The AI software, called Metamate after the internal name for the group’s employees, is built on the tech giant’s large language model, Llama, and is used for coding, conducting research, as well as drafting communications internally and externally.

Prashant Ratanchandani, Meta’s VP of engineering who leads Metamate’s development, said there were no plans to release the software for external use but that the project offered valuable insights into enterprise AI applications. He told the FT that Meta wanted to create “the world’s best enterprise assistant”.

The move comes as tech groups, including OpenAI and Google, are racing to bring AI-powered assistants to consumers and businesses as they look to drive revenues from the fast-developing technology.

The market for so-called AI agents, which can reason and complete complex tasks for people, is projected to grow from $5.1bn to $47bn by 2030, according to data from market research firm MarketsandMarkets.

It comes as Meta explores ways to make money out of AI through enterprise products. Last month, it recruited Clara Shih, former chief executive of AI at Salesforce, to head up a new “Business AI” group. Shih said her vision was to “make cutting-edge AI accessible to every business” in her new role.

Analysts welcomed the move as an opportunity for the company to monetise Llama, for example, by building off-the-shelf applications leveraging the open-source models. 

Meta chief executive Mark Zuckerberg has made clear he wants to be a dominant AI player, quickly integrating AI features to Instagram, WhatsApp and Facebook, and offering AI agents to users, creators and businesses using the platforms. 

However, Metamate is not yet as advanced as similar products offered by rivals Microsoft and Google, which their staff use internally and are also sold to businesses.

Unlike most consumer-facing chatbots, Metamate is primarily a text-based interface and does not currently support video or images.

It also lacks the more autonomous, agent-like features of its rivals, such as joining and summarising meetings, scheduling items in calendars or taking actions on behalf of the user.

Meta employees, including executives, said they used the tool regularly for different tasks, including coding, finding internal information from the company intranet, and preparing for client calls and follow-ups. The company has also been using other models internally including OpenAI’s ChatGPT.

Ratanchandani said Meta was considering ways to improve the product, including making it more like an AI assistant. He said it was too early to measure its impact on productivity.

Across the market, adoption of the technology is limited. A recent report by Goldman Sachs suggested that widespread uptake or productivity gains from AI tools was not likely until at least 2027.

“It is a very common playbook to develop something internally and ‘dog food’ [employees to test it]”, said Oren Etzioni, a computer science professor at the University of Washington.

However, he added: “[Just because] it works great inside Meta, that doesn’t mean that it’s easy for Meta — which is a very focused advertising-based social network — to become an enterprise vendor. They would not be at the top of my list for who will emerge as a winner in the space.”