SOC 2 + AI-agent certification · full launch early 2027

Show your auditor who approved each agent action.

And let them check it themselves. Clean Opinion holds your agent's risky, customer-impacting actions, gets a second check from a different model, takes a named person's signature on the exact action, and keeps a ledger your CPA firm and your customer's security team can verify without trusting you or us.

Try the 60-second demo · for CTOs, security and compliance leads at AI-agent vendors

Ledger · action A-0417fictional
Proposed by the support agentGrant admin to dana.r in the customer's workspace
Agent (model A)
proposed · signed
Second check (model B)
outside scope · B006
Named approver
not signed
Logged
E015

Mapped to AIUC-1 B006 and E015, both mandatory.

The name

"Clean opinion" is what auditors call an unqualified report: the auditor says, with no reservations, that your controls were described fairly and worked as described. It is the result a SOC 2 report is for. Your auditor decides it. Clean Opinion gives them the evidence for the part your agents touch. More on audit opinions

Demo · fictional data

Your support agent wants to make a user an admin.

You sell an AI support agent. During a customer's security review, their reviewer asks about one action: the agent granted a user admin rights in their workspace. Run it with logging only, then with Clean Opinion. Then switch to the Auditor tab and verify the ledger yourself.

Tallyhand Support Agent · in Juniper Freight's workspace (both fictional)
ready
  1. 1Logging only

    The agent's request: grant admin to dana.r@juniper-freight.example. Without a gate, the agent does it and writes a log line.

  2. 2Turn on Clean Opinion

    Same request. Now it goes through the gate first.

Everything in this demo is fictional: Tallyhand, Juniper Freight, the people, the emails and the workspace. Signatures use real Ed25519 keys generated in your browser, and the digests are real SHA-256. Nothing is sent anywhere.

What changed

Your agent is about to be audited like your software.

Oct 19, 2026 · AICPA

One engagement, two results

The AICPA announced a joint offering: one CPA-led engagement that produces a SOC 2 report and AIUC-1 certification, which it calls "the world's first standard for AI agent safety, security, and reliability." It adds quarterly AIUC-1 technical retesting, both renew every year, several CPA firms already deliver it, and full launch is expected in early 2027.

AICPA release

AIUC-1 · published controls

What is mandatory

  • E015 Log AI system activity: processes, actions and agent outputs. Mandatory
  • B006 Prevent AI agents from acting beyond intended scope and authorized privileges. Mandatory
  • E004 Assign accountability: formal approval of changes to the AI system across its lifecycle, with evidence. Mandatory
  • C007 Flag high-risk outputs for human review. Optional

Accountability · Security · Safety

Oct 28, 2026 · SOC 2 draft

AI in the SOC 2 criteria

An exposure draft revising the SOC 2 Trust Services Criteria adds points of focus on the risks of automated decision-making and the transparency of AI-generated outputs. Comments are open through Jan 15, 2027. If approved, it applies to engagements beginning on or after Dec 31, 2027, with early adoption permitted.

AICPA · Journal of Accountancy

2026 · security questionnaires

What reviewers are starting to ask

Published AI vendor questionnaire templates now include questions like "Can we require human approval per action type?" and "How does your platform ensure audit logs cannot be altered after the fact?"

Reco questionnaire, Q21 · Kognitos RFP template, Q8

To be precise: neither standard requires a person to approve every agent action. AIUC-1 makes human review optional (C007), and E004 is about approving changes to the AI system, not each action. What the mandatory controls ask for is a log of what the agent did and proof it stayed in scope. Your enterprise customers may ask for more. Clean Opinion covers both. Read the sources yourself; this page is not audit or legal advice.

How it works

Keep the agent fast. Make every risky action provable.

Clean Opinion sits between your agent and the systems it changes for your customers: workspace roles, retention settings, bulk ticket actions. Routine actions pass and are logged. Risky ones wait.

01

The agent proposes

The agent signs the exact action with its own key. That key can propose, but it can never approve, so an agent can't sign off its own work.

02

A different model checks

A model from a different vendor reviews the action against the scope you configured. Out-of-scope actions are held, and the hold is recorded against B006.

03

A named person signs

Your named approver sees the payload and its digest, and signs that exact action. Change one field and the signature no longer matches. A customer-named approver is optional.

04

Anyone can verify

Each step goes into a hash-chained, signed ledger mapped to E015 and B006. Your CPA firm and your customer get a free verifier and an export for workpapers.

Who it's for

AI-agent vendors selling to enterprises.

Support and voice agents, agent and RPA platforms, coding agents: if your agent changes things in a customer's systems, someone in their security review will ask who approved it. Clean Opinion gives you an answer they can check, and gives your auditor evidence that is already structured.

  • Different model, not the same one twice. The second check must come from a different vendor than the agent's model, and the ledger records which.
  • Bound to the exact action. A signature covers one payload by digest. Edit it after approval and it is refused.
  • Your approver by default. A named person on your side signs. Add your customer's approver only where a buyer asks for it.
  • Verifiable without us. Public keys travel with the ledger. A CPA firm or a customer verifies it without an account with us.
  • For audit firms. Firms running the joint engagement get the verifier free.

FAQ

Questions CTOs ask us first.

What is a clean opinion?

It is auditor shorthand for an unqualified opinion: the auditor's report says, without reservations, that your controls were described fairly and worked as described. It is the result a SOC 2 report is for. Your auditor decides it; no product can promise it. Clean Opinion supplies evidence for the controls your agents touch.

Does AIUC-1 require a person to approve every agent action?

No. On the published standard, E015 (log AI system activity) and B006 (prevent unauthorized AI agent actions) are mandatory. C007 (flag high-risk outputs for human review) is optional. E004 is mandatory, but it covers formal review or approval of changes to the AI system across its lifecycle, not approval of each action. Enterprise customers may still ask for per-action approvals in their own reviews.

Isn't logging enough?

A log shows what the agent did. It doesn't show that the action was in scope, or who decided it could run. That is the question in step 1 of the demo. Clean Opinion records the scope check and the approval next to the log entry, signed.

Why a different model for the second check?

A second pass by the same model shares the first pass's blind spots. A check by a different vendor's model is a more independent look, and the ledger shows which vendor made it, so an auditor can see the two weren't the same.

Does my customer have to approve my agent's actions?

No. By default the approver is a named person on your side. A customer-named approver is an option for buyers who ask for it, often in regulated industries.

How does an auditor check the ledger without trusting you?

Every entry is signed with Ed25519 and chained by SHA-256 to the one before it. The ledger carries the public keys, so the verifier recomputes every digest and checks every signature on its own. The Auditor tab in the demo does exactly that, in your browser.

Can I use it today?

It's in early access and in development. Clean Opinion is by Revelations Consulting and built on Seconded, our separation-of-duties platform for AI agents. Leave your email and your target audit date, and we'll be in touch.

Have an answer your customer can verify.

Early access is free while we build it. Tell us when your audit is, and we'll be in touch.