Grounded in your business, not the internet.
Proactive, not just responsive.
Deterministic signals on your records — overdue receivables, scheme gaps, PO quantity drift, dormant customers — that explain themselves, act with one tap through governed APIs, and can be snoozed. Plus the daily AI brief: one metered summary per user per day over facts they’re allowed to see, and a Sense home page that puts “needs your attention” first.
Ask what you’d ask a colleague.
Nine nudges, stated plainly.
Nine deterministic signal providers, each producing a finding with structured evidence and — where applicable — a one-click action applied under the user’s own permissions and written to the audit trail. The findings are computed deterministically; Sense explains them, answers follow-ups, and carries the action. Deterministic signals, AI delivery.
Tenants can also author agents that publish their own nudges alongside the built-in nine.
Charts when a chart is the answer. Your dashboard when it already is. The screen when you ask.
Sense is not only a text box. It draws, it runs what the business already trusts, and — with your confirmation — it operates the application in front of you.
The capabilities, stated plainly.
Every row ships today and traces to a capability described on this page.
| Capability | xMatix Sense Assist |
|---|---|
| Plain-language questions over your records | Native |
| Answers grounded in your data, with citations | Native |
| Runs with the asking user's permissions | Native |
| Document extraction and retrieval | Native |
| Drafts land as approvals, never silent writes | Native |
| Proactive nudges with one-tap actions | Native |
| Daily AI brief per user | Native |
| Works on web and in the field app | Native |
| Every action metered and audited | Native |
The questions evaluators ask.
Can Sense see records the user cannot see?
No — Sense runs as the person asking, through the same permissions that govern every screen. If you cannot open a record, neither can your assistant, and the answers it gives are grounded in the records you can access, with citations showing exactly which ones. There is no privileged service account behind the curtain answering from data you were never meant to reach.
Can it change data, or only answer questions?
Both — but it never writes silently. Sense drafts the change and lands it as an approval: the reminder emails, the report, the record update all wait for your confirmation before anything persists. Proactive nudges act the same way, executing through governed APIs with one tap from you. Answering is autonomous; acting always has a human on the trigger.
Does it need to be trained on our data?
No — there is no model training on your data. Sense reads the manual instead: module cookbooks and per-field AI descriptions teach it what your data model means, and for canned business questions it finds and runs your existing report rather than inventing the logic. Documents attached to records flow into the knowledge base automatically, searched with citations — retrieval, not training.
What stops a cleverly worded prompt from getting more access?
Permissions are enforced outside the model. Every read and write Sense performs goes through the same APIs and record-level security as the signed-in user — a prompt cannot grant access the account does not have, because the access check never consults the model. Writes additionally pause for human approval, and every action is metered and audited, so even the attempt leaves a trail.
Is what Sense does recorded?
Yes — every action is audited and metered. Each session is stamped with the user Sense was acting as, answers carry citations to the records that grounded them, and usage draws against visible credits and budgets. When someone asks “why did this change” six months later, the answer is in the audit trail, not in anyone's memory.
What does Sense do without being asked?
It watches for nine kinds of findings — the nudge library above — from a purchase-order line the pipeline already covers to an account gone quiet against its own pattern. Each nudge arrives with structured evidence and, where applicable, a one-click action applied under your own permissions and written to the audit trail. Tenants can author agents that publish their own nudges alongside the built-in nine, using the same delivery and audit path.
Are the suggestions generated by a language model?
No — and that is deliberate. The findings are computed deterministically by signal providers working over your records: the scheme-gap shortfall, the pipeline netting, the receivable ageing are arithmetic, not generation. What the model does is deliver them — Sense explains the finding, answers follow-ups, and carries the action. Deterministic signals, AI delivery: the numbers are checkable, and the conversation about them is natural.