inside a domain, so this unblocks everything else.
- Metric views for the 10 to 20 numbers people argue about: bookings, coverage, win rate. Define once, slice any way.
- Pages to explain the concepts behind those numbers, like what counts as a “qualified opportunity”. Genie cites them in its answers.
Give every asset a real owner with a route to fix things, set up permissions and audit logs while you have five users rather than five hundred, and write down 25 to 50 real questions with known-correct answers.
Get SHIKHER’s stories in your inbox
Join Medium for free to get updates from this writer.
That last step is the one I’d underline. Score Genie against those questions before the pilot and after every semantic change. The point they make, which I agree with, is that most wrong answers aren’t model failures. They’re gaps in your definitions, and each miss tells you what to govern next.
Phase 1: a small, honest pilot
Five to ten users who will tell you when it’s wrong. Review results weekly, log every incorrect answer and every “I don’t know”, and fix the root cause. Agree the exit criteria before you start: accuracy you’d base a real decision on, people choosing Genie over their old workflow without prompting, and a fix backlog that’s shrinking.
My favourite line in the piece is about the signal that the pilot worked: someone answers a colleague by forwarding a Genie response instead of raising a ticket.
Phase 2 and 3: widen on purpose
Phase 2 adds two or three neighbouring domains, each with its own owner and evaluation set. Shared terms like “revenue” move into the ontology; team-specific ones stay local. A light review gate checks permissions and sensitive columns before a new team goes live, and each Page gets a review cadence so definitions don’t quietly drift. A Slack channel and office hours do a lot of the adoption work.
Phase 3 is account-level access for everyone, synced from your identity provider. A central platform team owns standards and the ontology while domain teams own their agents. Unity Gateway extends governance to the tools Genie can call, and Genie shows up where people already work: Slack, Teams, Excel, Google Sheets and mobile.
My takeaway
The timeline is a guide, not a deadline. Move on when the evidence says you’re ready, and take the extra weeks if definitions still need work.
The organisations that get this right don’t necessarily have the cleanest data or the biggest platform team. They start with one team, one governed domain and a list of questions they’re willing to be graded on. Once that works, they’ve earned the right to expand.
If you’re planning a rollout, the full playbook and the Genie documentation are worth reading before you send that company-wide access email.
