Trustworthy AI

An analyst that can’t make things up.

Sarah is InCite’s built-in AI advisor. Because she works inside a financial and compliance platform, a made-up number isn’t just unhelpful — it’s harmful. So we built her to surface only what’s actually in your data, to cite where every figure comes from, and to say “I don’t have that” instead of guessing.

Your question Sarah Your data documents + validated records this tenant only General AI knowledge the web · training · benchmarks blocked for your figures × Answer — grounded & cited
For your organization’s numbers, Sarah may use only your data — never the open web or training-data guesses.
Grounded in your data

She answers from your records — not the internet.

Every answer is assembled, in the moment, from your organization’s own uploaded documents and validated records — scoped to your tenant and no one else’s. Sarah is explicitly instructed to answer only from that context, never to fill a gap with a benchmark, an industry average, or a plausible-sounding estimate.

  • Your data only. The context is built from your authenticated organization’s files — another tenant’s data can never enter it.
  • Verified snapshots, not live guesses. Documents are read and summarized once at upload; Sarah works from that stable, parsed text — not a fresh re-interpretation each time.
  • No web for your numbers. Sarah is barred from using web search for your organization’s own figures — web access is reserved for public, regulatory facts only.
Extraction & validation

Every number is pulled, checked, and kept with its source.

Key figures — revenue, expenses, headcount, funding, contract values — aren’t paraphrased from prose. They’re extracted into typed fields, validated against a schema the moment they’re ingested, and stored with a record of where they came from. A value that doesn’t pass validation is flagged, not saved.

Documentyou upload acontract or report Extractionmodel fillstyped fields Validationtype · range ·format checked Canonical recordstored withprovenance Cited answershown withits source Malformed → flagged, not stored
Figures become structured, validated records — or they’re flagged. They’re never quietly accepted.

And when two documents disagree, InCite doesn’t average them or pick at random — it defers to the more authoritative source.

MORE AUTHORITATIVE Form 990audited federal filing Audited financials Unaudited actuals Budget Forecast When sources disagree, the higher rung wins.
Provenance is tracked on every figure, so the most trustworthy source always governs.
Citations

If she can’t show you the source, she doesn’t claim it.

Sarah’s structured outputs require evidence. Every assessment must name the specific document or data point that supports it; if it can’t, the supporting evidence stays empty and the confidence is set to zero. Nothing is asserted without a traceable basis.

S

Your committed revenue drops about 68% by mid-2027 if nothing renews — driven mainly by two federal contracts ending in 2026.

Supporting evidence
USDOL_ETA_Youth_Workforce_FFY2026.pdf$1.25M · ends Sep 2026
WIOA_Essex_WDB_Agreement.pdf$540K · ends Sep 2026
Illustrative — every figure Sarah surfaces can be traced to the document it came from.
  • Evidence required. Each claim carries the source behind it — surfaced as supporting-evidence boxes you can open.
  • Honest confidence. Every assessment carries a 0–100 confidence score, set high only with strong, specific evidence; low-confidence items show their gaps instead of inflating.
  • Structured, not freeform. The model fills a validated schema — a field with no evidence stays empty, never invented.
  • Charts can’t lie. Visuals are drawn from canonical records by exact ID; incomplete inputs show a gap label rather than a fabricated shape.
Honest about gaps

“I don’t have that yet” is a valid answer.

Where the data is thin, Sarah tells you exactly what’s missing and why — a named list of the documents or data points she’d need. You always see the gap; you never see a confident answer built on nothing.

Human in the loop

Sarah flags. People decide.

When Sarah spots an inconsistency across your data, she surfaces it for review — open, visible, and dismissible. She never silently acts on a finding or edits your records on her own. The judgment stays with you.

Guardrails

The rules she runs under.

These instructions ship in Sarah’s system prompt on every message — non-negotiable, and applied before she ever sees your question:

“Answer only from the organization context provided.”
“Never speculate about figures not present in the data.”
“If the data is too thin to say anything useful, say so — briefly and honestly.”
“Do not search the web for organization-specific facts.”
  • Web research is off by default. A platform operator must enable it per organization, and even then it’s limited to public, regulatory information.
  • No memory across sessions. Each conversation starts fresh from your current data — so answers can’t drift or go stale on yesterday’s numbers.
  • Your question is kept separate. What you type is never folded into Sarah’s instructions, so a cleverly-worded prompt can’t talk her out of her guardrails.
Boundaries

What Sarah will not do.

× Invent financial figures that aren’t in your data
× Infer your headcount or revenue from industry benchmarks
× Cite a web source for your organization’s own numbers
× Give a confident answer when the underlying data is absent

When the data isn’t there, she says so — and tells you exactly what would fill the gap.

At a glance

Every risk, mapped to a control.

Hallucination risk
How InCite prevents it
Model invents financial figures
Your validated records are provided in every prompt; the model is instructed never to infer figures.
Model fabricates facts from the web
Web search is off by default and limited to public/regulatory info — never your numbers.
Charts show invented data
Visuals are built from canonical records by exact ID; missing inputs show a gap label, not a shape.
An assessment overstates certainty
Confidence scores and source citations are required; confidence is zero when nothing supports it.
Answers drift from your real data
Context is rebuilt from your records at every question; there’s no persistent model memory.
One tenant’s data leaks into another
Tenant identity comes from a verified sign-in claim; every query is scoped server-side.
Stale answers from old documents
Each document is snapshotted at upload; analyses re-synthesize on demand from current data.
Trust, by construction

Sarah is fast and plain-spoken — but never at the expense of being right. When she tells you something, you can click straight through to where it came from.