Take a week and read the actual product copy on every vendor pitching an "AI compliance agent" or "agentic SOAR" in 2026. The marketing is loud. The fine print is consistent. The agent suggests; the human approves.
This is the verified state of the market. Not the headlines. The product copy.
Under Trustworthy Autonomy™, the analysis-versus-action split lines up cleanly with the autonomy ladder from earlier in this series. Discovery, enrichment, summarization, drafting — autonomous everywhere. The agent runs through the cloud account, the policy library, the scanner stream, the questionnaire response. Taking the action that carries liability — human-gated everywhere.
Why the line is drawn here
The split is not arbitrary, and it is not a temporary state. It is what a responsible vendor ships when there is no published methodology for certifying an agent's wrong-attestation rate, no independent benchmark to score against, no runtime monitor whose detection recall has been measured, and no provenance substrate that lets an auditor reconstruct any single action.
Without those four things, the human approval gate is the only meaningful protection against an agent's wrong call becoming a real liability event. The market has converged on holding that gate because it is the responsible posture — not because models are not capable.
Every credible vendor in this category has settled on a version of the same answer. The agent does the analysis; the human does the action. The agents that claim more either gate "more" on different language ("approve once, runs forever") or are not in production at scale.
The verified state of the market
Here are the dated, fact-checked vendor positions. Every claim below cites the vendor's own product copy or press release at the date shown; none of it is inferred or reconstructed.
This article will land its main argument on the verbatim, dated product copy from the leading vendors in the category. Per the editorial rule for this series, no competitor claim appears here that was not fact-checked and provided directly. The structure is in place; the claims drop in.
The pattern that emerges
Once the claims are laid out, the pattern is consistent across the vendors. The agent's autonomy is broad on the upstream work — reading the environment, drafting the answer, framing the choice — and narrows to zero on the consequential action. Every vendor in the category draws the line in approximately the same place, on the same side of the same action types.
That convergence is information. It tells you that the gap between marketed autonomy and shipped autonomy is the same gap, in the same place, across the market. Nobody has crossed it yet because nobody has the four prerequisites: published methodology, independent benchmark, measured runtime monitor, provenance substrate.
Where Trustworthy Autonomy goes past it
The point of this series is not to disparage suggest-only. It is the right posture for the conditions today. The point is that the conditions are themselves measurable, and the methodology that measures them — the framework, the benchmark, the runtime monitor, the action chain — can be published and validated.
When that work clears, the action categories that earn promotion move past suggest-only on the evidence, not on the marketing. That is the substantive distinction. Trustworthy Autonomy is the methodology that lets a vendor move past the industry line responsibly, and lets a buyer ask for the evidence that the move was earned.
Until that evidence exists, the buyer's posture should be the same as the responsible vendor's posture: the agent suggests, the human approves. The next article in the series walks through what changes for a lean compliance team when, on a specific action category, that ceases to be the right posture — what scoped autonomy looks like in a working week.
Where this matters next
How often is your compliance AI actually right? — the pillar. The four prerequisites listed above are the technical reason every vendor is at suggest-only; this is the methodology that produces the evidence to clear them.
The trust ladder: from approve-everything to goal-oriented — what the alternative looks like once a category earns promotion. The same agent can be Tier 1 on the suggest-only side and Tier 2 on a different category, evidence-by-evidence.
Launch: Trustworthy Autonomy — the Trustworthy Autonomy launch release covering how the trust ladder frames "suggest-only" as tier 1 of 4.
APRI: AI-Powered Risk Intelligence — the AI diligence-analyst surface built to move past suggest-only via provenance + measured-action logging.