Inside Facebook’s M: How AI Coaches Turn a Chatbot into a Commerce Powerhouse

When the author received a Starbucks pumpkin‑latte delivered by Facebook’s AI‑driven virtual assistant M, it sparked curiosity about the most complex technology Facebook is building inside its Messenger app.

M has been on the author’s phone for a month in an experimental phase, handling tasks such as web search, flight booking, ticket reservation, free‑coffee delivery, songwriting, and even drawing, as illustrated by the image below.

The author admits a lack of understanding of M’s underlying principles and raises ethical concerns, noting that M repeatedly replies, “I am artificial intelligence, but humans trained me.”

During a meeting with David Marcus, Vice President of Facebook Messenger, Marcus demonstrated that M could even draw a portrait of him, confirming the assistant’s creative capabilities.

Marcus explained that M can be scaled massively, but human “coaches”—contract workers at Facebook—review every AI‑generated response, approving or rewriting it before it reaches the user.

The interaction flow consists of four steps: (1) user query, (2) AI‑generated answer, (3) coach review/modification, (4) scaled deployment. Coaches’ edits are fed back to the AI to improve future performance.

Initially, M struggled with flower‑ordering queries, but after iterative coaching it now handles most of the process, though a human still oversees the final suggestion.

Other companies, such as GoButler, have reduced staff after automation, highlighting the broader impact of AI‑driven assistants on employment.

Oren Etzioni of the Allen Institute praised Facebook’s systematic and effective AI progress, while Alex Lebrun, former CEO of Wit.ai (acquired by Facebook), discussed the challenges of extending AI creativity to tasks like painting.

From a technical standpoint, M combines automation, deep learning, and AI to mine massive data sets and generate recommendations or actions.

Beyond commerce, M builds a personal database of user information—addresses, payment details, preferences—raising privacy considerations about how deeply an AI can know an individual.

Facebook envisions merchants integrating M via APIs to capture new customers, citing examples such as Everlane’s automated order confirmations and KLM’s Messenger‑based flight notifications.

The company has concentrated resources on AI, hiring over 50 researchers led by Yann LeCun and acquiring Wit.ai, positioning M as a potential Google‑search alternative that could monetize search results for merchants.

With a multi‑billion‑dollar market at stake, Facebook’s strategy mirrors Google’s investment in DeepMind, aiming to make Messenger the hub for online commerce.

Ultimately, Marcus believes that while AI will automate many tasks, human involvement will remain essential for the foreseeable future, and M’s evolution will continue to shape the future of digital assistants.