Failure intelligence, not failure trivia Monday, July 27, 2026

Artificial Intelligence

Facebook M

Facebook M was a Messenger assistant that could book a table, order flowers, or plan a trip, because real people were answering behind the "AI." The plan was for those humans to train the software to take over. The economics never got there, M stayed a small Bay Area beta for two and a half years, and Facebook shut it down in 2018.

Failed strategy Discontinued Moderate
Company
Facebook
Started
2015
Ended
2018
How far M scaled past its Bay Area beta
never
Collapse speed
Gradual
Preventability
Medium
Lesson transfer
Industry-wide
Last reviewed
2026-07-24

Narrative

The story

The ambition

Facebook M was meant to be the assistant that actually did things. Announced by Messenger chief David Marcus in August 2015, it lived inside Messenger and, unlike Siri, Cortana, or Google Now, promised to complete real tasks: book a restaurant, buy and ship a gift, arrange travel, sort out an appointment. The trick that made it work was hidden. M was, in Marcus's words, "powered by artificial intelligence that's trained and supervised by people," and in practice people did the hard parts. The bet was that those human trainers would gradually teach the AI to handle more on its own, until the software could scale to Messenger's billion users without a matching army of staff.

The rise

For a while the idea looked magical. Testers could ask M for almost anything and a competent answer came back, because a blend of Facebook workers and software was fielding the requests. By late 2015 the most common asks were restaurant suggestions and reservations, local-business searches, and help making purchases, along with ordering flowers, arranging travel, and setting reminders. It was a genuinely useful concierge, and it captured attention as a glimpse of a more capable assistant.

The cracks

The catch was the one built into the design. The magic ran on human labor, and the automation that was supposed to relieve it never arrived fast enough. Contemporary coverage flagged the obvious risk: if the service stayed too costly to run, Facebook would burn a lot of money on it, and complex requests, the kind people most wanted help with, took the most human effort. So M never left a small test group in the San Francisco Bay Area. It could not open to the wider world without either automation it did not have or a workforce it would not pay for.

The collapse

Because it stayed a limited beta, most people never touched M, and the ones who could often did not lean on it. On 8 January 2018 Facebook announced it would shut the standalone M assistant down on 19 January, after two and a half years. The staff who had powered it were offered other jobs at the company.

The aftermath

Facebook kept the M name for lightweight "M suggestions" in Messenger, prompts to send money or share your location, but the ambitious do-anything concierge was gone. M became a standard reference for the limits of the "humans in the loop until the AI catches up" playbook: a way to fake a capable assistant before the technology could deliver one, and a reminder that the faking can cost more than the market will bear.

The lessons

A demo powered by people is not a product that scales. M worked in beta precisely because humans did what the AI could not, which made every impressive result a cost rather than a capability. The plan to have those humans train the software into self-sufficiency assumed an automation curve that did not arrive on schedule, and until it did, growth only multiplied the labor bill. Bootstrapping an AI service on human concierges can prove demand, but it is a bridge, not a business, and the bridge has to reach the far side before the money runs out.

Causal timeline

Failure Anatomy

  1. 2015-08-26

    A do-anything assistant

    In August 2015 Messenger chief David Marcus announced M, an assistant inside Messenger that completed real tasks (booking, purchases, travel), positioned against Siri, Cortana, and Google Now. [1]

  2. 2015-08

    Powered by people

    M was "trained and supervised by people," with human operators plus software fielding requests, on the plan that the humans would teach the AI to take over. [2]

    Unsustainable economics
  3. 2015-10

    Useful, and labor-heavy

    By late 2015 the top requests were restaurant reservations, local-business searches, and purchase help; the most valuable, complex tasks took the most human effort, raising the cost of scaling. [3] [4]

    Unsustainable economics
  4. 2016

    Stuck in beta

    M never opened beyond a small San Francisco Bay Area test group, unable to scale without automation it lacked or staffing it would not fund. [5]

    No real demandFailure to adapt
  5. 2018-01-08

    Shutdown

    On 8 January 2018 Facebook said M would close on 19 January, after two and a half years; the human trainers were offered other roles, and only lightweight "M suggestions" remained in Messenger. [6]

Structured analysis

What Went Wrong

Root causes

The magic ran on human labor. M's useful answers came from human operators backing the AI, and that labor was too costly to scale to a wide launch while the automation meant to replace it lagged. [2] [4]

Few ever used it. M stayed a small Bay Area beta, so most users never had access, and those who did often did not rely on it. [5] [6]

Contributing factors

Automation never caught up. The premise that human trainers would quickly teach the AI to handle complex tasks on its own did not materialize fast enough to relieve the labor cost. [4]

Immediate trigger

Shut down still stuck in beta. On 8 January 2018 Facebook announced M would close on 19 January, having never scaled past its limited beta. [6]

Visible symptoms

Never left limited beta. After two and a half years M was still open only to a small test group in the San Francisco Bay Area. [5]

Warning signs

Too costly to scale. Early coverage warned that if M stayed too expensive to run, Facebook would burn money on it, and the most-wanted requests were the most labor-intensive. [4]

Affected groups

EmployeesCustomers

Evidence

Claims & sources

Every numbered marker in the analysis links to the claim it rests on, and each claim to its sources.

  1. [1]

    Facebook announced M in August 2015 as a Messenger assistant, led by Messenger chief David Marcus, that completed real tasks such as booking restaurants, buying and shipping gifts, and arranging travel, positioned against Siri, Cortana, and Google Now.

  2. [2]

    M was "trained and supervised by people," with human operators plus software fielding requests, on the plan that the humans would teach the AI to take over increasingly complex tasks.

  3. [3]

    By late 2015 M's most common requests were restaurant suggestions and reservations, local-business searches, and purchase help, along with ordering flowers, arranging travel, and setting reminders.

  4. [4]

    M's reliance on human labor made it costly to scale, and the automation meant to relieve that labor did not arrive fast enough, with the most-wanted complex requests taking the most human effort.

  5. [5]

    M never expanded beyond a small test group in the San Francisco Bay Area, so most users never had access to it.

  6. [6]

    On 8 January 2018 Facebook announced it would shut down the standalone M assistant on 19 January 2018, after two and a half years, and offered the staff who powered it other jobs; only lightweight "M suggestions" remained in Messenger.

Sources