Why does Unloq give the same answer every time?
The short answer
AI tools give different answers to the same question when the model works out what a metric means each time someone asks. Unloq settles that meaning in advance: every metric resolves to one definition held in a semantic layer your team co-authors with ours. ARR for the third quarter of 2026 comes back as one number, traced to that definition, whether you ask in January or in April, as long as the data and the definition are unchanged. One business put the problem to us as two analysts, two stories.
What is the semantic layer?
The semantic layer is a runtime dependency: every query Unloq runs passes through it. The layer holds the following:
the canonical formula for each metric;
the mapping from business terms to SQL;
entity resolution rules, so a customer held under different names in two systems is recognised as one customer;
the version of every metric.
The language model plans the steps, the semantic layer resolves each metric to its one definition, and every query that runs is governed SQL against an approved list of tables. Annual recurring revenue, for example, is stored as one calculation, and a question that says “ARR”, “recurring revenue” or “annual recurring revenue” resolves to that one calculation.
Where do definitions come from?
Your team and ours write them together in the first week of onboarding, starting with the metrics the first decision uses. A definition holds the metric’s name, the other names people use for it, and the exact calculation behind it, down to the filters and the date rules. You own the definitions.
Microsoft’s own Power BI guidance says data preparation features can’t ensure a specific output every time because AI behaviour is nondeterministic. We think that diagnosis is right. Microsoft’s guidance is to prepare your data well; Unloq keeps each metric’s calculation in a stored definition, away from the language model.
What happens when two teams mean different things?
The difference gets settled in the definitions. If finance counts revenue when it is invoiced and sales counts it when it is booked, your business picks which one Unloq uses. Where two sources disagree, Unloq shows the mismatch so it becomes a visible decision, and a rule you agree decides which source wins. Two merged businesses resolve what “revenue” or “customer” means once, in a definition you own, and both keep their own systems.
Which number your board sees is decided by whoever owns the metric. In our experience the agreeing is the hardest part of setup, and it is work only your people can finish, because the definition has to mean what your business means.
Who confirms a definition, and is a change recorded?
Each definition is confirmed by a named owner on your side. Definitions are versioned in the semantic layer, so a change to how margin is calculated becomes a new version, and the earlier version is kept alongside it.
Does rephrasing the question change the answer?
No. “What was ARR last quarter?” and “how much recurring revenue did we have in the last quarter?” resolve to one definition, so both return one figure. The figure changes only when the data or the definition changes, and the trail on each answer shows the data it used.
What this page does not cover
How Unloq connects to the systems behind each definition is on how Unloq connects to your data. The definitions feed Describe, where Unloq pulls every system you connect into one picture of what happened, and Contextualise, where Unloq works out what drove a change and by how much. For how this compares with Copilot and ChatGPT, read what decision intelligence adds to Copilot.
Ask us your question
Bring the metric your teams have argued over longest and ask us your question about it. Our AI engineering team works the question through with our technical founders and replies with an answer.
Need this in writing for your review?
Request the security pack and tell us what your reviewer needs.
