The first ninety days, in named roles and hours
6 min readUpdated
A realistic rollout has three phases: connecting systems and confirming that people see only what they should, agreeing the small number of definitions that decisions actually rest on, and widening access to a second team. The work is mostly decisions rather than engineering, and the common failure is not technical: it is that nobody was made accountable for the definitions, so two teams keep quoting different numbers.
Weeks one and two: connect and check the boundaries
- Who is neededSomebody with administrator access to each system you want to connect, for about an hour each. No engineering build.
- What to verify before anything elseLog in as somebody junior and ask a question about something they should not see. Getting this wrong once, publicly, ends a rollout, so it is worth doing on day one rather than day sixty.
- What good looks likePeople are asking real questions of real data by the end of week two. If that has not happened, something is wrong with the plan rather than the pace.
Weeks three to six: agree the definitions that matter
This is the phase people skip, and skipping it is the single best predictor of a tool that gets abandoned. The work is not technical. It is getting a named person to decide what revenue means for reporting purposes, and writing that decision down where the tool reads it.
Do not try to do all of them. Find the ten or twenty definitions that decisions actually rest on, settle those, and leave the rest to be resolved when somebody asks.
- Who is neededOne accountable owner per domain, for a few hours in total, plus an analyst to write the definitions down.
- The trapTrying to model everything up front. It is a quarter of work, most of it for questions nobody asks.
Weeks seven to twelve: widen to a second team
A single-team pilot proves very little, because that team is motivated and close to whoever ran the trial. Extending to a second team with no personal stake is what tells you whether the thing is genuinely usable.
By this point you should be able to answer one question honestly: has anybody stopped asking somebody else for a number they now get themselves? If not, nothing has changed regardless of the usage figures.
The three ways this goes wrong
- Nobody owns the definitionsTwo teams keep quoting different numbers, both blame the tool, and trust never accumulates.
- It was rolled out to everyone at onceAccess without any agreed definitions produces confident, contradictory answers at scale, which is worse than the queue it replaced.
- It was measured on loginsUsage figures rise during any rollout. The only measure that matters is whether requests to the data team went down.