AI Support in Recovery
AI Support in Recovery
Context
Context
Starting in H2 2025 and through 2026, we’re rolling out Hacked Account Recovery, a cross-org effort between the Growth and Integrity orgs to unify Meta's fragmented account recovery flows to help hacked users recover their accounts in a secure, seamless, and scalable way.
In parallel, the Meta AI support assistant (MAISA) was launched in 2025, and its expansion and improvement remains a company-level priority.
Starting in H2 2025 and through 2026, we’re rolling out Hacked Account Recovery, a cross-org effort between the Growth and Integrity orgs to unify Meta's fragmented account recovery flows to help hacked users recover their accounts in a secure, seamless, and scalable way.
In parallel, the Meta AI support assistant (MAISA) was launched in 2025, and its expansion and improvement remains a company-level priority.
Problem Statement
Problem Statement
How might we use AI support to elevate the functionality of Hacked Account Recovery and improve persistent account access issues?
How might we use AI support to elevate the functionality of Hacked Account Recovery and improve persistent account access issues?
Current state
Current state
Today, MAISA can be accessed from various entrypoints in Meta products — from app settings, in chat threads, etc.
Today, MAISA can be accessed from various entrypoints in Meta products — from app settings, in chat threads, etc.
How it works: Tapping the entrypoint opens up a conversational chat UI over the app surface (referred to as the ‘scaled flow’) behind it. The user can start the conversation from a list of prompts, or type freeform into the input field.
How it works: Tapping the entrypoint opens up a conversational chat UI over the app surface (referred to as the ‘scaled flow’) behind it. The user can start the conversation from a list of prompts, or type freeform into the input field.
Constraints
Constraints
MAISA is still the early stages with clear functionality gaps, though with a lot of potential! The main limitation is a lack of context sharing, which results in a disconnect between the AI support flow and the scaled flow behind it.
MAISA is still the early stages with clear functionality gaps, though with a lot of potential! The main limitation is a lack of context sharing, which results in a disconnect between the AI support flow and the scaled flow behind it.
For example…
A user has already authenticated via code, then opens AI support on the password reset page. Because there is no context sharing between the scaled flow and the chat, the user essentially needs to start the recovery process from scratch.
For example…
A user has already authenticated via code, then opens AI support on the password reset page. Because there is no context sharing between the scaled flow and the chat, the user essentially needs to start the recovery process from scratch.
Collaboration model
Collaboration model
Until now, there wasn’t a clear collaboration model between the team owning support functionality and the teams owning the surface support would be implemented on - ‘scaled flows’.
This matters because access flows are highly optimized — ML recommendations, device-vettedness signals, and other logic run under the hood. Layering support on without integrating it into the scaled flow results in a lose-lose scenario where support can't meaningfully add value to the flow, and the flow's owners see it as clutter on their surfaces.
Until now, there wasn’t a clear collaboration model between the team owning support functionality and the teams owning the surface support would be implemented on - ‘scaled flows’.
This matters because access flows are highly optimized — ML recommendations, device-vettedness signals, and other logic run under the hood. Layering support on without integrating it into the scaled flow results in a lose-lose scenario where support can't meaningfully add value to the flow, and the flow's owners see it as clutter on their surfaces.
Before
AI support added “on top” of the scaled flow, running in parallel with no context sharing
Before
AI support added “on top” of the scaled flow, running in parallel with no context sharing
After
Scaled flow and AI support integrated with each other and sharing context
After
Scaled flow and AI support integrated with each other and sharing context
Opportunities
Opportunities
From UXR
When UXR participants ran into friction points or dead-ends in scaled flow, they expected personalized support and more flexibility (e.g., alternative pathways). Contextual AI entry points act as "safety nets," reducing anxiety even when unused. And lower-digital-literacy participants find AI especially empowering for navigating complexity.
From UXR
When UXR participants ran into friction points or dead-ends in scaled flow, they expected personalized support and more flexibility (e.g., alternative pathways). Contextual AI entry points act as "safety nets," reducing anxiety even when unused. And lower-digital-literacy participants find AI especially empowering for navigating complexity.
From data
The majority of funnel drop-off in recovery is code entry. Users struggle because they never get a code, they enter in the wrong code, their code expires, etc...
From data
The majority of funnel drop-off in recovery is code entry. Users struggle because they never get a code, they enter in the wrong code, their code expires, etc...
North Star
North Star
Anchoring on insights from UXR studies run on existing AI support experiences, we defined ideal aspects via Cursor of how AI support should function within recovery.
Anchoring on insights from UXR studies run on existing AI support experiences, we defined ideal aspects via Cursor of how AI support should function within recovery.
UXR INSIGHT #1
Contextual entrypoints
Participants prefer self-serve flows for speed but value AI for personalized guidance and flexibility—especially when they hit dead-ends. Lower-digital-literacy and older participants especially find AI empowering.
Contextual entry points should be subordinate to the main scaled flow, and prompts should be triggered at failure states or other moments users show they’re struggling.
Contextual entrypoints
Participants prefer self-serve flows for speed but value AI for personalized guidance and flexibility—especially when they hit dead-ends. Lower-digital-literacy and older participants especially find AI empowering.
Contextual entry points should be subordinate to the main scaled flow, and prompts should be triggered at failure states or other moments users show they’re struggling.
UXR INSIGHT #2
Continuity
Continuity is a critical user expectation. Participants expect one thread with context carried across AI ↔ scaled flow ↔ human, no repeating steps, and acknowledgement of completed actions and current state.
Continuity
Continuity is a critical user expectation. Participants expect one thread with context carried across AI ↔ scaled flow ↔ human, no repeating steps, and acknowledgement of completed actions and current state.
UXR INSIGHT #3
Dynamic greetings
Greetings messages should be dynamic based on the entry context, and name the problem and next steps clearly, since users often don’t realize they’ve been hacked, and they are reluctant to proceed without understanding why.
Vague or canned preset questions were called a "waste of time."
Dynamic greetings
Greetings messages should be dynamic based on the entry context, and name the problem and next steps clearly, since users often don’t realize they’ve been hacked, and they are reluctant to proceed without understanding why.
Vague or canned preset questions were called a "waste of time."
UXR INSIGHT #4
Transparency
Users want to be told the "quickest and surest" method to recover, but expect the chat to provide alternatives and to be able to explain why a given method was chosen.
Transparency
Users want to be told the "quickest and surest" method to recover, but expect the chat to provide alternatives and to be able to explain why a given method was chosen.
Visual variants
Visual variants
Playing with variations on placement of entrypoints (top right corner vs FAB style button) and animation styles (cross-fade, scale and fade) for cycling prompts.
Playing with variations on placement of entrypoints (top right corner vs FAB style button) and animation styles (cross-fade, scale and fade) for cycling prompts.
Next steps
Next steps
Negotiating with eng on feasibility and timelines!
Negotiating with eng on feasibility and timelines!