A category defined by one question: "Where did this number come from?"

Every sales leader has lived this moment. A forecast review, a dashboard full of confident figures, and one uncomfortable question: where did this number come from? The honest answer, for most sales data, is some mix of "the rep typed it in," "the AI summarized it," and "nobody remembers." The data looks precise. Its origins are vapor.

Provenance-tracked deal intelligence is sales intelligence built so that this question always has an answer. Every claim, number, and stakeholder attribute the system produces carries a verifiable evidence chain — a tag stating what kind of knowledge it is and where it came from. Three tags cover every case:

  • FOUND — taken from a source. The prospect said it in a call, wrote it in an email, or it sits in the CRM record. The claim links back to the exact place it came from, down to the transcript line.
  • ASSUMED — an AI inference. Plausible, useful, and explicitly flagged as something to verify rather than something known. The system tells you it is guessing instead of dressing the guess as fact.
  • CALCULATED — derived from found data, with the formula shown. An ROI figure is only as good as its inputs and its arithmetic; CALCULATED exposes both.

The discipline sounds simple. Almost no sales tooling does it — and that absence explains a great deal about why sales data is trusted by no one who has to act on it.

Why CRM data fails trust

CRM data is self-reported. The rep decides what to log, when to log it, and how optimistic to be. Stage fields reflect the seller's activity and mood, not the buyer's decision state. "Decision maker" and "champion" are free-text fields nobody maintains. None of it is audited — there is no mechanism by which a CRM entry confronts what was actually said in the room.

This is not a discipline failure by reps. It is structural: the CRM records interpretations, and interpretations drift toward hope. A pipeline review built on it becomes a negotiation between the rep's optimism and the manager's skepticism, with no evidence on the table to settle the difference. That is why deals "slip" every quarter — the slippage was always there, hidden inside unaudited fields.

Why AI notetakers fail trust differently

AI notetakers and conversation-intelligence tools were supposed to fix this by capturing the source itself. They capture it — and then summarize it into the same unaccountable mush. A summary is an interpretation with the receipts thrown away: you cannot tell which sentence came from the buyer's mouth and which the model helpfully inferred. When an AI confidently states "budget confirmed at 200K" and the transcript actually contains a vague maybe, there is no flag, no trace, and no way to catch it short of re-listening to the call.

So the market arrived at an odd place: tools that record everything, and outputs that can verify nothing. The raw material for trustworthy sales data exists in every call recording and email thread. What was missing is the layer that keeps each claim chained to its evidence — and admits when there is none. The honest question a buyer once put to us was whether the AI would simply invent data. With per-claim provenance the answer is structural, not aspirational: an invented number has no source to link to, so it cannot wear a FOUND tag.

What changes when every claim carries its evidence

Forecast reviews stop being interrogations

With provenance, a deal review starts from evidence instead of recollection. The Decision Gate status of each stakeholder is read from what they actually said — confirmed, assumed, or not mentioned — and every supporting metric shows its tag. The manager no longer cross-examines the rep; both look at the same evidence chain. Reviews get shorter and decisions get sharper, because the question "is this real?" is answered before the meeting starts. Deals advance on confirmed gates, not on a rep's CRM update — the full model is in the methodology.

CFO conversations stop dying in silence

Complex deals are won in rooms the seller never enters: the buyer's internal pitch to their CFO or management. Conventional collateral fails there because an approver's first instinct is to distrust vendor numbers — rationally, since those numbers carry no evidence. Provenance inverts this. A champion kit in which every figure is tagged FOUND with its source quote, or CALCULATED with its formula, and every assumption is visibly flagged, invites verification instead of resisting it. The CFO can audit the claim chain in minutes. Material that defends itself is the difference between an internal sale that takes a week and one that quietly evaporates — and it is a major reason deals built this way close in 2–3 calls instead of months.

Assumptions become work items instead of landmines

A flagged assumption is an agenda. If the stakeholder map says the technical lead's budget influence is ASSUMED, the next call should confirm it. Unflagged assumptions, by contrast, surface late as "surprises" — the blocker nobody mapped, the authority that was never real. Tagging inference as inference converts hidden risk into a to-do list.

The EU AI Act makes this architecture mandatory in spirit

European regulation is converging on the same principle from the compliance side. The EU AI Act's obligations for AI systems center on transparency (users must be able to understand what the system produced and on what basis), traceability (outputs must be reconstructable), and human oversight (people must be able to review and intervene before consequences attach).

An AI sales tool that emits unattributed summaries makes these obligations hard to satisfy: there is nothing to trace and little to meaningfully oversee. A provenance architecture maps onto them naturally — the FOUND/ASSUMED/CALCULATED chain is a per-claim audit trail, and a human confirmation step before any output executes is oversight by design rather than by policy document. Cosa is built in Germany and GDPR-aligned, and this alignment is architectural, not bolted on.

To be precise about the claim: Cosa is not an EU AI Act compliance tool, and using it does not make an organization compliant. Its provenance system aligns with the Act's transparency and oversight principles — which means a sales org adopting it is building on architecture that points the same direction as the regulation, instead of one it will have to explain away. The full picture for sales leaders is in the EU AI Act for sales teams.

How to evaluate any sales intelligence tool for provenance

Whether or not you evaluate Cosa, the category test is portable. Ask four questions of any tool:

  1. 1.Can I click from a claim to its source? Not to a call recording in general — to the line that grounds this specific claim.
  2. 2.Does it distinguish observation from inference? If everything is presented with equal confidence, inference is being laundered as fact.
  3. 3.Does it show its arithmetic? Any derived number — ROI, payback, pipeline value — should expose its formula and inputs.
  4. 4.Does it admit absence? A trustworthy system says "not mentioned" when the evidence does not exist, instead of filling the gap.

Tools that pass all four are doing provenance, whatever they call it. Tools that fail them are asking for trust they cannot earn — and in front of a CFO or an auditor, unearned trust is precisely what sales data can no longer afford. To see per-claim provenance on your own deal data, start with getting started with Cosa or review pricing.