Understand the question
Which vendors did we pay more than $50,000 to last year. What is our largest unbilled project. How many customers pay later than their terms.
AI capability
The value is not that you can type a question. It is that the answer arrives with the transactions behind it, computed within your permissions, and that the system declines when the data does not support the question rather than producing something plausible.
What it does
Which vendors did we pay more than $50,000 to last year. What is our largest unbilled project. How many customers pay later than their terms.
It composes against the underlying model rather than picking from a catalogue of pre-built reports, so the question does not have to have been anticipated.
Every figure links to the transactions that produced it. A number you cannot open is a number you have to trust, which is the wrong relationship.
Computed within the asker’s permissions in the data layer — not computed broadly and then redacted, which leaks the aggregate.
Any answer saves as a view, schedules as a report, or exports. Most people use conversation to discover the report they wanted.
Unclosed period, missing dimension, ambiguous phrasing — each produces a specific explanation rather than a confident approximation.
Natural-language querying has been demonstrable for a while and is not, by itself, the interesting part. The interesting part is whether you can check the answer without rebuilding it.
A number returned without provenance has to be trusted, and finance people are correctly unwilling to trust numbers they did not derive. So every answer here carries the filters applied and the transactions included, one click away. The first thing most new users do is open the transactions to see whether the figure holds — and after a fortnight of it holding, they stop.
Ask about revenue last quarter and there are at least three defensible readings: booked, recognised, or collected. A system that silently picks one is wrong a third of the time in a way nobody notices.
Ours asks which you meant, once, and remembers the preference. That is marginally slower and considerably more useful, and it is the same principle as the copilot declining rather than approximating.
Scope is applied when the query is composed, not to the results afterwards. The difference matters: a system that computes a company-wide total and then hides the departmental breakdown has already exposed the aggregate, which is frequently the sensitive figure.
A department head asking a company-wide question gets a clear refusal naming the scope limit, rather than a partial answer they might mistake for a complete one.
Limits
Every capability page on this site carries one of these, because a feature described without its boundaries is a claim rather than a description.
Questions about the future route to explicit modelling in the CFO agent. Conversational extrapolation is where confident nonsense comes from.
No phrasing recovers a department that was never coded. It explains that rather than deriving something adjacent and presenting it identically.
It cannot post, approve, or change anything, at any authority level. Asking returns a refusal and a link to the workflow that does it under controls.
Questions
Two periods of data and one real question is enough to judge whether the answers hold up.