Kalepa's AI Copilot for Underwriting: Sharper Risk Selection, Won or Lost in the Rollout

Commercial insurance underwriting is a numbers business built on messy inputs. An underwriter pulls a submission apart, checks loss histories, verifies the account against appetite, cross-references external data, prices the risk, and decides whether to bind. Do it well and the portfolio is profitable. Do it slowly or inconsistently and the good risks go to a faster competitor while the bad ones stay on your book.

A New York company called Kalepa has spent years building AI aimed at exactly this workflow. Its platform, known to the market as Copilot, reads submissions, surfaces the risk insights that matter, and helps underwriters select and price risk faster and more consistently. This is a deep look at what Kalepa has built, who is running it, and, just as important, what it actually takes to get value from a tool like this once the contract is signed. The second part is where most enterprise AI quietly succeeds or fails.

Underwriting productivity is capped by how fast a human can gather and reconcile information. Much of an underwriter's day is not judgment. It is retrieval: opening documents, classifying them, checking completeness, screening for sanctions, hunting through external sources, and re-keying data into a rating tool. The expensive expertise sits idle while the expert does clerical work, and the volume of submissions a carrier can profitably evaluate is limited by that friction.

Kalepa was founded in 2018 to remove it. CEO and co-founder Paul Monasterio came to insurance from an unusual direction, with a background in physics and a prior role as VP at Applied Predictive Technologies. That analytical lineage shows in the product: Kalepa's pitch is not a chatbot, it is decision-support built to make risk selection measurably better. The company is backed by IA Ventures and Inspired Capital, having raised roughly $16 million to date, anchored by a $14M Series A led by Inspired Capital, and it has been named to the InsurTech 100 by FinTech Global.

Kalepa is not a single feature bolted onto an inbox. It is a workbench that carries a submission through the full underwriting motion, and the platform is organized into the stages an underwriter actually works through. According to the Kalepa platform, those stages include:

- Submission ingestion that automatically classifies and extracts data from hundreds of document types.

- Clearance that detects conflicts, verifies completeness, and runs sanctions screening.

- Triage that prioritizes the submissions most likely to bind and aligns them to the carrier's appetite.

- Risk analysis that pulls exposures, controls, and risk factors into a single dashboard, drawing on billions of data points from loss histories to real-time external sources.

- Rating that incorporates the carrier's pricing models and auto-populates rating criteria.

- Quote and bind that generates decision-ready documents without friction.

- Portfolio management that connects account-level and portfolio-level underwriting in real time.

What ties it together is what Kalepa calls its AI Ensemble Engine, designed to deliver transparent and verifiable results rather than opaque scores. As the company puts it on its page for underwriters, Copilot is built for the person doing the work, surfacing the critical insight instead of replacing the judgment. That transparency is the reason a compliance-heavy business will let AI near the risk decision at all.

The direct beneficiaries are clear. Underwriters get their time back for judgment instead of paperwork. Carriers get faster, more consistent decisions, better risk selection, and more capacity from the same team. Brokers and their clients get quicker quotes.

Insurance is a conservative buyer, so the customer list is the most persuasive evidence. Kalepa's clients include Admiral Insurance, Paragon Insurance Holdings, Munich Re Specialty North America, Canopius US, Bowhead Specialty, Berkley, James River, AmRisc, Landmark, and Merchants Insurance Group, among others. Paragon has publicly expanded its use of Copilot to additional programs, and Bowhead Specialty deployed Copilot to power better risk selection. Expansion and repeat deployments are the strongest signal in enterprise software, because they only happen when the first rollout actually worked.

The reported results are the kind that get a project renewed: a 960 basis point improvement in combined ratio, more than 30 percent additional premium per underwriter, and a 58 percent reduction in quote time. In an industry where a few points of combined ratio separate a profitable book from a losing one, those are not marginal numbers.

Kalepa is a clean example of a pattern showing up across enterprise AI. The winners are not the general-purpose tools that promise to do everything. They are the vertical, deeply specialized systems that go narrow, learn the rules of one high-stakes decision, and earn trust by showing their work. Commercial underwriting is a near-perfect target: high volume, expensive expertise, document-heavy inputs, clear economic value in every point of improvement, and a hard requirement for explainability.

It also illustrates the real bottleneck in enterprise AI, which is not the model. It is adoption. This is the same shift reshaping how work itself gets delivered, the move away from renting labor by the hour and toward buying outcomes that are measured and proven. The intelligence is increasingly a commodity. The scarce, valuable thing is the ability to turn it into a result inside a real organization. That is a theme we have written about as the future of work, and AI underwriting is one of its clearest live examples.

Kalepa markets that its platform goes "live in weeks, not years," and that modular design is a real advantage. But even weeks of integration is a project, and getting value is more than switching the software on. Real adoption has at least five distinct pieces of work, none of which the software does for you.

1. Appetite and guideline configuration. An AI underwriting workbench is only useful if it reflects your appetite, not a generic one. Every carrier and MGA has its own risk appetite, guidelines, thresholds, and exceptions, often differing by program and line of business. Encoding that, testing it against real historical submissions, and tuning the triage and risk analysis until the output matches what your best underwriters would do is careful, expert work. It is also what determines whether the recommendations get trusted or quietly ignored.

2. Document ingestion and data plumbing. Underwriting inputs are messy. Submissions, loss runs, ACORD forms, broker emails, and external data feeds arrive in hundreds of formats. Getting the system to classify and extract them reliably means handling every document type your submissions actually contain, connecting the external data sources, and validating the extraction against ground truth. This is unglamorous integration work, and it is where timelines slip when it is underestimated.

3. Integration into core systems and workflow. A recommendation that lives in a separate tab does not get used. Value depends on the output landing inside the systems underwriters already work in: the policy administration system, the rating engine, the workflow queue. Straight-through quote and bind only works when that integration is genuinely done, not demoed. This is often the single largest slice of the effort.

4. Underwriter change management and trust. This is the piece that gets skipped, and the one that decides success. Underwriters are experienced professionals being asked to trust a machine on the core of their craft. If they do not trust it, they will work around it, and adoption dies no matter how good the model is. Building that trust takes explainability they can see, a rollout that starts where the wins are obvious, training that respects their expertise, and a feedback loop so their corrections visibly improve the system.

5. Measurement and continuous tuning. The value case has to be proven in your own numbers, not the vendor's marketing. That means baselining quote time, hit ratio, combined ratio, and premium per underwriter before go-live, then measuring the same metrics after, on a defined book, and tuning based on what the data shows. Without this loop, you cannot prove ROI to the executives who funded the project, and you cannot defend the next phase of the rollout.

Notice what none of the five require: a better AI model. The bottleneck is almost never capability. It is capacity. Getting an AI underwriting platform fully live and driving results is a cross-functional effort spanning underwriting, IT, data, and compliance, and most carriers do not have a spare team that understands both insurance workflows and AI deployment sitting idle waiting to run it. A lean vendor like Kalepa keeps its own people focused on the product, and the carrier's underwriters are busy underwriting. So even a platform that can be live in weeks stalls in the last mile, and the value the business case promised arrives late or not at all.

This is not a Kalepa problem or an insurance problem. It is the defining problem of enterprise AI right now, and it is exactly why the old models for adding capacity fall short. Hiring a permanent team for a one-time deployment push is slow and expensive. Traditional outsourcing or offshoring rents you bodies by the month and leaves you managing them. Neither is built to deliver a specific outcome on a deadline.

The carriers that win with AI underwriting treat implementation as the actual project and resource it properly. That usually means bringing in a delivery layer that sits behind the vendor's product and the carrier's team and owns the last mile: configuring the appetite and guidelines, building the document and data pipelines, doing the core-system integration, running the underwriter change management, and standing up the measurement loop that proves ROI.

This is exactly the work a Virtual Delivery Center is built to deliver. AiDOOS runs elastic delivery pods that plug in alongside a product like Kalepa and the carrier's own staff, take on the deployment and adoption work, and get the system live and proving value fast. The vendor keeps its focus on the product. The carrier gets to the outcome without building a permanent team for a one-time push. And because the work is scoped as delivered outcomes rather than bodies rented by the month, the carrier pays for progress, not effort.

The sensible way to start is small and provable: pick one program or line of business, get the platform fully live on it, measure the before and after on real numbers, and use that proof to fund the broader rollout. Prove the loop on a slice, then scale it with confidence.

Kalepa has taken the messy, high-stakes work of commercial underwriting and rebuilt the workbench around AI that shows its reasoning. The carrier logos, the program expansions, and results measured in combined-ratio points suggest it is real. The technology, from vendors like Kalepa, is ready.

The question every carrier should ask before signing is not whether the AI works. It is who is going to do the work to make it work here, and how fast. Answer that well, and the business case comes true. Answer it by assuming adoption will happen on its own, and even a platform that promises to be live in weeks joins the long list of enterprise AI that was capable but never quite delivered. The model is no longer the hard part. Getting it adopted is, and that is a solvable problem.