Five questions to answer before turning on Meta's new AI features

Jason Garrett works with healthcare and wellness brands on paid acquisition and privacy-compliant tracking. Connect with him on LinkedIn.

If you run paid acquisition for a healthcare or wellness brand, two new AI features arrived in your Meta Ads Manager this spring. Most teams I work with did not opt into either. Both are running.

The first is Manus, Meta's in-platform AI agent. It has been in Ads Manager since February and is available to every advertiser through the Tools menu. The second is the AI-powered Meta Pixel enrichment, announced on April 15 and active by default on every eligible account that did not opt out before May 15.

The question I get asked most often right now is whether to turn either of them on. The honest answer is that the question itself is too binary. The right configuration depends on five things, in order, and the answers determine which features you can use, which you should leave off, and what infrastructure you need underneath either.

This piece walks through the five questions. The order matters. The configuration that comes out the other end is what your team should be running this quarter.

Two pieces of context before the questions

Meta's posture on advertiser responsibility is not new. Meta has consistently argued, including in court, that the advertiser is responsible for ensuring that no sensitive data flows through their tracking implementation, regardless of what Meta's systems can or cannot detect. The April 15 announcement restated this in the same language: advertisers are responsible for complying with the Meta Business Tools terms when using the Pixel, which includes ensuring that sensitive data is not shared with Meta. Whatever configuration decisions you make, the legal exposure sits with the advertiser.

The AI Pixel enrichment is opt-out, not opt-in. Existing pixel users received a 30-day notification in Events Manager before the feature switched on by default. The notification was in-platform. Most teams I check did not see it in time, which means the feature is active on their accounts and the opt-out has to be done manually.

With that, the questions.

Question 1 — Are you a HIPAA-covered entity or business associate?

This is the threshold question. It determines whether HIPAA applies to your tracking practices at the federal level. A covered entity is a healthcare provider that transmits health information electronically, a health plan, or a healthcare clearinghouse. A business associate is a vendor that handles PHI on behalf of a covered entity.

If yes to either, both AI features are off. Meta will not sign a Business Associate Agreement with you. The data layer the AI features touch is the same data layer your standard tracking touches, which means everything Meta's AI does on your account inherits the same compliance problem your standard tracking already has. The configuration is server-side Conversions API through a HIPAA-compliant intermediary with payload cleansing, manual AI Pixel opt-out in Events Manager, and Manus restricted to aggregated competitive research that does not touch your account's data layer.

If no, continue to question two. HIPAA may not apply. Other regimes might.

Question 2 — Does your traffic touch patient portals or authenticated care environments?

Even if you are not a covered entity, your traffic patterns can create exposure. A wellness app that integrates with a third-party telehealth provider. A referral funnel that sends users to a provider's authenticated booking system. A content site with a patient portal entry point in the main navigation. Each one of these creates a touchpoint where the user transitions from a marketing context into a care context, and the pixel can fire on either side of the transition.

The AI Pixel enrichment reads the page at the moment the event fires. Whatever the page renders — provider directories, intake form answers, appointment booking interfaces, condition descriptions, treatment recommendations — can be extracted and attached to the event data sent to Meta. The exact technical mechanism is one Meta has not fully described in public communications. Research on the feature has consistently characterized it as the AI extracting content from the rendered page rather than reading a developer-controlled data layer. Whatever the underlying implementation, the operational reach is the same: content that used to be a UX decision is now signal.

If your pixel fires on pages that display care-context content, the posture mirrors a covered entity even though HIPAA may not formally apply. AI features off. Server-side architecture mandatory. DOM audit ongoing. The legal exposure may come from the FTC's Health Breach Notification Rule, state laws like Washington's MHMDA, or California's CMIA expansion — but the operational fix is the same.

If no, continue to question three.

Question 3 — Does your product implicate a health condition by purchase?

The condition-adjacency question. The test: if a user purchases your product, does the purchase imply they have a specific health condition or concern?

Yes in the obvious cases. A telehealth subscription for anxiety treatment. A continuous glucose monitor for diabetes management. An erectile dysfunction medication. A weight loss program with BMI eligibility requirements.

Yes in the less obvious cases. A sleep supplement marketed for insomnia. A mental wellness app whose category is depression management. An acne treatment. An ergonomic chair sold for chronic back pain. A fertility tracker positioned for couples trying to conceive.

Meta's classification system uses behavioral inference, not just keyword matching. Research on classified accounts shows that brands whose product description, ad creative, or landing page content implies a health condition can be classified into Health and Wellness regardless of whether they use clinical language anywhere on the site. The condition is the implied use, not the explicit label.

If yes, AI Pixel enrichment is off. The one-click Conversions API is off. Manus is restricted to upper-funnel work and aggregated analysis. The architecture is a server-side container or compliance-first customer data platform that filters payloads before forwarding to Meta. The DOM audit is mandatory, with priority on the post-purchase flow, condition landing pages, and any quiz or assessment that displays results.

If no, continue to question four.

Question 4 — Can you maintain a server-side cleansing layer between your site and Meta?

The infrastructure question. The configuration moves in the first three questions assumed an architectural capacity that not every team has. Server-side Conversions API through your own infrastructure or a compliance-first CDP that filters payloads before forwarding to Meta. The cleansing layer is what makes any upper-funnel use of Manus defensible. It is what makes the AI Pixel enrichment containable. It is the reason the one-click CAPI option Meta launched on April 15 is not the right answer for a regulated account — the convenience layer Meta provides removes exactly the surface where cleansing happens.

The 17.8 percent lower cost-per-result figure Meta is citing in the announcement is real and worth understanding correctly. It compares advertisers running Conversions API to those without. It is not a measurement of the AI enrichment's lift on its own. The reason Meta is anchoring the rollout on it is to drive CAPI adoption broadly — the one-click setup is the vehicle, the 17.8 percent is the pitch. For an ecommerce brand, the calculation is straightforward. For a regulated account, it is the wrong calculation entirely.

If yes to the infrastructure question, the framework can produce a nuanced configuration. Some AI features on, some off, with the cleansing layer doing the heavy lifting underneath.

If no, the configuration collapses to the conservative posture. Without the cleansing layer, you cannot reliably control what reaches Meta, which means you cannot reliably opt into any AI feature that depends on richer signal. AI Pixel enrichment off. One-click CAPI off. Manus restricted to analysis-only mode, with no campaign creation or audience building through the agent.

Question 5 — Does your team have the capacity to review AI outputs before they ship?

The operational question. Manus generates audiences, campaign structures, and reports faster than most teams can validate them. The human review gate is the procedural fix — but it only works if the team has the capacity to maintain it.

The capacity question is honest. A two-person agency managing twenty accounts does not have it. An in-house team of one running paid social for an early-stage company does not. A larger team with explicit review responsibilities does.

If yes, Manus is usable for the capabilities the prior questions did not rule out. With the gate in place, audience research and campaign creation can be done with the agent contributing rather than deciding.

If no, Manus stays in analysis-only mode, with outputs treated as research inputs rather than deliverables. The capacity question matters more than the capability question.

The two opaque Meta classification systems running on your account

The five questions assume something Meta's product design does not deliver: that you can know which configuration you are operating under at any given moment.

Two Meta classification systems are running on your account simultaneously.

The first is the data source category. Meta classifies your pixel into a category — Health and Wellness most relevantly for this audience — at some point of Meta's choosing, based on signals Meta does not publish, with consequences that appear in Events Manager after the fact. Categorization can move from no restriction to core to mid to full classification silently. Most teams I check have not been classified yet, which means they are operating without the protections that classification provides.

The second is the AI Pixel enrichment status. Either Meta has enabled it on your account or Meta has excluded it based on a classification you cannot verify from outside Events Manager. The exclusion criteria are vague. Meta's announcement says only that access may not be available to advertisers that have certain data source categories, without naming which categories qualify. Research suggests that Special Ad Categories including health are the likely excluded set, but Meta has not published the list and has not committed to one.

The advertiser is now monitoring both. Both decisions are made by Meta. Both are opaque. Both can change without advertiser-facing notice. Both affect what data flows where, with no audit trail visible from the account.

The unanswered question that follows from this is the one healthtech marketers should be asking and are not.

What happens if you are running with the AI Pixel enrichment enabled, and Meta later classifies your data source into Health and Wellness?

The possible answers, none of which Meta has confirmed publicly: the enrichment auto-disables on classification, and the historical event data Meta has already collected from your pre-classification pages stays in Meta's systems. The enrichment continues running because classification happened after enablement, and stays on until you manually disable it. Some categories of enrichment shut off and others continue, on a basis Meta has not published. You receive a notification through the same channel Meta uses for restriction tier changes — to the Business Manager email, days or weeks after the change has taken effect. You receive no notification at all.

Meta has not addressed the transition in the April 15 announcement, in the Business Help Center pages on Special Ad Categories, or in any community-facing communication I have found. The rollout is six weeks old. The classification system is itself silent. The answer remains unanswered.

The operational implication is that the responsible posture is to assume the least favorable version. If the answer turns out to be benign, you lose nothing by having disabled explicitly. If the answer turns out to be that Meta retains scraped page content from your pre-classification window and uses it to seed audiences afterward, you want to have been disabled the entire time.

Putting the answers together

The five questions produce a configuration, not a single yes or no. A few of the patterns I see most often:

A covered entity that touches no portals, sells a condition-implicating product, has a cleansing layer in place, and runs a team with review capacity. AI Pixel off, one-click CAPI off, Manus in analysis-only mode, server-side CAPI through the cleansing layer, DOM audit ongoing.

A health-adjacent direct-to-consumer brand whose product implies a condition by purchase, with a cleansing layer in place and review capacity. AI Pixel off, one-click CAPI off, Manus restricted to upper-funnel work, DOM audit on post-purchase and assessment pages, ad creative discipline aligned to outcomes rather than conditions.

A general wellness brand with no portal touch, no condition-implicating product, a cleansing layer in place, and review capacity. AI Pixel optionally on after a DOM audit confirms no condition-adjacent rendering, server-side CAPI through the cleansing layer (not one-click), Manus usable across all four capabilities with the review gate, DOM audit quarterly.

A B2B healthtech vendor selling to providers, payers, or enterprise health buyers, with no portal touch and no condition-implicating product. AI Pixel on, one-click CAPI defensible, Manus usable, DOM audit annual.

The framework is not a checklist. It is a synthesis. The work for the team is to answer the five questions honestly, then build the configuration from the answers — rather than starting with a posture and looking for justifications.

What changes in late 2026

Meta has been clear that fully autonomous campaign execution is the destination for the end of this year. At that point, Manus or its successor will no longer be drafting campaign structures for human review. It will be running them.

The five-question framework is a snapshot. The configuration work over the next two quarters is what determines whether your account remains in the same row of the synthesis when the underlying product ships. The opt-out toggles are still available. The DOM audit cadence is still establishable. The server-side architecture is still buildable. None of it gets easier when the agent moves faster.

The configuration question is not "should we turn it on or off." The configuration question is "which row of the synthesis are we, and what does the architecture underneath look like." The five questions produce the answer. The answer produces the work.

The work is what is in front of every healthtech marketing team for the rest of this year.