Could AI explain how our performance changed? Not yet: five definitions disagreed.
The owner wanted AI to explain why performance had changed over time. Before any explanation, we checked whether the numbers being compared measured the same thing. Five did not.
Why it matters to any businessAny business comparing figures across periods or systems fails the same way when definitions drift.

- Client
- Private property owner · Dubai, UAE
- Engagement
- AI adoption review · 1 unit
- Focus
- Consistent definitions before AI explains a trend
- Sector
- Luxury property
Trend explanations wait for an approved data dictionary. Until then, differences are labelled, not explained.
- 5
- definition conflicts found
- 20
- fields in the minimum data dictionary
- 0
- figures the assistant may produce itself
- The request
- Let AI explain how the property’s performance changed over time.
- What we found
- Five definition conflicts between historical reporting and the later dashboard.
- The answer
- Not ready for AI trend explanation until a 20-field data dictionary is approved.
The five conflicts
Historical management reporting and the operator’s later dashboard looked comparable. They were not:
- Different occupancy bases. One measure never documented its denominator; the other counted unavailable and blocked nights. They cannot form one trend.
- Different reporting windows. The two sources covered overlapping but different periods.
- Conflicting length-of-stay figures. Two views gave different averages under different filters, and neither can be relied on until the data settles which applies.
- An unclear period boundary. A chart label ran past the end of its filter, so the final month had to be confirmed at source.
- Weak baselines. Some series held zero values for early months, so percentage changes against them could not be read as growth.
None of these is an error in the owner’s own reporting. Each is a reason to fix a definition before a model explains a movement.
Why AI makes them dangerous
A person reading two inconsistent charts tends to hesitate. A language model tends to explain. Asked why occupancy rose, it would have produced a fluent, plausible answer built on two different denominators, and the fluency would have made the error harder to see.
The temptation runs in one direction: combining the figures would have told a stronger performance story. We kept them apart and labelled, and left unsupported comparisons open.
What must come first
An approved data dictionary, before any AI analysis of an owner’s figures. At minimum it defines 20 fields: property, owner, currency, transaction date, service period, booking status, booked value, earned revenue, refunds, channel fees, taxes and charges, expense category, invoice, payment status, management fees, owner allocations, opening and closing balances, reserves, distribution authorization and report version.
Generated answers are then tested against cases where denominators, currencies or date windows differ. The assistant labels what is final and what is provisional, says when a question has no answer yet, and puts figures into words without ever producing one.
Key takeaways
- Before explaining why a number changed, confirm that both numbers measure the same thing.
- An unresolved discrepancy stays visible until someone accountable reconciles it.
How this fits the RoadmapIn an AI Adoption Roadmap, the data behind each AI idea is checked before it is recommended, and sometimes the answer is to fix the data first.
Property owners’ identities and confidential financial and operational data are withheld.
Make your next AI investment an informed decision.
Begin with a 30-minute Clarity Call about your own business.