Insurance Spends $70 Billion on BPOs. Pace Wants 80% Gross Margins Instead

Pace, Jamie Cuffe tells Sequoia hosts Lauren Reeder and Pat Grady, is the agentic process outsourcer for insurance. For a couple of decades the industry shifted work from onshore teams to BPOs. His thesis for the next decade is a shift from outsourcing to people toward outsourcing to AI. The company works only with insurance carriers, automating back-office operations that today sit with business-process outsourcers.

The path looks unlikely on a résumé—top of his Princeton class, Sequoia investor, rising star at Retool, then insurance BPOs—but Cuffe says he actually started there. He was born in London and grew up between New York, London, and Bermuda, the capitals of insurance. His father ran operations for a reinsurer and then for a Lloyd's coverholder in London, so the problem was never abstract. He did not run at it out of college. The turn came at Retool, whose customers included far more financial-services firms than the startup stereotype suggests, because operationally intensive businesses need custom internal tools that off-the-shelf software never covered.

As a forward-deployed engineer he sat in the front row: legacy carriers and tech-forward ones alike, from Ethos and Vouch through Progressive and Berkshire Hathaway. People were even building Retool apps to streamline BPO operations. He watched the same workflows his father had lived with—policy administration, claims, submission intake. Software at that point could dent BPO spend by about 10%. That was the ceiling of OCR and RPA. After the ChatGPT moment, while he was working on Retool's AI products, the economics flipped. Instead of augmenting outsourced services with software, you could replace the vertical service entirely.

Insurance is a particularly good starting point, he says, because the lingua franca of the industry is emails and PDFs. Too many intermediaries sit on different systems; with no common API the fallback is email, PDFs, and phone calls. The same manual document-and-email handling that created the BPO industry is now what AI agents are well positioned to do. Longer term he wants the agentic process outsourcer across many industries. Insurance is the beachhead.

Grady asks how you build a better business than the last generation of BPOs while taking the same workflows and the same spend. What keeps Cuffe up at night, he says, is exactly that question. Pace is landing impressive customers. It could become a very large company by remaining a BPO. There are already multiple public companies worth tens of billions, running at 10–15% gross margin, with tens to hundreds of thousands of people doing this work. Those are excellent businesses with exceptional teams. For Pace, building a multi-billion-dollar BPO without changing the industry's economics is the failure case he talks about with his own team.

The goal is to take that 10% gross-margin business to 80%, with hundreds of thousands of AI agents and a small team of insurance experts instead of hundreds of thousands of people. That, he thinks, is a much larger company. The step after that is to do it not in one vertical but across many. People tell him to push gross margins to later, or that they do not matter much. He disagrees. Margins do not have to be high on day one, but you need a path. The flywheel he wants is simple: every product improvement lifts gross margin. Even if the number is lower today, as Pace takes on harder tasks, that figure should keep moving up.

The workflows Pace takes are mission-critical processes that customers already want done end to end. Pre-AI, extracting data from unstructured documents was useful but only one slice: you still had to reason, apply judgment, and write back into internal systems—the part that previously required humans. One tempting product is a DAG-style workflow builder with AI nodes for extraction or email. Cuffe thinks that stance is short on AI. BPO work exists because the process was so hard to codify deterministically: human judgment and endless edge cases. If you shoehorn a model into tiny boxes and wrap a code layer around them, you miss the point of what can be done today. Pace is building for agents that can take any standard operating procedure and run it end to end.

On the vast majority of workflows there is no human in the loop; AI processes the work end to end. Pace is building an operations team for the most complex cases, with escalation to the customer available, but the main human role is onboarding. Out of the box an agent might sit around 90% accuracy. Customers in these critical workflows want 99.5% plus. The gap is closed with labeled examples and a self-improvement loop. Over time the human effort shrinks and the operations team moves to the next harder task.

Pace cannot take every outsourced workflow on day one. It starts with work that is already nicely outsourced at scale: the SOP is already written, accuracy is already checked internally or by sample QA, and spinning down a BPO is easier change management than retraining a W-2 workforce. It hunts the high-volume line items with the biggest ROI, usually at the top of the funnel—submission intake of new risk, extracting data, applying business logic, writing into underwriting platforms, and the claims equivalent in first notice of loss. Customers then pull the company down the policy life cycle: endorsements, billing, payout, quality assurance.

In the BPO world the conversation starts with cost. NPS on incumbent vendors is often low, and the line item is huge—eight figures plus for many of Pace's customers. Working with some of them, Pace has hit 50–75% cost savings in a short period. What usually happens next, Cuffe says, is that customers notice things BPOs simply could not do, which agents running at superhuman speed now make possible.

A submission rarely goes to only one carrier. Whoever answers faster can close the business. Seasonality is violent: January 1 enrollments, hurricane season, a spike of claims that used to mean spinning up extra BPO headcount. Agents can work around the clock and chew through the backlog before it becomes one. Observability is the third shift. BPO work used to be a black box—hard to see what was done, which stage a task sat in, where the bottleneck was. With agents you can surface the steps and the reasoning, which is old-fashioned software hygiene, and customers start redesigning workflows instead of copying what the BPO did.

The human baseline is also less flattering than the industry talk. Average BPO error rates, Cuffe says, run around 5–10%, and some customers sit much higher. The usual AI question is whether the model can match a person. In a lot of Pace's use cases he thinks it can be far more accurate. Picture claims or policy QA: a 300-page file, hundreds of rules, Friday at 3 p.m. A person will miss something on page 298. The model applies the same quality every time. The promise is that you can scale your best BPO rep. Grady reaches for Waymo in San Francisco: a driver with perfect attention who has seen every corner case a million times. Properly trained AI, he says, already has an existence proof that it can do a lot of jobs better than most people.

Reeder cites the MIT finding that 95% of AI pilots fail, and the contrast with Sequoia companies that do not. Getting AI into production is not a done deal when you walk in, Cuffe says; other vendors and in-house builds fail often enough. Pace's record is the other extreme: every pilot has been successful and gone through to production, which he is proud of. The company value behind it is closing the distance. They fly on-site, they only do these insurance workflows, and they have enough insurance experience on the team that customers hear their own language. The hard work happens up front, not after a seven-figure contract is live—especially in a market still burned by exciting demos that did not pan out.

He built one of the larger forward-deployed engineering teams of the Palantir era at Retool, and the profile that works is former founders: engineers who want to be commercial, or commercial people who are technical. The job is whatever it takes to make the customer successful—prompt-tuning a model, diving into operations data to build a clean eval set, integrating internal systems. Hiring is both sides of the aisle, sat close together beside the customer: operators from mission-critical workflows at top-five carriers next to former YC founders from the agent frontier. There are not many people already fluent in both. People from the BPO world often feel the thesis in their bones, and Cuffe has been impressed at how fast each side picks up the other skill set.

Are forward-deployed engineers a temporary patch until models get reliable enough? Cuffe and Grady agree on the end state: software-codified workflows. In the short term Cuffe would rather have more of these people, not fewer. They expand inside a customer's BPO budget, and they ship into the production codebase—Pace tests that in the on-site interview. The feedback loop is brutal and useful: see a problem, fix it, do not open a ticket for someone else. Deployment times are already falling; leverage is rising as work gets pushed back into software. After the first one or two workflows, some customers have a business analyst who has taken the count to nine or ten in production. About 80% of Pace is engineering, with forward-deployed engineers above that and go-to-market on top. Those layers have to stay in balance so the product actually gets investment.

Grady zooms out. In this market you want to be pegged at one end of distribution, not stuck in the middle: ChatGPT and OpenEvidence are bottoms-up tools people love and invent uses for; Harvey, Sierra, and Pace are top-down solutions. If you take the solution route, what you are selling right now is trust as much as software. Customers can see the magic and the failure modes. The vendor's job is to say we will make sure you are good, and to bring the messy reality of the workflow back so the company is not theorizing from headquarters. Insurance, Cuffe adds, is already the business of trust. Another Pace value is be the rock: carriers show up on their insureds' hardest days, and Pace intends to show up for those carriers every day. The GTM motion matches Harvey and Sierra—win the highest-volume, highest-ROI logos at the top of the market and let results carry the next set.

The third company value is set the pace, Cuffe's favorite. He has a high sense of urgency—impatience, he admits—and feels lucky to be building now. Naming the company Pace was about speed: the value customers already see, the internal tempo, and a deliberate step toward a larger mission. That mission traces to a Sequoia internship, when he learned that partner Andrew Reed's favorite company was Constellation Software. (Grady jokes that Reed's favorite is now whatever is in the Sequoia portfolio—Figma, Robinhood—and they settle on favorite non-Sequoia public company, not investment advice.) Mark Leonard ran Constellation for about thirty years, aggregating niche vertical software businesses that were not legendary on their own into a federated compounder, one of Canada's largest companies, on the order of $70–80 billion.

Cuffe wants the same move in services. BPO spend in banking, financial services, and insurance is a little under $400 billion a year, about the same as global cloud software; insurance is about $70 billion of that. A platform that aggregates those niche vertical services across BFSI, and maybe beyond, is a huge market. The more interesting part is not just aggregation but transforming the industry's economics. Grady calls these markets sneaky big: much larger than they look because they are not household B2B SaaS names.

Going after BPO spend, Grady adds, also gives Pace a clean interface. The customer already has a contract surface with the BPO, objectives are already written down, eval is built in because work comes back and gets checked, and many buyers run champion-challenger setups across vendors. Cuffe agrees. Carriers already know that high-volume, repetitive work with human judgment lives in the BPO line. SOPs can be modeled in the product, accuracy can be scored, change management is faster than moving employees, and the baseline is sitting there in the open—even when that baseline is worse than it should be.

Engagements usually start in the office of the COO. Cuffe dislikes the phrase proof of concept; POCs have a bad reputation and are the wrong move in AI. The job is to get the customer to success as fast as possible, on real data, not demos. Pick something, then define success criteria up front—and encourage the customer to make those criteria as hard as they want. Pace wants to show it is up for that. Teams fly on-site and take existing SOPs, often 50 to 100 pages with around 60 steps, complete with red boxes around the field you pull from one system and type into an admin panel. Because the work was already handed to a BPO, it is fairly well codified, but there is always a gap, shadow practice the document does not capture. Closing the distance is how you find it quickly.

Those SOPs become what Pace calls agent operating procedures: a Notion-like document of natural-language steps, with tools hung off them. There is long-context retrieval over nasty document sets, reasoning over large rule piles, and write-back into vertical systems of record. When there is no API—because it was never built, or because the carrier does not own the other end, as when a broker works a carrier portal—Pace uses web agents on the interface. If there is an API they use it; if not, the web agent still has to make the customer successful. On mission-critical use cases, several customers have told them Pace was the fastest company they had ever taken to a pilot and the fastest to production.

His wishlist for the labs starts with gratitude for reasoning-model scaling, which already gives better-than-human reliability on much of the document extraction, even if the problem is not fully solved. The next frontier is web agents. He would like the best ones sent his way. The canonical insurance task is CRUD on admin panels, which looks nothing like booking a restaurant or a flight. Longer term he wants RL and reinforcement fine-tuning. Because Pace does not trap the model in a single DAG node—the analogy is you would not put your best person in a box and tell them only to classify documents with no sense of before and after—the agent sees input through output, with an expected result you can grade. That makes a reward function, and per-customer, per-workflow RL, a natural fit. The long-term product is not freezing the process into deterministic code. It is a well-modeled world the agent can act in, with RL across the tool calls.

Two things have surprised him about building with AI. One is how much fun it is. Cursor, Claude Code, and Codex help the team move, but the deeper shift is that work which used to take decades of knobs and requirements can now be offloaded to an agent. The second is advice: go be a forward-deployed engineer, ideally at Pace. In most companies you either build or you sell; FDEs do both, which is founder training. At Retool, CEO David Hsu and Sequoia's Bryan Schreier bet on him early to stand up post-sales and take new products from zero to 100, which made this second company feel tractable. He interned at Sequoia and still loves the screen by the New York elevators: we are only as good as our next investment. Pace translates it to every quarter resets; you are only as good as your next customer. This time he is a solo founder. Co-founders matter in the desert; once the work is execution risk, involving the whole team in values, hiring, and product lets you hire stronger people. He met Schreier in college by staying up all night on a 20-page memo after Schreier mentioned a visit to Princeton's president. Schreier has watched every step since. Finding people who back you again and again, Cuffe says, is the fortune you actually need.