# Insurance agents can learn repetitive tasks in legacy systems

Markdown copy of https://realalpha.app/s/insurance-agents-can-learn-repetitive-tasks-in-legacy/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Jamie Cuffe](https://realalpha.app/builders/jamie-cuffe/) · Founder & CEO, Pace · Built AI agents that automate insurers’ back-office workflows, turning standard operating procedures into agent operating procedures.
- Host: [John Coogan](https://realalpha.app/people/john-coogan/) · Co-host, TBPN · Co-founded Soylent and Lucy; OpenAI bought his show TBPN in April 2026.
- Host: [Jordi Hays](https://realalpha.app/people/jordi-hays/) · Co-host, TBPN · Founded Branded Native; co-founded fintech Capital, which Rho bought.
- Show: TBPN
- Episode: [🦞 Moltbot / Clawdbot Creator Peter Steinberger, Palmer Luckey, and more](https://realalpha.app/v/c4kLgSWUfC8/), Jan 27, 2026
- Moment: 1:28:45 to 1:32:42 of the episode (4 min), playing on https://realalpha.app/s/insurance-agents-can-learn-repetitive-tasks-in-legacy/; the episode from there: https://realalpha.app/v/c4kLgSWUfC8/?t=5325

## Quote

> “It's much more about, you know, take this repetitive task and **fill out this form** thousands of times.”
> — [Jamie Cuffe](https://realalpha.app/builders/jamie-cuffe/) at [1:29:37](https://realalpha.app/v/c4kLgSWUfC8/?t=5377) · TBPN

## Takeaways

1. Small models can locate relevant parts of long documents.
2. RL environments target repetitive form-filling and CRUD tasks.
3. Forward-deployed engineers ship reusable code for customer workflows.

## About this segment

Pace combines models and builds reinforcement-learning environments for repetitive insurance work. Its forward-deployed engineers turn customer workflows into reusable product capabilities.

From “🦞 Moltbot / Clawdbot Creator Peter Steinberger, Palmer Luckey, and more”, Jan 27, 2026. [Watch the full episode](https://realalpha.app/v/c4kLgSWUfC8/)

## Transcript

From the transcript TBPN publishes with the episode.

[1:28:52](https://realalpha.app/v/c4kLgSWUfC8/?t=5332) Yeah. So I think we use a mixture of kind of models to be able to achieve what we need for the right task. If you think about parsing a many 100 claim error page claim file or policy document, you might use a small model to kind of zoom in on the areas that are most important and then use an extensive model to say, synthesize, you know, what's the most important and what's the answer. Sure. Sure. And then, you know, I think what's more interesting, you talked a little about kind of fine tuning an RL. Yeah. You know, we're kinda like the perfect set of use cases for for web agents being able to navigate, you know, these applications that don't have APIs, be they, you know, green screen CLIs or desktop apps. And so we've actually been, you know, working on a handful of fine tunes and creating basically RL environments that are these sort of CRUD applications.

[1:29:39](https://realalpha.app/v/c4kLgSWUfC8/?t=5379) They're very different from, say, know, the book me a restaurant or this flight type of demo. Yeah. Yeah. It's much more about, you know, take this repetitive task and fill out this form thousands of times. It's kind of the ideal case for, I think, a lot of these computer use systems. So we're seeing a lot of success there. I think we are just kind of rounding the corner of the scaling laws starting to work on those.

[1:29:59](https://realalpha.app/v/c4kLgSWUfC8/?t=5399) Kinda feels the same way that maybe, like, voice did, you know, six to twelve months ago.

[1:30:03](https://realalpha.app/v/c4kLgSWUfC8/?t=5403) Yep. You guys have pursued a forward deployed engineer I was about to market. Do you sounds like it's working well to date. Do you think this is something that will still be a part of the core motion in a few years? Or is this something that makes sense to get off the ground till you understand what the typical stacks are within your within your market, and then eventually, you'd move beyond it.

[1:30:31](https://realalpha.app/v/c4kLgSWUfC8/?t=5431) Yeah. I mean, I think the forward deployed motion is definitely something that, you know, really came naturally to me. It's kind of like the DNA of of our company because that's, you know, also what we were doing at Retul Yeah. When it was, you know, much earlier, I think, on that. I think for us, you know, going and working with our customers, being able to speak their language, getting to, you know, really understand the the complexities of these many 100 page standard operating procedures and make them successful as agents Mhmm.

[1:31:00](https://realalpha.app/v/c4kLgSWUfC8/?t=5460) I think it's really important for us to be sitting, you know, side by side with them. That being said, I think a lot of companies sort of talk about FTEs as sort of being the end state or, like, that is, like, you know, the best part about my business and why I'm, like, Palantir. I think for us, like, it is you know, these our FTEs are also really able to ship code in the code base, and so they are actually constantly making the product better. And that's really important. Like, I think our product should just be continually giving our forward deployed engineering team more and more leverage Yeah.

[1:31:31](https://realalpha.app/v/c4kLgSWUfC8/?t=5491) So that they can, you know, be able to to complete more for our customers. So some some FTE goes into a business. They're badge. They're in there. They're interfacing with the green screen, you know, local on premise mainframe software. They write some binding for it or some screen scraper or whatever you need to do, and then they can ship that back to HQ. And then the next time you run into that, you have you have an integration effectively. Is that, like, the basic pattern of that?

[1:31:57](https://realalpha.app/v/c4kLgSWUfC8/?t=5517) Exactly. I think there's just a lot of things that are required to get these agents running at scale that, you know, customers don't need to build from scratch every time. I I do think it's really important, that we also enable our customers to be able to build their own, you know, agents. And for us, you know, some of our customers are at tens of agents live in production. We may have built you know, help them build the first one or two in an FD motion.

[1:32:18](https://realalpha.app/v/c4kLgSWUfC8/?t=5538) But just because it's a natural language, they're able to sort of turn around two through 10 without us having And they're

[1:32:24](https://realalpha.app/v/c4kLgSWUfC8/?t=5544) still running those. And so they're running number three through 10 on your platform through your system, but they're defining them and architecting them themselves. Exactly. Yeah. Just a nice Very cool. Good business. I love it. Congratulations.