Gather the Reports, Trace Each Source and Decide What AI Can See

Over this course you will produce one reconciled figure for each contested metric. A contested metric is one where departments report different numbers, such as revenue or churn. Each reconciled figure comes with three things attached:

• its definition: what it counts and what it leaves out
• its period: the dates it covers
• its source: the system it came from

You will also write a short note on what still differs and why. With that note, the meeting can discuss the gap instead of arguing over whose number is right.

Some things are out of scope. You will not build dashboards, forecast next quarter or choose strategy from the numbers. Those come after you have figures everyone accepts.

This lesson is the groundwork. You collect every figure, trace where each one came from, and decide what an AI tool is allowed to see. If you skip this step, later comparisons rest on guesses.

Before you compare anything, list every figure in one table. Use one row per figure and six columns:

1. Metric name as written, for example "Net revenue" or "Bookings"
2. Figure
3. Department
4. Report name
5. Period label as written, for example "Q3 FY25", "September" or "Last 30 days"
6. Source system: the software the number was pulled from

Copy the labels exactly as each report shows them. Don't tidy them up yet. Different wording is often the first clue to a different definition.

Here is an example. Finance reports "Net revenue, September: $4.20m". Sales reports "Revenue (closed won), Sep: $4.65m". Operations reports "Billed revenue, Sept: $4.38m". That makes three rows and three different names, and the figures are up to $0.45m apart. Churn gets its own rows too, for example sales at "Logo churn 3.1%" and operations at "Churn 2.4%".

At this stage you are not deciding which figure is right. You are making sure every figure is on the table.

Each department usually draws from a different source system. These systems record different events.

• General ledger or accounting system (the company's official books): usually feeds finance. Revenue is typically recorded when it is earned under your accounting rules.
• CRM (customer relationship management, the sales team's record of deals): usually feeds sales. It often records a deal when it is signed or marked as won.
• Billing, subscription or product usage systems: often feed operations. They record invoices raised, subscriptions cancelled or customers who stopped using the product.

So "revenue" from the CRM might mean deals signed in September. From billing, it might mean invoices sent in September. From the ledger, it might mean revenue earned in September. All three can be correct for what they measure and still disagree.

This is why the source column matters. A gap between two figures is not automatically an error. It may be two systems answering slightly different questions. Your company's systems may be set up differently, so confirm with each report owner rather than assuming.

Add two more columns to your inventory: extract date and manual adjustments.

An extract is the moment someone pulled the data out of a system. Systems keep changing after month-end. Late invoices get posted and cancelled deals get updated. Suppose finance pulled September on 3 October and operations pulled it on 6 October. Operations will then include anything posted in those three days, and finance won't.

A manual adjustment is a change someone made by hand after the extract. It might be removing a disputed invoice, adding an accrual, or excluding a one-off customer. These are often sensible. But if they aren't visible, they make two reports disagree for reasons nobody can see.

So two reports can share the same definition and still show different figures because of timing or hand edits alone. Try the sliders. Both reports here use the same definition. The only differences are when each was pulled and one adjustment.

Many reports don't show their source system or extract date. When a detail is missing, write "not stated" in that cell. Don't guess. A guessed source looks just like a confirmed one in your table, and it will mislead you later.

Then ask the report owner, the person who produces that report. Keep the question short and neutral so it reads as curiosity, not an accusation. For example:

"Hi Priya, I'm lining up the September figures before Thursday. For the $4.65m revenue in the sales report: which system and which date was this figure pulled from, and was anything adjusted by hand afterwards? Thanks."

This message works because it names the exact figure and asks for facts. It doesn't hint that the number is wrong. People answer questions like this quickly. They get defensive when they think they are being audited.

When the answer arrives, record it in your inventory and replace "not stated". If no answer arrives before the meeting, leave "not stated" in place. That is honest, and it is useful to say out loud.

Before you paste anything into an AI tool, check your company's policy on approved tools. These are the tools your organisation has cleared for business data, often with contracts that control how the data is stored and used. A free consumer chatbot is usually not one of them.

Some details should stay out of unapproved cloud tools:

• Board discussions, including draft board figures
• Acquisition plans, including targets or deal values
• Employee details, such as names tied to performance or pay
• Customer names, for example a churned client listed by name

Go through your inventory and mark each row as in, remove or replace, or keep out. A row reading "Churn 2.4%, operations, Sept" may be fine in an approved tool. A note reading "includes loss of Acme Ltd" names a customer, so it should go.

If you are unsure whether a tool is approved or a detail is sensitive, ask IT or legal before you share it. A short delay is cheaper than a data leak.

You can still get real help from AI without an approved tool. You just share less.

AI is useful here for structure: laying out comparison tables, spotting that two labels probably mean different things, or suggesting questions to ask. It doesn't need your real numbers to do that.

Share only:

• Metric labels, such as "Net revenue" and "Revenue (closed won)"
• Definitions, as each report describes them
• Period descriptions, such as "fiscal September, cut-off 30 Sep" or "last 30 days"

Replace real figures with placeholders, which are stand-in labels like "[Finance revenue]" or "[Sales churn %]". Then take the AI's table back into your own spreadsheet, fill in the real numbers and do the arithmetic yourself.

Whatever tool you use, check every figure. AI can misread a label or get a subtraction wrong. You are the one presenting the numbers, so you need to be able to say how each one was produced.