Blog · ← All posts
Passing the Enterprise Security Review: The Agent Evidence Pack Your Buyer's Assessor Actually Wants
By the Ingenious AI team · Published: November 14, 2026 · Reviewed: November 14, 2026 · 7 min read
The short answer: Enterprise assessors evaluating an AI feature ask six questions: what can the agent do, what has it done, how do you know the record is complete, how do you know it was not altered, what happens when it misbehaves, and who is accountable for it. The pack that answers all six in one session is: the agent's mandate and lifecycle record, a scoped ledger export with chain verification and the period certificate, the measured coverage figure from downstream reconciliation, the anomaly and incident file with dispositions, and the data-handling note (capture policy, retention, masking). Vendors who hand this over proactively shorten their security review from months to weeks.
Replay a recorded agent incident free → Step through a real recorded run in the replay view, or paste ledger records into the public verifier and get a VERIFIED/BROKEN verdict in seconds — no signup.
Every enterprise deal with an AI feature in it now contains a second sale: the security review, where an assessor who will never use your product decides whether it may be used. Assessors are not hostile; they are pattern-matchers under time pressure, working a questionnaire built from the last decade's vendor incidents. AI agents alarm them for a specific, legitimate reason: an autonomous actor inside the customer's processes, whose behavior the vendor is usually unable to evidence beyond dashboards and assurances.
That inability — not the agent's capability — is what stalls deals. This guide lists the six questions as assessors actually ask them, and the artifact that answers each, assembled as a single pack an account team can hand over before the questionnaire even arrives.
What is an evidence pack? the pre-assembled set of artifacts a vendor provides to demonstrate an AI agent's accountability to an external reviewer: mandate and lifecycle records, tamper-evident action history with verification, measured completeness, incident and anomaly history with dispositions, and data-handling documentation. Its defining quality is that its claims are checkable from the artifacts themselves, rather than resting on the vendor's attestations.
Questions 1–2: what can it do, and what has it done?
The assessor opens with scope: the agent's mandate — tools, data classes, action caps, escalation paths — stated as a document, not a demo. Then the follow-through that separates serious vendors: the lifecycle record showing the mandate's history (when granted, when changed, by whom) and the action history itself, sampled by the assessor against their own test transactions. A vendor whose pilot interactions are already in a verifiable ledger answers question two by handing over the export; a vendor whose answer is a dashboard tour has told the assessor how the next five questions will go.
Pre-answer both in the pack: mandate document plus a sample export from your own production agent (your support agent's real, masked history is more persuasive than any synthetic dataset, precisely because it is slightly messy in the ways real operations are).
Questions 3–4: how do you know it is complete, and unaltered?
These are the discriminating questions, the ones the questionnaire was written to reach. Completeness: the measured coverage figure — downstream reconciliation across payments, mail, and CRM witnesses, with the method described and the current figure stated, imperfections included. Integrity: the hash chain explained in one paragraph, the verification report for the sample period, the period-close certificate, and the standing offer that the assessor's team re-run verification themselves on the exported records.
Most vendors fail here not because their systems are weak but because they have never measured either property. The assessor's report language differs accordingly: 'vendor provided independently verifiable records with measured completeness' versus 'vendor attests that logging is enabled'. Only one of those sentences survives a risk committee.
Question 5: what happens when it misbehaves?
Assessors assume misbehavior; they are pricing your response, not your perfection. The pack's answer is the incident file: the anomaly rules that watch the agent (bursts, first-seen tools, after-hours), the alert-to-pause path with its named authority, and — decisively — one or two real incident records from your own operation, with timelines, bounding, root cause, and the regression case that followed. A vendor with zero incidents to show is claiming either inexperience or concealment, and assessors write down which they suspect.
The first-hour playbook, included as a document, converts the answer from anecdote to system: containment in minutes, evidence freeze, scoped communication. Assessors have seen many postmortems; a response procedure that begins by preserving evidence reads as maturity in a dialect they trust.
Question 6: who is accountable — and the data note
The closing questions are organizational and legal: a named owner for the agent, the lifecycle discipline around it (reviews, retirement), and the data-handling note — what the records contain (masked summaries, digests), what they never contain (raw personal data), retention periods per tier, access control over the ledger, and the erasure procedure. This is where the capture policy pays its commercial dividend: the privacy review and the security review converge on the same document, and a vendor who can hand it over in the first meeting has clearly built the system rather than the slide.
The pack as a sales asset, maintained like one
Assemble the pack once, as a living artifact with an owner and a refresh rhythm — exports regenerated monthly alongside the verification and coverage figures, incident file updated as incidents close. Version it. The commercial return is measurable: review cycles shorten, the questionnaire's agent section closes in the first exchange, and the assessor becomes an internal advocate armed with artifacts their committee can check. In a market where every competitor claims agent capability, the pack is the differentiation that survives contact with a skeptic — because it is designed to.
Five actions you must take today
Build the pack before the next enterprise questionnaire lands:
- Write the mandate document per agent. Tools, caps, data classes, escalation — the scope answer, signed by the agent's named owner.
- Generate a standing sample export. One real month of your flagship agent's masked history, with verification report and certificate attached, refreshed monthly.
- Measure and publish your coverage figure internally. Reconciliation method and current number, ready to state — including the honest residue and its dispositions.
- File two real incidents beautifully. Timeline, bounding, cause, regression case — chosen, documented, and cleared for assessor eyes.
- Finish the data note. Capture policy, tiers, retention, access, erasure — one document serving the security and privacy reviews together.
Worked example: a review that closed in nineteen days
Fictional B2B platform Surveyline sells analytics with an agent that drafts and sends customer-facing summaries. Its first enterprise prospect's security review was quoted at 'eight to twelve weeks, typical for AI features'. Surveyline's account team sent the evidence pack with the proposal instead of waiting for the questionnaire: mandate for the summary agent, a month of its production ledger (11,842 records, verification PASS, certificate attached), coverage measured at 100% across the mail platform and CRM for the trailing quarter, the anomaly rules and the one real incident in the file — a misdirected summary, bounded to a single recipient from the records, remediated with an allowlist the same day — and the two-page data note.
The assessor's agent section closed in the second meeting; their written summary singled out the ability to re-verify the export independently and the incident file's candor. Total elapsed review: nineteen days, with the agent questions — usually the long pole — finishing first. Surveyline now tracks 'pack sent' as a pipeline stage, because its effect on cycle time proved larger than any pricing lever they tested that year.
FAQ
Won't showing real incidents scare assessors off?
The evidence runs the other way. Assessors price concealment risk above incident risk: a vendor with real, well-documented incidents demonstrates detection, bounding, and remediation — the exact controls the review exists to verify. The vendor claiming zero incidents is claiming their detection is perfect, a claim no assessor can write into a risk report with a straight face. Choose your incidents well — bounded, honestly caused, demonstrably fixed — and let them work for you.
How much of our own production data should the sample export contain?
Enough to be checkable: a real period, a real agent, complete for its scope, masked per your capture policy. Assessors test authenticity by sampling against witnesses (their pilot usage, your status page's incidents) and by the texture of the data — real histories have refusals, errors, and boring Tuesdays. A curated highlight reel fails the texture test in the first ten rows.
What if our coverage or verification story has a known gap right now?
Disclose it in the pack with the investigation and remediation state — the same rule as audit bundles. An assessor who finds a vendor-disclosed, in-remediation gap writes 'control environment detects and addresses deficiencies', which is close to the best sentence available in their vocabulary. The fatal version is the gap they find themselves, which converts every other artifact in the pack into a question.
Who should own the pack inside our company?
The agent's platform owner, with the account team as its distributor and a monthly refresh tied to the verification calendar. Ownership matters because packs rot: last quarter's export and a certificate from two closes ago quietly become the 'highlight reel' failure mode. A dated pack, refreshed on rhythm, is itself evidence of the operating discipline it describes.
Does the pack replace certifications like SOC 2 or ISO 27001?
No — it complements them, and assessors treat the two differently. Certifications attest that a management system exists; the pack exhibits the agent-level evidence the certification's scope usually gestures at. Vendors with both move fastest: the certificate opens the door, the pack closes the agent questions. Vendors with neither should build the pack first — its artifacts are the raw material the certification audit will ask for anyway.
Start logging your agents free → The free tier gives you the full hash-chained ledger for 1 agent and 1,000 actions a month, with chain verification and replay included.