Back to resources

AI & Reporting · 10 min

AI Business Reporting: Ask Better Questions of Your Data

, Founder, Evolve StrategistsPublished Updated

The useful question isn't whether AI can write a report. It's whether your team can ask a business question and get an answer they can check. That's the standard we work to in AI consulting, when we connect AI to a business's sales, churn and marketing data.

That depends less on the AI model than on what sits behind it: the right data connected, agreed definitions for the numbers, controlled access and the business context behind your decisions. This guide covers each of those, with a worked campaign comparison you can check yourself.

What Can You Ask AI About Your Business?

With the right setup, people can ask questions like these in plain English:

How many net sales did we make last week?
Which campaigns produced paying customers, not just leads?
Which group of members tends to cancel earliest?
Where do handovers from sales to delivery slow down?

This is often called AI business intelligence: plain-language questions over connected, well-governed business data. The system can retrieve records, run calculations and summarise the results. How useful the answers are depends on which data is connected and how clearly the numbers are defined. Answering the campaign question, for example, needs sales outcomes linked to advertising data, which our guide to Meta offline conversions covers.

Agree What the Numbers Mean First

Most reporting arguments are about definitions, not arithmetic. Before connecting anything, agree what each important number means:

Questions that need a definition first
QuestionDecide explicitly
How many sales did we make?Is a sale a signed contract, a first payment, or a payment that wasn't later cancelled?
What's our churn rate?Which customers count in the denominator, and over what period?
Which campaign is most profitable?Revenue, or profit after advertising and other specified costs?
Which rep keeps customers longest?Are you comparing like groups, with similar lead sources and start dates?

Write each core metric down as a short metric contract. Here's an illustrative format:

Illustrative metric contract

Metric: Net sales

Formula: Paid sales in the period, minus sales cancelled within the agreed cancellation period

Source system: Payment platform, matched to CRM deals

Time window: Calendar week, Monday to Sunday, Perth time

Refresh: Daily, early morning

Exclusions: Test transactions and internal accounts

Owner: The finance manager

Connecting a pile of spreadsheets to an AI tool doesn't create a single source of truth. Agreed definitions and checked data do.

Build a Controlled Path to the Data

A dependable setup has distinct layers:

From source systems to an answer

  1. 01Source systems: CRM, bookings, payments, accounting and advertising platforms.
  2. 02Integration and validation: data is brought in on a schedule and checked for gaps and duplicates.
  3. 03Approved metrics and access rules: the agreed definitions, and who may see what.
  4. 04Query layer: questions are answered using those approved metrics.
  5. 05The answer: shown with its sources and the time the data was last refreshed.

Where your systems don't connect directly, the integration and validation layer may need custom integrations, with each sync logged so any total can be traced back to the records behind it.

Join payments and refunds to orders using stable order IDs, then join each order to its CRM deal and captured campaign ID. Check that one payment can't appear twice after a join. Keep unmatched orders and unknown campaign sources in a visible exception report, with their counts and values, rather than silently dropping them from totals.

Core totals such as sales, revenue and churn should come from defined calculations, not from the model's own arithmetic. The AI's job is to interpret the question, choose the right approved metric and explain the result.

Access needs the same care:

Use read-only connections with the least access the reporting needs.
Enforce limits on rows and fields for each role in the application or database, not only through instructions to the AI.
Share only the data the questions require, and keep a log of the questions asked.
Check the AI provider's hosting and data-retention terms, and make stale data visible.

Connecting AI to business data doesn't make it secure by itself. Security comes from these controls, and they need reviewing whenever the system changes.

Give the Answer Business Context

Numbers become useful when the system knows what good looks like for your business. Give it approved context such as:

What counts as a good sale, and your target margin
The normal cancellation period for each product
How different service lines should be compared
The thresholds that should prompt someone to act

Keep these rules written down, versioned and open to inspection, so anyone can see why an answer said what it did. Documents can inform an answer, but they don't replace accurate joins between data sources and correct calculations.

Microsoft's guidance on preparing data for AI in Power BI makes a similar point: describing the data, setting verified answers and adding instructions helps AI give more relevant, grounded responses. The principle applies whichever tools you use, and answers still need checking.

A Useful Answer Shows Its Working

Here's one question answered from start to finish, using a single fictional acquisition cohort. The answer comes first, then the data and assumptions behind it. Open each question to check the sums and see why an answer needs more than the cheapest cost per lead.

Worked example: illustrative, fictional data

A fixed calculation walkthrough, not client results or a live AI tool. It isn't connected to any business systems.

Question: Which campaign produced better sales outcomes?

Answer: Neither, on these definitions. Campaign A has cheaper leads, but each campaign keeps 24 orders after refunds, spends A$250 in ads per retained order (retained CPA) and leaves A$8,400 contribution after ads and delivery.

Fictional cohort data
MeasureCampaign ACampaign B
Ad spendA$6,000A$6,000
Leads acquired300150
Paid sales3025
Full refunds61
Retained orders2424

What this answer relies on

Cohort
Leads acquired 1 to 30 September 2026, with sales and refunds observed through 31 October 2026, AWST.
Matched records
Every lead is matched to one campaign, and every paid sale and refund to one order. Duplicates have been removed and nothing is left unmatched.
One order per customer
Each customer places one order, so paid sales, orders and customers are the same count.
Amounts
All amounts exclude GST. Each retained order is A$1,000, with A$400 delivery. Refunded orders are assumed to have zero delivery cost, for illustration only.
Late sales and refunds
Sales and refunds after 31 October aren't counted, so a later sale or refund could change these results. Only full refunds are modelled.
Contribution
Net sales minus delivery and ad spend. It isn't net profit: overhead, sales commissions and tax aren't included.

Check the working

Which campaign has cheaper leads?

Campaign A: A$6,000 ÷ 300 leads = A$20 cost per lead (CPL).

Campaign B: A$6,000 ÷ 150 leads = A$40 cost per lead (CPL).

Campaign A's leads cost half as much. That doesn't tell us which campaign produced better sales outcomes.

What happens after refunds?

Campaign A: 30 paid sales − 6 full refunds = 24 retained orders. A$6,000 ÷ 24 = A$250 retained CPA.

Campaign B: 25 paid sales − 1 full refund = 24 retained orders. A$6,000 ÷ 24 = A$250 retained CPA.

Both campaigns have the same retained CPA. Stopping at paid sales would favour Campaign A, at A$200 per paid sale against A$240.

How much contribution is left after ads and delivery?

Campaign A: 24 × A$1,000 = A$24,000 net sales. 24 × A$400 = A$9,600 delivery. A$24,000 − A$9,600 − A$6,000 ad spend = A$8,400 contribution.

Campaign B: 24 × A$1,000 = A$24,000 net sales. 24 × A$400 = A$9,600 delivery. A$24,000 − A$9,600 − A$6,000 ad spend = A$8,400 contribution.

Both campaigns leave the same contribution. That's what remains after ads and delivery, not net profit.

What should someone check next?

Refunds: Campaign A lost 6 of 30 paid sales (20%). Campaign B lost 1 of 25 (4%). Review the refund reasons before changing either budget.

Follow-up effort: Campaign A produced twice as many leads for the same retained orders. Sales time isn't in this calculation, so compare it before treating the campaigns as equal.

The numbers show where to look. They don't show why the results came out this way, so these are questions for a person to investigate.

In a real implementation, every answer should link to the permitted source report and show the refresh time, attribution rule and any excluded records. If payments haven't refreshed, campaign IDs are missing or the refund window is incomplete, the correct response may be "we can't compare these campaigns yet". It should identify the missing evidence and who needs to resolve it.

Questions about people need extra care. If you ask why one rep has more cancellations, the data can compare tenure, lead source, product mix and time to cancel. It may flag a pattern worth discussing. It can't establish the cause, and it shouldn't be used on its own to judge someone's performance.

A good system will also say when it can't answer reliably from the current data. That's more useful than a confident guess. Decisions with real consequences still need a person to review them.

Start With a Small Set of Repeatable Questions

Choose five questions the business asks every week or month: net sales, retained CPA, contribution by campaign, cancellations by cohort and records that can't be matched. Define what decision each answer supports before connecting more data.

Roles in an initial reporting implementation
RoleResponsibility
Business owner and financeAgree the five questions, definitions, comparison windows and a trusted reference report.
EngineerConnect sources, reconcile totals, handle duplicate and unmatched records, and enforce permissions.
AI query layerInterpret the question over approved metrics and explain the returned calculation and limitations. It doesn't own decisions.
People using the reportsCheck exceptions, challenge explanations and decide what action to take.

Before release, use an acceptance checklist:

Check each answer against reports you already trust, including awkward cases such as refunds and customers who joined part-way through a month.
Test that each role sees only what it should, and what happens when data is missing.
Reconcile payment and refund totals to the agreed reference report for the same dates and timezone.
Show missing-source records and stop comparisons when a required source is stale or incomplete.
Confirm that unauthorised records stay inaccessible even when a question explicitly asks for them.
Record the failures, assign an owner and fix the definitions or data behind them before release.

Measure how often the system answers the agreed questions correctly, and how long it takes to reach an answer someone has verified. Faster confident guesses aren't an improvement.

If your existing dashboards already answer these questions well, improve those first. Plain-language questions are worth adding when they meet a real need, such as investigating a problem the dashboards don't cover.

Revision note

Updated 3 November 2026: we replaced the answer format with a worked campaign comparison using fictional data, and added guidance on joining payment, refund and CRM records, who does what in an initial implementation, and acceptance checks before release. Originally published 5 October 2026.