Where an AI policy becomes a decision

A company can write that an AI agent must not disclose sensitive customer information. That statement matters, but it does not yet tell a system what to do when an agent asks to send a particular file to a particular tool.

At that moment, the policy has to become a decision. The action may be allowed. It may require a person to review it. It may be blocked. Access granted earlier may need to be revoked. Whatever happens, the organization should be able to explain which rule applied, who owned the exception and what evidence remains.

That translation—from written intent to a decision inside real work—is the focus of this week’s Weekly Job Intelligence. The series examines the consequential work organizations are defining as AI changes their products, operations and responsibilities. The evidence this week points to governance becoming part of product and workflow design, but not to a broad hiring boom or a solved technical problem.

Two controls at different boundaries

On September 29, OneTrust introduced CORIE (opens in a new tab), which it describes as a shared intelligence layer for AI-agent governance. The stated design converts organizational policy and context into runtime decisions about tool calls: allow an action, block it, or send it for human review. It also creates a separate evidence record.

That is a useful formulation of the problem. A policy document is too general to decide every case by itself. A runtime control needs relevant context: which agent is acting, which resource it wants, what information is involved and which conditions require a person to intervene.

OneTrust’s announcement describes intended product behavior. It does not independently establish customer deployment, measured risk reduction or consistent control across every system an agent may reach.

Okta’s Resource Access Certifications for AI Agents (opens in a new tab) operate at another boundary. Announced as generally available on September 30, the capability is designed to expose which resources an agent can access, assign an owner, route that access for review, revoke it when necessary and record the review decision and context.

The distinction matters. Reviewing whether an agent should retain access to a customer system is not the same as deciding whether one file transfer should occur. Both are governance decisions, but they happen at different times and answer different questions. Access review, action control and evidence need to connect without being treated as interchangeable.

These are vendor accounts of their own products, not independent outcome studies. Still, their specificity helps reveal the work underneath the word governance: translating policy, defining decision conditions, assigning responsibility, managing exceptions and preserving evidence.

The same work is appearing inside product roles

Two current product mandates reinforce that picture from another direction.

G2’s live Principal Product Manager, Content (opens in a new tab) posting combines content and research evidence with provenance, rights and approvals, human control, discovery, testing and commercial outcomes. The role is not framed as a policy office separated from product delivery. The controls are part of the product system.

Talkdesk’s live Forward Deployed AI Product Manager (opens in a new tab) posting combines observation of customer workflows with scope review, specification ownership, user acceptance testing, compliance gates, observability and value handoff. It places the product manager close to both the work as performed and the point where an AI implementation becomes operational.

Those postings do not establish a market-wide hiring trend. G2’s August 18 posting and Talkdesk’s September 23 posting both predate this reporting week. They are useful as current examples of how two employers have assembled responsibilities, not as a count of growing demand.

The broader labor data argues for restraint. The September employment report (opens in a new tab) showed U.S. payroll employment increasing by 29,000, unemployment at 4.2 percent and little change across major industries. July and August gains were revised down by a combined 60,000. The August Job Openings and Labor Turnover Survey (opens in a new tab) reported 7.1 million openings, 5.2 million hires and 5.1 million separations, all broadly little changed.

Those national figures cannot isolate senior product or consulting demand. They can, however, keep a few specialized roles in proportion. The signal is not that hiring has suddenly accelerated. It is that some organizations are defining senior work around the boundary between policy and action even in a restrained labor market.

Trace one decision all the way through

The practical test is smaller than an enterprise governance program. Choose one consequential action an AI agent may take: send a customer file, update an account, approve a refund or publish a response. Then trace five things.

1. Source policy and context. Which written rule, operating constraint or customer commitment governs the action? What facts about this case matter?
2. Decision conditions. What makes the action allowable, reviewable or blocked? These conditions need to be specific enough for the system and the reviewer to use.
3. Named owner. Who owns the resource or decision? If human review is required, who is responsible for making it rather than merely receiving a notification?
4. Exception and revocation path. What happens when the ordinary rule does not fit, or when access already granted should be removed?
5. Evidence record. What will show which policy and context were used, who decided, what action followed and whether anything changed afterward?

Run the trace twice: once for an ordinary case and once for an exception. In the ordinary case, the system should be able to explain why the action proceeded. In the exception, the organization should be able to show where automated handling stops, who takes responsibility and what permits the work to continue.

This is not a claim that five fields produce universal control. A rule can be incomplete. Context can be wrong. A connected service can behave differently from the initiating system. A person can approve an action without enough information. The exercise is valuable because it makes those boundaries visible before the organization relies on a policy statement as proof of control.

The strongest evidence is not that a policy exists. It is that the team can inspect one consequential decision from source rule to recorded outcome and explain where human judgment enters.

If you chose one AI-assisted action in your organization, could you name the rule, the decision owner, the exception path and the evidence left behind? Which of those is hardest to make concrete?