AI evidence engine · built for consulting firms

Every finding traces back to the exact words that prove it.

Trace turns the exhaust of an engagement — calls, interviews, documents — into a base of findings where every claim carries a verbatim quote, located at an exact character offset by a deterministic verifier. If the quote isn't in the source, it never becomes a finding.

no finding without a source span
nothing written to the record without review
contradictions surfaced side by side, never resolved for you
FINDING · F-042 ✓ verified

Onboarding runs about eleven days before a new consultant reaches billable work.

“…honestly it's about eleven days before someone new is on anything billable, sometimes a bit longer…”
S. Hahn · client as of Q1 2026
P-STARTUP · onboarding interview · 2026-02-14
exact match · chars 8,214–8,289
● the cost of an unsourced number

One fabricated figure is a partner-level event.

In 2025 a major firm delivered a government report that contained AI-fabricated citations, and refunded part of its fee. When synthesis can't be traced, the liability lands on the partner who signed it. Trace inverts the order of operations: the verbatim quote comes first, and a claim only exists once that quote has been located in the source.

A$440K
government report · references generated by AI could not be found in any source · fee partially refunded. The failure mode Trace is built to make structurally impossible.
How it works

A pipeline that treats every quote as a fact to be located.

Five stages, each behind its own interface so any vendor or model can be swapped. Order matters — verification happens before anything is attributed, answered, or drafted.

01

Capture one intake

Teams calls arrive through Microsoft Graph — no bot in the room, native consent. External calls, manual uploads, and engagement documents land through the same contract.

02

Extract & verify the moat

A model proposes each claim with its verbatim quote. A deterministic verifier finds that quote in the source at an exact character span — or rejects it as a hallucination. Offsets emitted by the model are never trusted.

03

Attribute who & when

Every finding is tied to who said it, from a stakeholder registry, and to the period it is about. Dates are normalized in code against the engagement's fiscal calendar — never guessed by the model.

04

Reconcile time-aware

Contradictions are shown side by side and never resolved for you. A newer fact supersedes an older one only after you confirm it — and the superseded fact drops out of reads while staying in the record, never deleted.

05

Answer & draft cited

Ask a question and get an answer that cites [F#] markers back to verified spans. Draft a firm-branded deck from the evidence — labeled a draft, pending your review.

✓ the trust layer

Six rules the product will not break.

These aren't settings. They're invariants — the reason a partner can put a Trace finding in front of a client and stand behind it.

No finding without a source span

The verifier locates every quote deterministically. Unlocatable quotes are rejected, not quietly down-weighted.

Never a silent write

Low-confidence output goes to a review queue. The record changes only when a human accepts it.

Contradictions stay contradictions

Conflicting statements are surfaced side by side. Trace never picks the winner for you.

Newer wins only when you say so

Supersession needs human confirmation. Superseded facts leave the current view but are never deleted.

Every edit is evidence

Each human change to an AI output is captured, append-only, with a reason. The audit trail is a feature.

Nothing client-facing ships unreviewed

Answers and decks are drafts until a person signs off. The AI never has the last word.

See it work

Answers that show their sources. Disagreements kept honest.

Ask the evidence and every claim carries a citation. Select a marker to see the verbatim quote it stands on.

Q · How long is onboarding, and how has it changed?

New consultants reach billable work in about eleven days , an improvement over the two-to-three weeks the team described a year earlier . Both figures come from client-side interviews, each anchored to the period it describes.

F1 · onboarding duration✓ verified
“…about eleven days before someone new is on anything billable…”
S. Hahn · clientas of Q1 2026
F2 · onboarding duration✓ verified
“…back then it was two, maybe three weeks before they were useful on a file…”
S. Hahn · clientas of Q1 2025
difference of opinion
Ops lead
“the approval loop is about three days end to end, on a normal week.”
CFO
“realistically that sign-off is a two-week cycle once finance is in it.”

Same process, two roles, two answers. Trace shows them together with both quotes verified — and leaves the reconciliation to you.

Methodology, delivered

Your firm's way of working, inside your consultants' own AI.

Trace serves your firm's methods and each engagement's brief over MCP — so Claude, ChatGPT, or Copilot answer the way your firm actually works. Skills are versioned, updated once for everyone, and every fetch is logged. Connecting is one click: paste the Trace URL, approve in the browser, done. No token to copy.

list_skills get_project_brief search_findings ask_evidence get_contradictions
connected · claude
You · draft the interview guide for a cost-diagnostic engagement.

Pulling interview-guide v1.0.0 and the project brief from Trace… structured by workstream, with coverage tracked against the stakeholder registry. Draft ready for your review.

skill v1.0.0 fetch logged
Fits your Microsoft world

Built for M365 firms — and for the people who approve them.

Trace meets an IT reviewer where they already are. No new meeting bots, no client data used for training, and a pilot that keeps every control inside your own tenant.

Microsoft 365 setup guide →

Teams transcripts via Microsoft Graph. No bot joins the meeting; capture rides native consent.
Lives in a Teams tab and in the browser — nothing new to install for consultants.
Zero-data-retention models. Confidential work runs Sonnet-class under ZDR; nothing trains on your clients.
Data hosted in Canada (ca-central-1), on Postgres you can point to.
Single-tenant pilot. The app registration and every kill switch stay in your tenant.

Bring verifiable synthesis to your next engagement.

Trace is piloting now with a Toronto consulting firm. If your work ends in a deck a client relies on, it belongs on evidence you can defend — quote by quote.

Request a pilot design partners onboarding · single-tenant · M365-native