Since their inception, intelligent agents like Siri, Google Now, and Cortana have reverted to internet searches when they could not understand the meaning of a user request. A few years ago, I envisioned an intelligent agent that would seamlessly route the query to a human being if it could not understand it or fulfill it. Backed by a cadre of humans, such an agent could deliver a leap in service performance.

A colleague, formerly of Tellme Networks, which was acquired by Microsoft, assured me that such a model could not possibly work, because of the cost of employing the human beings. I pointed out that those costs would decline as fast as the intelligent agent’s performance (on its own) improved, and that ongoing improvement of its models and algorithms could be built in to the overall operation, but he insisted that this had all been analyzed previously and deemed infeasible.

Facebook M

It turns out that I was right. The Wired article entitled Facebook’s Human-Powered Assistant May Just Supercharge AI describes just such an agent. A number of things could have rendered this feasible for Facebook:

  • Facebook started with sufficient investment to make the venture profitable in the long run.

  • Facebook M is a text messaging agent, while my colleague and I were considering speech recognition systems, that are vulnerable to unintelligible audio.

  • Facebook simply hasn’t realized yet that it is infeasible!

The article was doubly interesting because it mentioned that was acquired to help build Facebook M. provides an attractive platform for crowdsourced intent classification. In other words, a developer can specify a limited set of “intents” to be recognized from various natural language statements. The system will generalize from this. When given a new natural language statement, it will generate a probability distribution across the intents. The system is not perfect, but it provides a dashboard for the developer to correct its mistakes. It will approach perfection over time. If a sufficient number and diversity of developers participated, then the system could recognize a corresponding number and diversity of intents, and become generally very useful.

It is not clear why was ultimately sold to Facebook. Shortly before their acquisition, they open-sourced their Duckling parser. I do know that one of its competitors, Ask Ziggy apparently ran out of money. Another competitor, Speaktoit recently re-branded itself as mobile application, backed by currently offers its service for free, which is likely to make things difficult for In any case, the platform business is bound to be difficult unless there is an end-to-end application that is sufficiently compelling and also promotes competition.