We have run eight data and AI engagements with specialty insurers over the last two years — Lloyd's syndicates, company-market carriers and one MGA. Underwriting discipline and regulatory posture differ from retail, and the AI patterns that travel are narrower than the market sometimes suggests. These are the five that have reproducibly moved numbers in our engagements.
Pattern 1 — submission triage and data extraction
Every underwriting team we have worked with is drowning in submissions. The core task — extract the salient risk signals from a heterogeneous pack (emails, slips, spreadsheets, SOVs) and decide whether this is a risk the team should quote — is genuinely well-suited to structured LLM extraction with a small human-in-the-loop.
Measured results on a cyber underwriting team: submission throughput +42%, quote response time down from 48 hours to under 9 hours on the 80% of well-formed submissions. The remaining 20% still require human diligence; that is correct.
Pattern 2 — clause extraction and wording diligence
Treaty reinsurance, facultative placements, MGA binders. The clauses matter, they differ subtly, and no one has the time to compare them carefully across a book. An LLM with a clause taxonomy and retrieval against precedent is useful for surfacing deviations. Compliance is very supportive because it replaces sampling with near-census.
One caveat: the vendor space for this is crowded and some offerings are thin. We have built custom and integrated vendor; custom tends to win for specialty because taxonomy precision matters.
Pattern 3 — outwards programme analytics
The reinsurance-purchasing side. The data problem is that the cede-to-treaty matching is a mess across legacy systems. A well-modelled cede data warehouse makes a materially different conversation possible with reinsurance brokers. This is not AI — it is data engineering — but it is the category with the largest underwriting-margin impact we have seen in the last two years.
Pattern 4 — loss reserving support
Carefully framed. Actuarial reserving is regulated; you are not replacing the actuary. What AI can do is flag anomalies in the claims development triangle that merit investigation, and compare current-period deterioration against historical case-level drivers. The output is an attention map for actuaries, not an alternative reserve.
Pattern 5 — broker and distribution intelligence
Which brokers are sending quality submissions; which are sending volume but poor conversion; which are shifting mix in a way that merits a conversation. This has been done in spreadsheets for years; the value is in doing it in a live, trustworthy way. Again not AI — data engineering and BI — but consistently valuable.
Patterns that keep appearing and keep under-delivering
Pure pricing models
For specialty, the data is sparse, the heterogeneity is high, and the underwriting art is significant. We have seen pricing ML projects that marginally beat a strong underwriter on narrow lines; none that have replaced the underwriter's judgement across a book. Useful as a second opinion, not a substitute.
Fraud detection beyond retail patterns
Specialty fraud is rarer and more bespoke; the retail ML playbook does not transfer. Forensic accounting and SIU workflow support are useful; generic fraud detection less so.
Chatbots for underwriting
Every eighteen months a vendor pitches an "underwriting copilot". The category has a future but the 2026 products are not yet where the underwriting teams we know would rely on them. Revisit in a year.
The governance posture
PRA and FCA expectations, Solvency II, SM&CR accountability — every AI deployment in an insurer has a named senior manager, a documented model risk assessment, and a monitoring plan. The bar is higher than in most sectors. The posture that works: build as if the model is a new underwriting tool, with validation, documentation and change control from day one.
The reference architecture
A tenancy-aligned data platform (most of our insurer clients are on Snowflake or Azure Fabric); a well-governed semantic layer for KPIs; a small set of ML and LLM services, each with model documentation and evaluation; integration into the policy, broker and claims systems. Nothing exotic. Discipline, not novelty, is what makes specialty AI work.