The Agency Builds Its Own AI, Self-Hosted Models Included

Mauricio Umansky did not hedge. "AI is the biggest shift our business will see in a generation, and we refuse to rent our future from someone else's playbook," the founder and CEO of The Agency said in a statement to Inman.

So the luxury brokerage is building its own AI platform. According to Inman's Oct. 23 report, it runs open-source large language models on infrastructure The Agency controls, alongside commercial models from companies such as OpenAI and Anthropic. It pulls MLS and market data, transaction management, brokerage financials, listing syndication, marketing tools and internal communications into one data lake that agents and leaders can query in plain language. The Agency says it will eventually reach its more than 4,000 agents worldwide.

The split is about who holds the data. Commercial models handle work like market analysis and forecasting. Sensitive client, transaction and financial data stays with models the company hosts itself, said Zane Burnett, The Agency's EVP of digital strategy.

"We don't necessarily want to run sensitive data through commercial models, because we have no idea what's happening there," Burnett told Inman.

It is early. The platform is in alpha testing with internal staff, and a beta with some leadership, including Umansky, is about a month out. The Agency plans to unveil it at its leadership summit in Miami in November, give all leadership access in the first quarter of 2027 and bring agents on later that year. Burnett declined to say how much the company has spent or how large the team is.

Nor is it ditching vendors. The Agency uses Rechat today and will keep some existing tools, reached from inside the new interface. Burnett compared the old single-sign-on dashboard to the knobs on a 20-year-old car stereo, and the new system to the controls on a modern steering wheel. "We're past dashboards," he said. The business case, in his telling, is recruiting and retention: make the work easier and agents stay.

That makes three build-it-yourself bets we have covered this month. Pokeworks built its own agent and stopped hiring at headquarters, and Real put 36,000 agents on pay-per-task AI. As our column on Wingstop argued, most brands cannot staff a model team.

That gap is where most franchise and multi-location operators sit, and it is where Revscale, which publishes Revscale Media, takes a different route. Its AI agents are configured in plain language for whatever workflow a business can describe, such as answering inbound inquiries, qualifying and routing leads, or re-engaging leads that went quiet. Every interaction is logged in Revscale CRM and can flow on into the systems a business already runs. The trade is control for speed: an operator gets working agents without hiring engineers, but it does not self-host its own models the way The Agency plans to.

Burnett admits he does not know what rivals are building. The real test arrives later in 2027, when the people using the platform are agents rather than executives.

Revscale Media is published by Revscale.