Trust by construction
The most-claimed and least-shown property in this category is auditability. We show it. Every sentence in a proposal drafted in BidAngel can be traced backwards to the source document it answers, the asset or claim it draws from, the model run that produced it, and the human who approved it.
This is the answer to “but AI hallucinates” — not a promise about better models, but a set of structural commitments: each cited claim is checked against the source it cites, a section with thin evidence is answered honestly rather than padded, and nothing reaches Submitted without a traceable trail.
The provenance trail, end to end
One horizontal spine runs left-to-right: opportunity discovered, requirements extracted, pursuit opened, draft authored, reviewed and approved, submitted, outcome recorded. The Review gate is the literal advisory-to-approved transition. Below the spine, three layers attach to every stage — the evidence cited into it, the actors attributed to it, and the audit ledger recording it.
- Source documents (RFP + attachments)
- Content assets (your library)
- Capability claims (your assertions, with freshness)
- Model run — version, prompt, timestamp
- Reviewer attestation at every approval
- External agent identity (when present)
- Tenant-scoped audit ledger
- Every tool call, not just writes
- Substrate for incident response
What it means in practice
AI output is advisory until a human approves it
Every draft, every extraction, every requirement match is generated into an advisory state. The Review gate is the only path into approved state. This is a literal state transition, not a UX label — enforced at the data layer, the tool boundary, and the audit ledger.
Every claim cites its evidence
Drafts cite content assets and capability claims by reference, not by paraphrase. Requirements cite the source-document paragraph. When a CFO asks why your proposal claims X, the answer is one query away — including which model version produced the draft, which prompt was used, and who attested the approval.
Cited claims are checked against their source
Citing an asset is not the same as being supported by it. Each cited sentence is checked against the specific asset it cites and marked supported, partially supported, or unsupported — with a one-line rationale — so a reviewer sees which claims the evidence actually backs, not just which paragraphs were referenced. A claim whose cited asset does not substantiate it is surfaced for review, never quietly trusted.
Thin evidence produces an honest answer, not filler
Before a section is drafted, the platform judges whether the approved evidence actually substantiates an answer. When it does not, the agent answers only the part the evidence supports and names the gap — or declines to draft and routes the section to an evidence-gap state — rather than padding with plausible, industry-standard language the evidence never stated. A short, fully-supported answer is treated as better than a long, partly-unsupported one.
Capability claims supersede, never silently rewrite
A capability claim ("we hold ISO 27001") is a first-class assertion with provenance, authority, and freshness. New evidence supersedes — old assertions become superseded, not deleted. Edits to approved content demote it back to advisory until re-approved. Your library cannot drift past review.
Every action is recorded, not just writes
The audit ledger captures every tool call — reads included — for forensic completeness. Tenant-scoped, queryable, and the substrate for our incident-response runbook. This is what makes the trail defensible to a SOC 2 auditor, a procurement officer doing vendor due diligence, or a CFO asking how a claim got into the proposal.
Agents traverse the same governed surface as humans
A first-party drafting agent, an external agent connected over MCP, and a human team member all reach the system through the same typed tool registry. Same tenant isolation (your workspace is kept apart from every other customer’s), same role gating, same approval requirements, same audit trail. There is no AI side door.
How BidAngel’s provenance posture compares
Almost every tool in the category cites a library entry at best. Almost none can show the full trail from source RFP through every model run, every reviewer attestation, and every audit-ledger entry. The difference is structural, not feature-level.
| Tool category | Provenance posture | What this trail exposes |
|---|---|---|
| RFP writers (Loopio, Responsive, Qvidian, AutogenAI, Inventive, Arphie) | Library entries cited at best — model output usually un-attributed | A citation points to a source document at best — never a per-claim check that the cited source actually supports the sentence. No model + asset + claim trail into the draft; confidence indicators limited or absent. |
| Discovery tools (GovWin, GovTribe) | Have source documents — nothing downstream | Trail stops at source. The whole right half (draft → review → submit → outcome) is missing. |
| Generic AI assistants (M365 Copilot, ChatGPT, Claude) | No domain artefacts to trace to or from | No evidence layer. No approval gate. No audit ledger. |
| Custom-built internal proposal stacks | Often have some pieces — never all of them, rarely audited | Disconnected fragments. No single ledger. No single state boundary. |
| BidAngel | Every stage attested, every cited claim checked against its source, every action audited | — |
Trust is the governance side of the boundary
BidAngel owns both the canonical procurement backbone and your tenant’s workflow on top of it, with a clean governed boundary between them. The security page shows the data-isolation side of that boundary. This page shows the governance side: how advisory becomes approved, how AI output is constrained before it reaches a buyer, and how every action is recorded.
Same boundary, two audiences. Same architectural commitment.
Doing vendor due diligence? Our Data Processing Agreement publishes the standard DPA and the authoritative subprocessor list — every third party that processes customer data, what it processes, and the conditions under which it is engaged.