Insurance is one of the most operationally intensive industries in the world. Behind every policy is a mountain of work - claims to investigate, submissions to triage, renewals to process - and almost all of it still moves forward by hand, one document, phone call, and system update at a time.
For decades, automating insurance work meant a tool for each step. OCR helped digitize documents, RPA wrote results back into policy systems, and IVRs routed calls. Each task still required real human work to move from first notice to close.
Pace does the whole task. Today we're launching one integrated, multimodal agent that runs every step itself, end to end. It reads the loss runs, reasons over your procedures, operates your policy system, picks up the phone with the broker, and closes the file. One agent for the entire unit of work.
Working across every modality
To keep the agent accurate across all modalities, we’ve built a few core capabilities:
Procedures
The Agent Operating Procedure (AOP) is our single surface area for describing the work. You write the steps in plain English, and the agent carries them out across voice, computer use, document extraction, APIs, and even deterministic code, holding every result in one continuous case.
When the work repeats and is specific to your book, you can abstract it into its own procedure. These are the parts no generic tool fits, because they’re shaped not just to insurance, but to the exact way you format and run the work. For example, a census in a layout no one has seen before or an endorsement that has to match your template down to the cell. You document each once, and every task that needs it inherits it.
Subagents
Accuracy comes from focus, so Pace agents work several channels at once by splitting the work into focused pieces. Each are handled by a subagent is scoped to a single task. Some are pre-tuned for a specific modality: a voice agent for the broker call, a computer-use agent for the write-back, a document agent for the loss run. Others spin up on demand for focused work that the main thread shouldn’t carry inline. As a result, you get specialist’s accuracy, while still following the AOP closely.
Shared context
The agent keeps one running thread of context the whole way through, no matter how often or when it moves between different modalities. A number it pulled from a document is still in hand when it’s on the phone an hour later, and what a customer says on that call flows straight into the system update it makes next. Because nothing is lost when the work moves from one channel to the next, the agent carries a single case all the way through, instead of losing track every time it changes tasks.
A demo of a Pace agent autonomously completing a premium audit across voice, email, document extraction, and computer use
Trusting every answer
Getting from demo to production takes more than just a capable agent. When a single answer is stitched together from documents, systems, and calls, trusting it means being able to check any figure and know it hasn’t drifted. So Pace agents show their work, in a few ways.
Pace agents cite and log every decision they make. Take any figure and you can follow it back through the calculation that produced it to the exact source behind it: the page in a loss run it was read from, the line in a call where it was confirmed, the field in your policy system it was written to. Calls come with the recording and transcript, and computer use with a full session replays and step-by-step reasoning.
Where citations show the work, graders check it. Graders are automated quality checks against the agent’s work. You define what correct looks like, and the graders runs automatically on every case. Some graders read the agent's work and decide whether it followed your procedure, like confirming the policy number before quoting a renewal. Others are queries that check the math, like whether the audited premium equals the sum of the class codes.
Finally, test cases let you check the agent across many runs at once. A test case is a real example paired with the answer you already know is correct, and a suite of them becomes your benchmark for accuracy. Every run, and every correction your team makes, becomes another test the agent is measured against, so it grows more reliable on your work over time.
A demo of Pace’s citations, graders, and test cases to ensure our multimodal agents stand up in production
In production today
For as long as anyone can remember, the size of an insurance operation has been capped by how many people it could hire. Today, that changes. The expertise and judgment stay with your team; the work runs alongside them, and the protection that work makes possible can finally reach further.
In the next few days, we'll share more about the product and engineering story behind our multimodal agents: specifically, a closer look at our voice and computer use agents, and a deep dive on how we built one agent that reasons, navigates, and speaks across modalities and months.
Pace agents are running tens of thousands of these tasks a week in production at customers like Prudential, WTW, Palomar, and Convex. If there’s one in your book still waiting to be automated, reach out.