Lazarus AI Raises $43 Million to Give Boston’s Legacy Software an Exit Interview

Imagine a Boston insurance office where the newest employee is artificial intelligence and the oldest business rule lives inside software nobody is entirely comfortable touching. The robot has arrived to accelerate transformation. The mainframe would first like someone to explain the exception added several reorganizations ago.

This is a surprisingly good place to spend an AI funding round. Boston-headquartered Lazarus AI announced $43 million on November 3, with participants including QBE, State Farm Ventures, Mucker Capital, and One Way Ventures. The company says the money will support partnerships, research and development, and new products. The announcement does not identify a round stage or valuation.

The pitch is customer control over models, data, deployment, and costs. Its three highlighted businesses address document analysis, cybersecurity, and modernization of old software. My verdict is a serious enterprise software bet: useful problems, credible institutional interest, and substantial execution questions that will require more than a handsome declaration of independence.

The revolutionary act of keeping your own keys

Lazarus’s current product positioning puts ownership and operating choice at the center. The company promises deployment on customer infrastructure or in the customer’s cloud, model flexibility, and freedom from per-token charges. “Independent AI” is its organizing phrase.

Translated out of conference language, the attraction is straightforward. An organization may want to choose when a model changes, where confidential records are processed, and how much a workload costs. Those are practical purchasing concerns. A buyer can be enthusiastic about automation while declining to hand an outside vendor the steering wheel, the map, and authority to change the tolls.

That makes the Massachusetts connection substantive. This is a locally headquartered software company selling into insurance, healthcare, government, and financial institutions. Its market overlaps neatly with the region’s appetite for difficult institutional work. Boston does love a breakthrough that comes with an implementation meeting. Preferably several, so everyone can ask the question they prepared on the train.

The broader theme also runs through Tamr’s work on getting enterprise records straight: useful AI depends on the systems surrounding the model. Knowing what information means, who can use it, and how an output can be checked is commercially important even when it photographs poorly.

The mainframe has institutional memory and no LinkedIn

Constructor is the most interesting part of this story because it confronts a problem that more fluent code generation does not automatically solve: understanding existing behavior.

In its July 20 technical explanation, Lazarus describes reconstructing an application’s business logic before producing modern software. Constructor generates specifications, diagrams, and migration documentation, then creates Java from that reconstructed understanding. The company describes a COBOL emulator that exercises the original application and compares the replacement against the same scenarios, with traceability back to the original code.

Consider a hypothetical insurer with a peculiar premium adjustment that only applies to a narrow class of policies. Translating the syntax is one job. Identifying the adjustment, understanding when it applies, and confirming that the replacement preserves it are several more. A beautiful new application that quietly changes the calculation has modernized the interface and created an entirely different problem.

The approach is sensible. Its limits matter. Matching tested scenarios does not establish correctness for every possible input or operating condition. The announcement says Constructor is in pilots with two major mainframe makers, without naming them. That is promising access, not proof of a completed migration at production scale.

This is why mabl’s emphasis on verifying AI-written software belongs in the same conversation. Faster creation makes evidence more valuable. The old system may contain mistakes worth removing, too; deciding which behaviors to preserve still requires people who understand the business. Nobody should accidentally immortalize a bug because it has seniority.

The document needs a receipt; the receipt needs checking

Lazarus says its Applied Intelligence System links findings to source pages. Its funding announcement reports a 12.5-point improvement in field-level accuracy over standalone models across 150 customer use cases. That is a company-reported result, not an independently established accuracy rate for every buyer’s documents.

The product idea is useful even before accepting the performance claim. If an analyst asks why a system extracted a particular figure, a source reference makes review possible. It does not make the interpretation correct. A page can contain the right number under the wrong heading, or a footnote that changes what the number means.

QBE’s October 2024 investment account gives the commercial interest some history. The insurer discussed evidence-based reasoning and the potential to support underwriting and claims work. Its separate 2024 sustainability report described investing in and working with Lazarus to explore longer-term uses. This is an established relationship; this week’s financing did not invent it.

That evidence supports a serious buyer problem. It does not establish a current savings figure or justify treating an investment as a blanket product endorsement. As our Thomson Reuters legal-AI coverage explored, specialized knowledge and inspectable results become especially valuable when someone remains accountable for the answer.

Independence still has an electricity bill

Cybersecurity adds another version of the same evidence question. Lazarus’s September account of ClearWing describes an open-source security-research project intended to support inspectable, reproducible defensive work. It explicitly frames the tool as a contribution alongside professional teams and other infrastructure. That is a healthier ambition than pretending a scanner makes an organization invulnerable.

For prospective customers, control has operating costs. Someone must maintain infrastructure, manage access, evaluate updates, and handle failures. A pricing model without token fees can be attractive, but it does not eliminate compute, integration, or staff time. Buying independence can mean acquiring responsibility, which procurement will eventually discover even if the keynote skips that slide.

The meaningful test is the cost and reliability of a completed task: a correctly reviewed document, a reproducible security finding, or a migration whose behavior the engineering team can explain. These are proposed evaluation criteria, not results established by the funding announcement.

For readers outside Massachusetts, Lazarus illustrates an important buying preference: organizations want AI they can operate and scrutinize. This financing is a meaningful vote for that direction, with the proof still arriving customer by customer. Boston can reasonably enjoy the moment without declaring itself the capital of intelligence.

Give the old software a careful exit interview. It may be the only employee who remembers why the calculation works that way.