The question at this stage Which channels bring customers who stay, and does every team trust the numbers that say so?
Record now, or lose it: every change to a customer, deal or subscription, with ad spend joined to revenue by channel. Without it, you cannot tell which channel's customers stayed.
The five steps at this stage
- Step 1Choose the problem
Frame channel questions as groups of customers: who came from each channel, who stayed, and what it cost. The data person challenges each question before anyone builds the answer, and checks each goal for how someone could game it.
If you skip it: The team optimizes for cheap sign-ups instead of customers who stay, and people hit targets by gaming the number.
- Step 2Collect what you need
Record every change to a customer, deal or subscription, and join ad spend to revenue by channel. Keep a written, tested list of what you track, and keep accounts separate from the people who use them. A team that changes what its system records announces the change and tests it before release.
If you skip it: A renamed field breaks reports without an error.
- Step 3Choose the number
Define each number once, have every tool read that definition, and review every change to it. Give each number one owner. This shared layer of definitions waited until now on purpose.
If you skip it: Each tool drifts to its own definition and the dashboards disagree.
- Step 4Count it right
Move data in steps from raw to cleaned to business-ready. Reprocess any past period safely, including late data. Run automatic checks with owners, test changes away from the live setup, and put alerts on data spend.
If you skip it: A fix to one number quietly changes history somewhere else.
- Step 5Read it right
Compare groups of customers over time, from the shared definitions. The agent drafts the analysis and a person reviews it before it travels. Write a short note whenever a number turns out wrong.
If you skip it: A confident agent-written story reaches the board unchecked, and the same mistake returns when the person who knew leaves.
What can wait
- Heavier tools can wait until something specific breaks.
- Safe by default
Give the agent written rules and keep a list of what data exists. The agent suggests changes, and a person approves anything touching definitions, access or personal details. Review who can see what on a schedule.
- Who runs it
Your first data person, full-time or half-time, runs it while an AI agent builds and maintains the setup.
We think one person can now cover what used to take three.