The agent did the work. Can you show what it did?

On 15 October 2026, Help Net Security reported that Optro had expanded its portfolio of autonomous AI agents for enterprise risk teams. The headline claim was specific: every agent action, decision, model entry and evidence trail is logged and cited, so the platform can support audit readiness rather than merely claim it (Help Net Security, 15 October 2026). Happy Wang, Optro's Chief Product and Technology Officer, told the outlet the company built governance and automation "into one platform with intention" and that it does not "just claim quality" but tests agents before release and monitors them once they run in a customer's own environment.

The detail worth sitting with is not the agent count. It is the design decision to log the decision, not just the output.

Why this is not an isolated announcement

Three independent signals point the same way.

First, the named customer numbers in the same report. Reddit told Optro it automated 175 controls in its first three months, covering over 40% of its SOX scope, cutting testing time per control by roughly 60% on average, and later expanding automated coverage past 75% of that scope (Help Net Security, 15 October 2026). Carmen Lam, VP of Internal Audit and Compliance at Klaviyo, said testing now runs continuously through the year instead of at a single point in time, and that the audit trail shows how each conclusion was reached. Matt Sewell, Sr. Director and Head of Internal Audit at FJ Management, described the requirement in exactly those terms: an "auditable, human-in-the-loop standard." Three separate compliance functions, independently, converged on the same ask.

Second, the gap this is answering. Reco's State of Agent Security 2026 (26 August 2026), built from enterprise telemetry and analysis of 500 public Model Context Protocol servers, found 80% of AI tools running with no oversight at all, and 414 unsanctioned AI tools per 1,000 employees in smaller organisations. Agentic tooling is spreading faster than anyone is writing down what it touches.

Third, the regulatory direction. The EU AI Act's record-keeping obligations assume an organisation can state what a system did and on what evidence. NIS2 makes cybersecurity risk management a personal accountability for management, not a policy exercise. Both assume a trail exists. Neither assumes you built one on time.

What it means for a regulated enterprise

An audit function that adopts agentic tooling without a logged decision trail has not saved time. It has changed who is making judgement calls without changing who is accountable for them. If an agent assigns a risk score, clears a control as satisfied, or tells a vendor their onboarding evidence is incomplete, a regulator or an external auditor will eventually ask three questions: what did the agent decide, what evidence did it use, and who signed off on it being correct. If the answer to any of those is "we'd have to reconstruct that," the control itself is the finding.

This is sharper for SOX and SOC 2 testing specifically, because both regimes rely on sampling and attestation. Optro's own framing, through Sewell's phrase, is instructive: agentic systems can do the testing, but a human still owns the attestation. That line between "the agent gathered and scored the evidence" and "a named person stood behind the conclusion" is exactly the line most agentic rollouts blur first, because it is the line that is easiest to automate away under deadline pressure.

What actually addresses it

What matters here is not the agent but the append-only decision log sitting underneath it, paired with a hard boundary on which steps require a human sign-off before they become a record. Three properties separate a usable version of this from a dashboard:

• The log records the decision, not just the output. "Control marked satisfied" is not evidence. "Control marked satisfied, citing document X, against policy clause Y, scored by model Z at timestamp T" is.
• The record is tamper evident. If an entry can be quietly edited after the fact, the log is not evidence, it is a claim.
• Human attestation is a separate, visible step from automated evidence gathering. The agent can propose a risk score or flag a satisfied control. A named person has to be the one who accepts it, and that acceptance has to be its own logged event, not inferred from silence.

The same requirement applies to any agent making a decision that a regulator, an auditor, or a court might later ask about, not only agents in GRC software.

What to check on Monday

Before buying anything, an enterprise can answer four questions about agents already running in its risk and compliance functions:

1. For every agent touching audit, risk scoring, vendor due diligence or AI governance work, can you produce the evidence it cited for its last ten decisions, not just the conclusions? 2. Is there a visible, separately logged step where a named person accepts or overrides each consequential output, or does approval happen by default when nobody objects? 3. If an entry in that log were altered after the fact, would anyone know? 4. Does the answer to question 1 take minutes, or does it require someone to go and reconstruct it from memory and chat history?

An organisation that cannot answer the first question quickly is governing its agents retrospectively, if at all.