We’re on a mission to build certainty in a relentlessly uncertain world. If you run an insurer and share our vision, let’s build the future together.

Enduring institutions weren’t built by standing still. At Pace, we’re changing the economics of serving insurance so the world’s leading insurers can cover more risks, build products that weren’t before possible, and make the next era of progress inevitable.

Insurance workflows demand highly performant computer use. Our team has made major strides in building cost efficient, self improving harnesses. Read more from @aryanjn9 below:

Computer use is one of our most demanding workloads at @pacecom. Our agents drive real browsers and desktop applications through long, multi-step workflows, where every action depends on correctly understanding the screen before it. It's the one place where we usually have to make tradeoffs between model intelligence and speed. When we tested Sonnet 5, it ran through customer workflows quickly without giving up accuracy, landing at the top of our evals. That combination lets us ship faster, more capable agents that deliver real value to our customers.

Bridging the gap between demo and production requires real engineering depth. Multimodal agents in particular require in-depth context management and robust orchestration pipelines in order to keep them reliable over time. In the below post @yeauxgi shares a behind the scenes look of some of the engineering work that goes behind our multimodal agents and how we've been able to reliably stand these up in production over the last few months.

Most insurance work still lives in systems that were never built to talk to anything else: policy admin platforms, claims systems, the mainframe you reach through a green screen. Still, these are the platforms that hold the data the mission-critical work depends on. Computer use removes that limit. If a person can operate a system, a @pacecom agent can too: it sees the screen, reasons about what's on it, and acts. We've pointed it at the trickiest UIs out there - cluttered policy platforms, 40-year-old mainframes, systems behind two-factor logins - and because it reasons over screens instead of following hardcoded clicks, it doesn't break when a field moves. This is the work nobody could automate, till now. See how we build production-grade multi-modal agents here

Writing and servicing policies doesn’t just happen in one channel. Endorsement requests, cancellations, and first notices of loss often get scheduled by phone, confirmed by text, and followed up via email. @pacecom agents work the way the work actually happens, and in the front office, that often means voice. Voice is also one of the hardest channels to get right in production, because live calls are messy. People interrupt, go off-script, give the wrong information, or call back three days later and expect you to pick up where you left off. At the same time, voice is unforgiving on speed. A caller expects an answer in less than a second, and it still has to be right. So we built the infrastructure to make sure our agents hold up against that reality: In Prompt Studio, you design and iterate on agent behavior in a controlled environment, instead of tuning blindly in production. You set up the graders that score the agent on latency, quality, correctness, and tone, and optimize against them. Then in Simulations you run the agent against hundreds of realistic conversations and those same graders measure how it holds up before a single real customer is on the line. This is why top insurers trust Pace’s voice agents to handle over 90% of calls in their contact center fully autonomously. See how we build production-grade voice agents here.