Security / AI and accuracy
The AI analyst
The dashboard figures are computed without any AI, as the pipeline page sets out. The AI analyst is a separate layer you can connect to ask questions in plain English. This page covers what it can see, why it cannot change anything, and how we keep its answers honest.
The analyst can read and explain your data. It cannot change it.
The analyst reaches your data only through a defined set of read tools. It can run a query and read the result back to you. It has no tool that writes, edits, or deletes, so there is no path through which it could alter a record even if asked to.
Read-only is enforced at the database, not just the tools.
Two independent controls back this up. The query interface only accepts read statements, and the login the analyst uses is granted read access and nothing more. A write cannot succeed even if something upstream tried to issue one, because neither control depends on the other.
Queries are bounded so a heavy request fails cleanly.
Each query has a limit on how long it can run and how much it can return. A pathological or runaway request times out and stops rather than holding a connection open or pulling the whole database back.
The analyst reads finished numbers. It does not invent them.
The figures already exist, computed by the pipeline. The analyst's job is to find them, combine them, and explain them. When it cites a number, that number came from a query you can re-run, not from the model's own guess.
Revenue is reconciled against your accounting system's own report.
Month by month, the revenue Clarity shows is checked against the official P&L report from your accounting system. When they match, you can see that they match. This tie is on revenue. We do not claim a full profit reconciliation, because the transaction-level view omits items like payroll, accruals, and journal-only entries; for profit, the dashboard reads your accounting system's own report rather than rebuilding it.
Where it can be wrong, and what catches it
The analyst is a language model, so it can misread a question or reason poorly. Three things keep that in bounds. It cannot change your data. Every number it gives you traces to a query you can inspect. And the dashboard is the verified baseline, so if an answer disagrees with the dashboard, the dashboard stands. The analyst recommends and explains; it does not act on your behalf.
The written analysis is reviewed by a person
The number pipeline is automated. A written narrative report is an AI-assisted draft that a person reviews before it goes to you. We do not ship fully automated written analysis.
Related: how a client's AI connection is authenticated, and the sub-processors involved when you connect the analyst.