Risk Automations
The agent layer: Extractor, Verifier, Triage, and Analyst — four agents that read, corroborate, prioritize, and reason over vendor evidence, with the Analyst endpoint as the interactive portfolio Q&A interface.
What it is
Risk Automations — in detail.
Risk Automations exposes the four agents that run the TrustVendor evidence pipeline as a first-class product surface. The Extractor agent reads trust centres, subprocessor pages, SOC 2 reports, and incident disclosures and emits structured claims grounded to exact character spans — validated by a Go post-processor before any claim reaches the database. The Verifier agent corroborates high-impact claims across independent sources before they move a score or fire a critical alert. The Triage agent scores materiality per relationship: the same subprocessor addition is urgent for a PHI vendor and informational for a telemetry-only tool. The Analyst agent answers portfolio Q&A by calling a bounded, read-only tool set over the evidence graph — every factual statement it returns carries a citation chip linked to the source span.
The model cost ladder keeps inference spend rational. Tier-0 is deterministic Go parsers for structured feeds — about 65% of extraction volume at zero cost. Tier-1 small models handle classification and deduplication. Tier-2 mid models handle trust-centre extraction and diff interpretation. Tier-3 frontier models handle SOC 2 exception analysis and Analyst Q&A — about 1% of volume, running only when the payload warrants it. The routing decision is a pure Go function made before any model call. This architecture is what makes TrustVendor auditable: score movements are always traceable to a specific agent version, a specific artifact, and a specific model tier.
All four agents are prompt-versioned. Every claim records the prompt file, model, and schema version that produced it. When a prompt changes, you can identify exactly which claims need reprocessing. A precision regression is a rollback — you cannot silently corrupt the corpus because every claim traces back to the version that created it. The Analyst agent's system prompt forbids it from asserting a fact it cannot cite. When tool results are incomplete, the response envelope carries a warning and the answer reflects the gap honestly — it is never silently filtered.
How it works
Three steps,
fully auditable.
The Extractor agent emits claims with mandatory span metadata: character start and end offsets plus a quote hash. A Go validator checks the hash against the stored artifact before any claim reaches the database — hallucinated span locations produce ErrSpanMismatch and go to the dead-letter queue, never the graph. Track ErrSpanMismatch rate per extractor version as your hallucination signal.
The Verifier agent checks high-impact claims against independent sources and marks each as corroborated, uncorroborated, or contradicted. Uncorroborated critical claims still surface — flagged "single source, unverified" — never suppressed. The Triage agent then scores materiality per relationship and hands an ordering hint to the deterministic scorer.
The Analyst agent answers portfolio questions by calling a bounded, read-only tool set: searchVendors, getVendorProfile, listRelationships, queryClaims, getEvidence, computeExposure. It never generates facts — it retrieves them and cites the artifact and span for each. Every session is logged and replayable from the raw tool call trace.
What you get
Built for teams that
have to prove things.
Portfolio Q&A with citation chips
Ask complex cross-vendor questions in plain English. The Analyst agent decomposes the query into tool calls, retrieves results from the evidence database, and returns an answer where every vendor finding links to the claim and span that backs it. The Analyst endpoint is also accessible via the Vendor Pulse API.
Span-validated extraction
The Extractor agent's span validation runs in Go before any claim reaches the database. ErrSpanMismatch is your hallucination detector — it costs nothing to run and produces a per-version quality metric you can gate CI on.
Prompt-versioned, replayable
Every claim records extractor_version = prompt_file@model@schema_version. Change a prompt, identify affected claims, re-run exactly those. No silent corpus corruption. Replayable from the raw artifacts indefinitely.
What the agents will never do
Agent guarantees,
in plain terms.
Agents produce claims, never numbers. Scoring is deterministic Go arithmetic over those claims — models never run in the scoring path.
Every claim cites a hashed, immutable snapshot span. No claim is written to the database without an artifact_id and a validated quote hash.
Extractor agents are prompt-versioned. Every claim records the prompt file, model, and schema version that produced it. A regression is a rollback, not a data loss.
The Analyst agent's system prompt forbids it from asserting a fact it cannot cite. Warnings surface in the response envelope when tool results are incomplete — the answer is never silently filtered.
Sample
What it looks
like in practice.
Common questions
Answered plainly.
What makes Risk Automations different from a general AI assistant pointed at vendor documents?
Can the Analyst agent access data from vendors outside my portfolio?
How do I know the Analyst agent is not making up facts?
Can Risk Automations generate vendor assessment narratives for auditors?
Which model tier runs which extraction?
See it live
Risk Automations
on your vendors.
Book a 30-minute demo. We’ll run it live on vendors from your register — not demo data.