// Private equity
We can read an insurance asset the way we ran one.
Dearborn Labs is an AI-native software development firm built specifically for insurance. Our founders built and operated an AI-native carrier for a decade — underwriting, claims, servicing, distribution, regulators, a live P&L. That is the credential a diligence read on a carrier, MGA or brokerage actually needs, and it is the one an engineering firm rarely has.
Three questions bring sponsors here: whether an asset's technology claims are real, whether the operating thesis on expense ratio is achievable, and what the first hundred days should contain.
// The situation
The dispersion in this industry is operational.
The loss ratio gap between top-quartile and laggard carriers is 26 points — 47% against 73%.
McKinsey Global Insurance Report 2025.
That spread is not explained by product or by rate filing alone. It is largely how the work gets done: which risks are seen, how fast they are decided, what gets re-keyed, and what the organization can measure about itself.
That is also where an operating thesis either holds or does not. So the useful diligence question is not whether an asset uses AI. It is which workflows are instrumented, which decisions are already automated in production, and what would have to be true for the expense line in the model to move.
// What a read actually looks at
The artifacts, not the roadmap deck.
Two patterns recur in insurance assets and both are visible in a week. The first is a modern platform on some lines with the historical premium and loss data, the part any model needs, still sitting on a retained legacy stack. The second is a production process whose real system of record is a spreadsheet and a mailbox, which is not automatically wrong and is always material to a build plan.
A third thing is worth checking alongside them: whether the asset’s core vendor is about to ship what the asset is planning to build. Guidewire and Duck Creek are both shipping their own agents now, which changes the build-versus-wait answer inside a hold period.
// The build
Four engagements a sponsor actually uses.
A technology read on an asset, pre-LOI or confirmatory
What is in production against what is described as in production. Which integration surfaces the builds depend on, and whether those surfaces are generally available or early access. Where the historical data actually lives. What the asset can measure about its own operations today, which is usually the sharpest signal.
Surface: the asset's own systems, contracts and artifacts, read by the engineer who writes the memo.
The expense-ratio thesis, instrumented
The model in the investment committee memo assumes a cost line moves. The read names which workflow that line belongs to, what the current touch count and cycle time are, and which of those numbers the asset does not currently collect. Every gap comes back with the instrument that would close it.
Surface: the workflow as observed, with the measurement definitions written down.
A hundred-day plan with a build sequence in it
Instrumentation first, because a baseline defined after a build is not a baseline. Then intake and document reconciliation, which is where the unavoidable manual volume sits. Then write-back into the system of record. Decisioning last, when the operating data exists to support it. The plan also names what not to build, including where waiting for the core vendor is the honest answer.
Surface: the operating plan, sequenced against the asset's change-control calendar.
The build itself, with transfer written into the scope
Senior engineers embedded with the asset's team, building alongside them. What gets built belongs to the asset — code, mapping layers, documentation and governance file — which is what a buyer diligences at exit as an asset rather than as a vendor relationship.
Surface: the asset's repositories, its sanctioned APIs, and its compliance file.
What we measure
Expense ratio decomposed to the workflow that owns each component rather than to the department. Touch count and cycle time per file, on a cohort definition stated in writing. Rework rate and correction rate on any automated output. And for exit readiness: whether the asset’s team can run and extend what was built without the firm that built it.
No forecast attaches to any of those. They are definitions and baselines. A projected saving printed before the baseline exists is a number nobody can defend in a data room.
// We ran one of these
Reading another carrier's operations quickly means knowing where the work actually goes, what the systems can and cannot do, and which parts of a technology story are load-bearing.
The operating record →// What's hard about this
Two limits worth stating before a read is commissioned.
A short read cannot validate model accuracy, and should not claim to.
Whether an asset’s pricing or triage model is well calibrated is an actuarial validation exercise against held-out experience, and it takes data and time a diligence window rarely has. What is checkable in days is the workflow around the model: what is in production, on which surface, at what volume, with how much human touch, and whether the correction loop is measured at all. A read states which of those it verified and which it did not, and an unverified claim stays unverified in the memo rather than becoming a finding.
Some of the operating data a thesis needs is not in any filing.
In delegated-authority business a material share of premium never appears in structured regulatory filings, and effective dating makes a naive premium or exposure extract wrong — commercial policies are amended constantly through endorsements, audits, cancellations and reinstatements. So the read works from the asset’s transaction-level records rather than from a summary, states the version semantics it used, and names the instrument that would produce each missing number in the first hundred days.
// What ships with it
An AI claim with no governance file is a liability.
Roughly half the states have adopted the NAIC AI Model Bulletin. Connecticut requires an annual AI compliance certification attested by a named officer, Iowa formally defines bias and outcomes testing, and New York’s Department of Financial Services states that an insurer cannot rely solely on a third party’s claim of non-discrimination.
NAIC and state bulletins · current at September 2026.
Responsibility is non-delegable, so a portfolio company running models in a regulated decision without a model inventory, lineage, testing evidence and a named human decision-maker is carrying an examination exposure into the hold period. A read checks whether that file exists. A build produces it.
What the governance file contains →Send us one asset and one workflow.
One senior engineer, the asset's own artifacts, and a written read of what is real, what is measurable and what the first hundred days should contain. Signed by the person who wrote it.