01 the business
Wholesale distributor
150 staff, one data analyst
Managers have data questions every week, but only one analyst can get answers out of the reports. Requests queue up, private spreadsheets multiply, and two managers quote two different revenue figures in the same meeting.
02 before
The problem
Most managers have a question for the data every week, but only one or two people know how to get the answer out of the reports. Requests queue up, private spreadsheets multiply, and two people quote two different revenue figures in the same meeting.
- Questions wait days for the analyst.
- Metric definitions differ between teams.
- Spreadsheet copies drift from the source.
03 the build
What the AI does
- Step 1: Connects to your reporting sources, such as your accounting system, CRM, data warehouse or Power BI datasets, with read-only access.
- Step 2: Works from agreed definitions for terms like revenue, active customer and gross margin, written down once with your finance lead.
- Step 3: Turns a plain-English question into a query, runs it and replies with the figure, a short explanation and a simple chart.
- Step 4: Shows the filters and date range it used, so anyone can see how the answer was built.
- Step 5: Says when it cannot answer reliably and passes the question to your data owner instead of guessing.
- Step 6: Follows existing permissions, so people only see figures their role already allows.
paused: waiting for a person to approve
How it was put together
- Connected reporting sources with read-only access.
- Wrote down definitions for revenue, active customer and gross margin with the finance lead.
- Turned plain-English questions into queries and replied with the figure, an explanation and a chart.
- Showed filters and date ranges used, and passed unreliable questions to the data owner.
04 human in the loop
Where a person stays in charge
Your finance or data lead approves the metric definitions before launch and reviews flagged questions each week. Any figure headed for the board, a lender or a client is checked by that person before it goes out.
- Finance approves metric definitions before launch.
- Existing permissions apply to every answer.
- Figures for the board, lenders or clients are checked by the data owner.
05 rollout
From kickoff to live in 4 weeks
- Weeks 1–2Agree definitions and connect sources.
- Week 3Test with the analyst against fifty real questions.
- Week 4Open to managers in Teams.
06 after
What changes
Managers get answers without waiting in the reporting queue, and the whole business works from one set of definitions. Your analyst spends time on the harder questions instead of re-running the same report.
Assume 25 managers each asking two data questions a week, at 20 minutes of analyst time each.
25 × 2 × 20 minutes = 1,000 minutes, or about 17 hours a week of analyst time on ad-hoc requests.
What to measure
- Questions answered without the analyst.
- Answers matching the analyst's check.
- Analyst time on ad-hoc requests.
- Conflicting figures quoted in meetings.
07 evidence
What the research says
Gartner research put the average cost of poor data quality at $12.9 million a year per organisation.
Gartner, Data quality: why it matters and how to achieve it (opens in a new tab) · 2020
In a field experiment with 758 BCG consultants, those using GPT-4 on tasks within its abilities finished 12.2% more tasks, 25.1% faster and at over 40% higher quality, but did worse on a task outside its abilities.
A survey of 4,050 sales professionals found the average seller spends about 40% of their time actually selling, and 51% of sales leaders using AI say disconnected systems are slowing their AI work.
Salesforce, State of Sales report, 2026 edition (opens in a new tab) · 2026
08 lessons
Mistakes to avoid
- Skipping definitions. Without them, you automate the confusion.
- Hiding how answers were built.
- Letting it guess. 'I cannot answer that reliably' is a good answer.
09 questions
Common questions
- Will it get numbers wrong?
- It can, which is why definitions are fixed, workings are shown and external figures are checked by a person.
- Which data sources work?
- Power BI datasets, BigQuery, Snowflake, Xero, HubSpot and others with read-only access.