Unify credits task-specific OpenAI models for 30% pipeline jump

Most sales teams still treat prospecting like an art project — lots of manual research, guesswork, and spray-and-pray emails. Unify, a go-to-market platform, decided to treat it like an engineering problem instead. The bet appears to be paying off. By matching specific OpenAI models to discrete jobs within their sales workflow, the company now generates 30% of its own pipeline through AI-driven outreach, while supporting hundreds of millions in pipeline for customers.

The key wasn’t just throwing a single chatbot at the problem. Unify’s system rests on three distinct components, each powered by a different model. An observation layer, running on OpenAI o3, continuously scans public data for high-signal triggers like executive hires or tech stack changes. A research agent leans on GPT‑4.1 for planning and the Computer-Using Agent (CUA) to actually browse and navigate websites — the kind of dynamic, UI-level work that old-school scraping can’t handle. Finally, GPT‑4o synthesizes all that context into hyper-personalized email drafts.

“There’s something really special about human-to-human interactions that isn’t going away,” says co-founder Connor Heggie. The goal, rather, is automating the soul-crushing busywork that eats up a seller’s day. Software engineer Kunal Rai pointed to signal identification as the hardest part of the search problem, noting that “reasoning quality really matters in those early steps.” For Unify, o3’s multi-turn reasoning outperformed open-source alternatives when it came to grasping nuance across diverse sales scenarios, which is why it sits at the top of the funnel making classification calls that cascade into everything else.

Unify actually architected its platform in anticipation of models getting dramatically better at reasoning and tool use, scoping its observation engine before the models to power it even shipped. When o3, GPT‑4.1, and CUA dropped, the team had structured evaluations ready to go — not just measuring accuracy or latency, but reasoning quality in real-world GTM tasks. That kind of forward deployment thinking, where you swap in new brains as they prove themselves against your specific tests, is what turns growth from a guessing game into something closer to a science.