Ask three people in the same company "what were our sales last month?" and you may get three answers. Finance counts invoices issued. Sales counts orders confirmed. Operations counts what actually left the warehouse. Each number is right by its own rules. The trouble starts when two of them end up in the same meeting.
Connect an AI assistant to those systems without agreed definitions, and it may use a different rule from the one in your finance report. When its answer conflicts with a report people already trust, confidence in the assistant falls quickly.
A better model cannot settle what "sales" means inside your business. That is a decision your business has to make, and it is worth making before anyone, person or AI, answers with the number.
One question through a governed data architecture. Figures are illustrative.
What a definition records
A useful definition of a business number usually records six things:
- Name: the one name everyone uses, with the other names people say listed as aliases ("revenue", "turnover", "sales").
- Business meaning: one sentence a manager would agree with.
- Formula: how it is calculated, including what is excluded, such as returns, credit notes, internal transfers or GST.
- Scope: which entities, sites and customer types, and which date counts: invoice, order or delivery.
- Source: which system, table or report is the reference when two disagree.
- Refresh: how often it is updated, so nobody compares today's figure with last night's.
Add the breakdowns people ask for, such as by site, product or customer type, so everyone splits the number the same way.
Many disagreements come down to the formula, the scope or the source. "Sales including GST, by invoice date, from the accounting system" and "sales excluding GST, by delivery date, from the warehouse system" are different numbers that share a name.
One name, one meaning, one calculation
Two rules help keep definitions consistent as they grow. A name should mean one thing. And that thing should be calculated the same way wherever it appears. When a team needs a variant, give it its own name, such as "net sales" or "shipped sales", rather than quietly changing the old one.
Each number also needs an owner: the person who approves its definition and decides when it changes. When a definition does change, record the new version with the date it takes effect, and retire the old one. Otherwise last year's figures get compared with this year's on different rules, and nobody notices.
Sometimes the mismatch is in what is being counted, not how. If the same customer exists twice, under different codes in two systems, a "sales by customer" report may split their sales across two records or count them twice. Agreeing who a customer, a product or a supplier is, and keeping those records in step across systems, has to be worked on alongside the definitions whenever a question depends on them.
Two ways to let AI answer
There are broadly two ways to let staff ask an AI assistant about business data.
Let the AI write the query. The assistant reads your table and column names and writes a database query for each question. It is quick to set up and flexible, which makes it useful for exploring. Its weakness is the problem above: it has to infer the definition each time, and how well it does depends on how clearly your tables are named and documented.
Give the AI the definitions to use. You first record the agreed numbers, their formulas, breakdowns and the names people use. The assistant then maps a question to those definitions instead of inventing a calculation. It takes more work up front, and questions that need a number nobody has defined yet still need extra work. In return, the assistant and the report use the same calculation, and a change to a definition is made in one place. Differences can still come from the data itself or from refresh timing, but the definition is no longer one of them.
One practical path is to use direct queries for exploring, then define the numbers people rely on every week. What tends to go wrong is skipping definitions entirely and hoping the AI will work them out.
How these projects tend to stall
The pattern is familiar. Someone senior sees a demo. A trial on sample data looks impressive. Then it is pointed at real data, its answers stop matching the existing reports, and the team starts adding a rule for each wrong answer. Every rule makes the next one harder to maintain, and the project quietly stops.
A more manageable start is narrower:
- Choose one group of users and the ten or twenty questions they actually ask.
- Write definitions for the numbers those questions depend on, and have each number's owner approve them.
- Check answers against the reference report before anyone relies on them. When an answer fails, the owner finds out why and corrects the definition, the data or the source.
- Treat each correction as an update to the definitions, not a patch to the assistant.
- Add questions and numbers only after the first set holds up.
Expect most of the effort to go into definitions and data, not the AI.
Where to start next week
You don't need a data platform to begin. Take the five figures your managers most often have to reconcile. For each, write down the six things above, the owner, and the report it should match. Those definitions help managers and analysts straight away, and any dashboard or AI assistant you add later will need them.
When you are ready to let an assistant work with your own system, our AI connect service starts with one task. Each thing the assistant can do is a named action with clear limits, it has its own access separate from your staff, and anything that changes money, stock or customer records waits for one of your people to approve it.
See how it works: AI connect.
Originally published on dynokore.com.


