Meta puts an action-taking support agent inside Facebook and Instagram

An assistant that finishes the task, not one that hands you a link

Most in-product help experiences end where the real work begins: they surface an article, a settings path, or an external page and leave the user to execute. Meta's framing here is a deliberate break from that.

When you have an account issue, you need a solution — not just a suggestion.Montana Labs

The stated capability list is where the frontend ambition shows. The assistant can reset passwords, update profile settings, manage privacy settings, and report scams, impersonation accounts, or problematic content — all directly within Facebook, and Instagram in the future. That is an agent wired into account-mutating actions, not a chatbot bolted onto a help center.

It also folds in the appeals surface: seeing why content was taken down, what appeal options exist, and tracking what happens next. Bundling the reporting and appeal flow into the same assistant that answers questions is a consolidation of several previously scattered UI paths into one conversational entry point.

Latency and placement are the product decisions

Two concrete design commitments do most of the work here. The assistant is built into the apps so help is "always just a tap away," and it "can respond to requests typically in under five seconds."

For a support surface, sub-five-second response is the difference between a tool people use and one they abandon for a web search. Meta explicitly positions this against "traditional help center searches or seeking answers on external websites" — the friction it is trying to remove is the tab-switch to a slow, generic help portal.

The rollout is scoped carefully: globally in countries where Meta AI is already available, across iOS, Android, and desktop Help Center, in all supported languages for support topics. But the highest-stakes case — logging in when you're locked out — is gated to select cases in the US and Canada first. Login recovery is exactly where a helpful assistant and an account-takeover attacker look similar, so a narrow start there is a sensible sequencing choice.

The enforcement claims are less a frontend story than a roadmap

The second half of the announcement covers more advanced AI for content enforcement, and the figures are specific: 5,000 scam attempts a day caught that no review team had found, an over-80% reduction in reports of the most impersonated celebrities, twice the adult sexual solicitation content caught with a 60%+ lower mistake rate, and a 7% drop in views of ads with scams and serious violations after broader testing.

These are presented as early-test and broader-test results, not a shipped state. The deployment framing is explicitly cautious — Meta says it will deploy these systems "over the next few years" and only "once we've seen them consistently perform better than our current methods." It also signals reduced reliance on third-party enforcement vendors in favor of internal systems and workforce.

For engineering teams the notable technical claim is language coverage: these systems reportedly work in languages spoken by 98% of people online, up from roughly 80 languages previously, with the ability to adapt to code words, emoji meanings, and slang. That breadth is what makes a conversational support surface plausible across Meta's global user base, since the assistant is shipping in all languages Facebook and Instagram support for these topics.

What an action-taking support agent obligates you to get right

The specific implication of this announcement is that once a support assistant can change passwords, edit profiles, and file reports, its failure modes are no longer "wrong answer" — they're "wrong action taken on a real account."

Meta's own enforcement example underlines the tension: it describes AI recognizing an account takeover from a new-location login plus a password change plus profile edits. Those are precisely the actions a user-facing support agent is being empowered to perform. The same signal that flags an attacker is a signal the helpful assistant could trigger on a legitimate user.

That is why the human-in-the-loop carve-outs matter more than the automation headline. Meta keeps people on the highest-risk decisions — appeals of account disablement and reports to law enforcement. Any team building agents that mutate user state should read the boundary Meta drew here as the real design lesson: let the agent handle the reversible, high-frequency tasks, and hold the irreversible, high-impact ones back for human judgment. The claimed under-five-second speed only earns trust if the destructive-action perimeter is drawn conservatively.