Frequency Advisor: An AI Agent for Send Frequency Recommendations

AI email frequency recommendations work best as advice a person can test, not as a setting a model changes. The Frequency Advisor proposes caps and send spacing from aggregate fatigue signals at L2, the propose level, for lifecycle leads using any customer data platform (CDP). It never changes a cap itself, targets people by vulnerability, overrides a suppression or reads an individual's send history.

Rail level: L2 · Risk: medium · Function: lifecycle marketing

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

All AI agent rails

Here is the rail as one part of the AI agent rails method.

Rail at a glance

The Frequency Advisor's rail is short because the agent holds so little. A cohort data view enforces the minimum group size, and caps live in normal configuration, which sits outside its tool allow-list.

Field	Value
Job	Recommend frequency caps and send spacing from aggregate fatigue signals.
Trigger	A monthly review, or a rise in unsubscribes or complaints.
Data scope	Aggregate send counts by cohort, unsubscribe and complaint rates, and the campaign calendar. No individual histories.
Tool scope	Read: cohort aggregates, calendar, current caps. Propose: cap and spacing changes with expected trade-offs. Execute: none.
Output contract	Recommendation, evidence, expected trade-off, and how to test it.
Gate	The lifecycle lead accepts or rejects, and caps change through normal configuration.
Limits	Minimum cohort size, and a budget for each review.
Ledger	Query references, recommendation, decision, model version.
Evaluation	Back-test on past periods, and held-out cohorts for accepted changes.
Owner and kill switch	Head of lifecycle marketing. Stopping the review schedule leaves existing caps in force.

What it does and who uses it

The Frequency Advisor reads the calendar and the fatigue numbers together, a step that gets skipped when several programmes share one list. Lifecycle leads use it to see the combined load that no single programme owner sees.

An illustrative monthly review at a retailer: customers in their first 30 days receive nine emails a week, against four for everyone else, and their unsubscribe rate is three times higher. The advisor recommends a cap of five a week for that cohort, with 48 hours between promotional sends. The stated trade-off is fewer promotional clicks in month one, against possibly lower list loss. Its test plan holds out one-fifth of the cohort for six weeks. The lifecycle lead accepts the test, and the configuration owner makes the change.

Signals differ in what they can support, as the table shows.

Signal	Can support	Misleads when
Unsubscribe rate by cohort	Comparing cohorts of similar tenure	Tenure differs, since newer subscribers may behave differently whatever the volume
Complaint rate	A hard ceiling: Google asks senders to stay below 0.10% and never reach 0.30% 1	Providers use different denominators, so rates cannot simply be added together
Opens	Very little	Apple Mail Privacy Protection inflates them, so a rise may mean nothing
Clicks and purchases per recipient	The short-term value of an extra send	Future reach lost to unsubscribes stays out of the count

Field evidence argues for caution. In an experiment with a charity, Damgaard and Gravert found that a reminder email raised donations and sharply raised unsubscribes 4. That is one charity and one reminder, not a rule for retail, but it explains why the trade-off field is mandatory.

Capping tools also differ. Braze documents that a capped message is skipped instead of held, and that a campaign can switch capping off (checked 1 November 2026) 3. A recommendation should therefore state which sends it assumes the cap covers.

Why this level

The Frequency Advisor is fixed at L2 because a cap looks reversible when its cost is not. A setting can change back the same day, but someone who unsubscribes is generally gone, and regulators warn against asking them later if they changed their mind. The blast radius is a whole cohort.

Very little of Meta's Agents Rule of Two applies to this session 6. Campaign names in the calendar may be untrusted text, aggregates contain no individual histories, and no tool changes state. A write tool would save minutes and remove the one person weighing the trade-off, so the rail has no promotion path.

A cap can only limit mail that is otherwise allowed, and it can never bring a suppressed person back into an audience.

Three ways it fails, and the control for each

The Frequency Advisor's failures share a theme: the numbers support an answer that the data cannot actually prove.

Failure	What it looks like	Control
Selection effect read as tolerance	Heavy-mail cohorts look healthiest because active customers trigger more mail, so the advice is to send more.	The output marks evidence as observational and must include a test. Any increase goes through a held-out cohort first.
Pressure on people at risk	A cohort that responds strongly to daily offers may be under strain, such as people in debt or with a gambling problem.	The cohort allow-list excludes vulnerability signals and proxies. Article 5(1)(b) of the EU AI Act bans exploiting certain vulnerabilities where significant harm is likely, so take advice 2.
Short-term win hides list loss	Extra sends lift clicks this month while quiet attrition shrinks next year's reach.	The trade-off field is required, and the test window must span more than one campaign cycle.

How to evaluate it

Evaluate the Frequency Advisor against held-out cohorts, because before-and-after comparisons absorb seasons, sales and list changes.

Back-test on past periods. See whether its advice would have come before known fatigue episodes. A back-test shows plausibility, not effect.
Hold out for each accepted change. Fix the metrics and the window first, then compare unsubscribes, complaints and purchases between treated and held-out groups.
Read acceptance rate with outcomes. High acceptance with no measured effect suggests rubber-stamping.
Keep a ceiling in view. Share the spam rate of treated cohorts with the Deliverability Watcher owner.

Sometimes doing nothing is the honest result. If the held-out group behaves the same, leave the caps alone.

Where it runs

The Frequency Advisor's aggregates allow an external model if the contract permits, and a private model suits regulated teams, a choice the boundary architecture patterns help you weigh. PrivyCDP's AI-assisted review of groups and reports applies limits on data access and review of results, and it does not include this advisor as a product.

AI agents for lifecycle and CRM marketing: this agent's playbook and use case.
Deliverability Watcher: the rail that watches provider limits.
Rail card template: a card for your cohort rules.

Other lifecycle profiles are listed at AI agent rails. Download the rail card template to set your own test standard.

Browse use cases Function playbooks Agent roster

FAQ

How often should we send marketing email?

No single number holds across lists. Compare cohorts of similar tenure, test a change against a held-out group, and judge it on unsubscribes and complaints as well as clicks.

Can AI set a frequency for each person?

This rail uses aggregates only. Per-person optimisation needs individual histories and amounts to profiling for direct marketing, which people may object to under Article 21 of the General Data Protection Regulation (GDPR) 5.

Why are opens left out of the evidence?

Apple Mail Privacy Protection stops opens from showing that a person read a message, so a rise can be an artefact. Clicks, purchases, unsubscribes and complaints are closer to what recipients actually did.

Sources

This article is general information, not legal advice.