Consent Conflict Reviewer: An AI Agent That Explains Consent Conflicts

An AI agent for consent conflicts explains why consent and suppression records disagree; it does not settle them. The Consent Conflict Reviewer helps DPOs and compliance reviewers on any customer data platform (CDP), at L2, L4 for hold only: it proposes fixes and can act alone just to raise a hold. A model never decides who may be contacted. Rules enforce, and this agent can never release a hold or change consent state.

Rail level: L2, L4 for hold only · Risk: high · Function: lifecycle marketing

By Art Kukharenko · Published 2026-11-02 · Updated 2026-11-02

All AI agent rails

Read it as one rail in the AI agent rails method.

Rail at a glance

The Consent Conflict Reviewer's rail separates enforcing from explaining. A send-time decision service applies consent rules with no model involved, the tool allow-list holds one action, and only the approval service can record a clearance.

Field	Value
Job	Explain consent and suppression conflicts, propose fixes, and raise a hold where written criteria match.
Trigger	A rule-check result or a reconciliation report.
Data scope	Conflict categories, counts, purpose and channel definitions, and pseudonymised samples only where a case needs them. No free text.
Tool scope	Read: conflict reports, purpose and channel definitions, rule text. Propose: a plain explanation and proposed fix. Execute: raise a hold on a launch.
Output contract	Conflict, plain explanation, rule reference, proposed fix, and any hold raised.
Gate	A compliance reviewer clears holds and accepts fixes.
Limits	A cap on holds per day. No consent or suppression write tools exist for this agent.
Ledger	Rule version, conflict ID, hold raised, who cleared it and why.
Evaluation	The consent test cases plus seeded conflicts.
Owner and kill switch	DPO or compliance lead. Disabling the hold action leaves rule-based enforcement unchanged.

What it does and who uses it

The Consent Conflict Reviewer answers the question a reconciliation report leaves open: why do two systems disagree, and what should a reviewer look at first? Compliance reviewers read its explanations, and campaign owners see any hold.

An illustrative case: a nightly reconciliation finds 312 people in a renewals SMS audience with an objection to direct marketing in the consent ledger, yet still active on the gateway's list. The send-time rule would deny them anyway. The agent explains that an objection reaches all direct marketing, and that the pattern fits a propagation failure after a gateway outage. It proposes re-running propagation and, because a written criterion matches, holds the launch. A compliance reviewer checks a pseudonymised sample, accepts the fix and clears the hold with a reason.

The written rule already has an answer for most conflict types, as the table shows, following the suppression architecture. The agent's contribution is naming the type.

Conflict	What the rule does	What the agent adds
Objection in the ledger, destination list still active	Denies at send time	Points to a likely propagation failure
Old file re-imports a suppressed address	Blocks reactivation	Names the import source
A merge joins profiles with different consent	Keeps suppression on the identifier	Explains which identifier carries it
One-newsletter unsubscribe read as all marketing, or the reverse	Applies the recorded scope	Sets out both scopes for a reviewer
Opt-in after withdrawal with no fresh evidence	Stays suppressed until a new affirmative action	Flags the missing evidence

Why this level

The Consent Conflict Reviewer's two levels reflect an asymmetry: a wrong hold costs a delay, while a wrong contact cannot be taken back and may break the law. So the only thing it does alone is make a launch wait.

Consent records are evidence, not model output. Article 7(1) of the General Data Protection Regulation (GDPR) requires a controller to be able to demonstrate consent, and Article 7(3) lets a person withdraw it at any time 1. Under Article 21(3), personal data may no longer be processed for direct marketing once someone objects 2. Applying those rules belongs in deterministic code.

Two of the properties in Meta's Agents Rule of Two are present here 4. The session holds private data, because pseudonymised records stay personal while they can be attributed to someone 3, and its one state change is the hold. Untrusted input must stay out, so free text does too. If a free-text field were ever added, the hold tool would have to go.

L4 usually comes with auto-revert, but reverting a hold means releasing it, and a model may never do that. A timeout changes nothing either. No promotion path adds release rights or write tools.

Three ways it fails, and the control for each

The Consent Conflict Reviewer's real risk is people or systems treating its words as a decision, so each control keeps enforcement out of its hands.

Failure	What it looks like	Control
Recency read as consent	A re-imported file looks newer than a withdrawal, so the model proposes that the latest record should win.	Rule text says suppression wins, and proposals have no write path.
Hold pressure	Too many holds in a launch week push teams to lean on reviewers for fast clearance.	A daily cap, after which matches alert the owner. Each clearance needs a reason, and no bulk-clear tool exists.
Silence read as clearance	Nobody looks further because no hold was raised, or an explanation gets treated as legal sign-off.	Rule results appear separately from the agent's text, which cites rules and never calls anything lawful.

How to evaluate it

The Consent Conflict Reviewer is tested first against the thirteen cases in the suppression architecture article, such as a re-imported CSV or an objection raised through a support ticket. Add seeded conflicts from your own reconciliation history, then measure:

Hold precision. Holds a reviewer upholds, as a share of all holds raised.
Caught before launch versus blocked at send. A conflict blocked at send shows the rule worked and the review came late.
Refusal tests. Have the agent try to clear a hold and call a consent write tool. Both must fail at the tool layer.
Kill-switch drill. Disable the hold action, then confirm that a suppressed test identifier is still denied at send.

The ledger names who cleared each hold for audit purposes, and that record should not become a league table of reviewers.

Where it runs

The Consent Conflict Reviewer runs only on a private model in the customer's own environment, and this data never goes to an external one. The boundary architecture patterns cover the setup. PrivyCDP's immutable audit ledger names the human or system actor behind each AI-assisted action, but the hold rail is a method and not a shipped feature.

Consent and suppression checks before send: the use case this agent supports.
AI agents for lifecycle and CRM marketing: where consent review sits.
Rail card template: a card for your hold criteria.

The wider method is set out at AI agent rails. To agree your hold criteria with your DPO, start from the rail card template.

Browse use cases Function playbooks Agent roster

FAQ

Can an AI agent decide whether someone has consented?

No. Consent state comes from recorded events, and deterministic rules apply it at send time. The model explains disagreements between records and has no vote on whether any individual may be contacted.

Why can it raise a hold but not release one?

Raising a hold only makes a launch wait, and a person can undo that quickly. Releasing one could allow contact that should not happen, so release stays with a compliance reviewer.

What happens if the agent is switched off?

Rule-based enforcement carries on. Suppressed people are still denied at send time, and reports still reach compliance, without the plain-language explanation.

Sources

This article is general information, not legal advice.