Thinking about new GenAI initiatives can be dauting, it is not a problem of what to build but what to build first. Because the technology is deceptively simple to pilot and can seemingly address a wide-ranging number of workplace processes, enterprises often experience AI Sprawl* (which is the unchecked proliferation of AI tools across departments without a unified strategy). The goal here should be mundane consistent reliability, not “razzle-dazzle” and to synergise enterprise efforts instead of unchecked AI sprawl along the way. GenAI can only deliver value if it’s targeting real problems, properly grounded in enterprise data and can be shipped into day-to-day work.
Why now?
Many insurers have pilots; but few have made measurable impact. The differentiator isn’t in the model they have selected but it’s selecting a problem worth solving, such as the intelligent automation of the entire client onboarding process, shortening of claims handling times, or legal processes. At the same time, uncoordinated pilots and team-level experiments create AI sprawl leading to overlap (reinventing the wheel multiple times), compliance gaps, and cost inefficiencies.
What’s changing
Early wins in marine insurance skew toward operational efficiencies: summarising large amounts of underwriting information, voyage/sanctions checks and summarising claims reports. These succeed when outputs are auditable with useful first drafts that remove tedious manual effort without bypassing the human in the loop evaluation.
How to cold-start—practical steps
Begin with return on investment, not a demo. Select use cases where a 10–20% improvement can immediately equate to operational savings or unlock opportunity. Define the metrics and a go/no-go gate before you write a line of code. Find a problem worth solving with proven utility. Onboarding processes for new clients can serve as an ideal such use case (including sanctions checks) , as its benefits are visible, measurable and immediate.
Identify the knowledge necessary. GenAI agents are not truth engines, they’re probability engines that require contextual knowledge to analyse. Useful projects begin with enterprise knowledge, policies, documentation, databases—indexed with access controls, freshness rules and citations. If your task requires specialised information, such as correlations of customer data with risk, or the analysis of an underwriting portfolio in a line such as cargo, then make sure you can supply proprietary knowledge.
https://www.moodys.com/web/en/us/insights/banking/good-ai-without-good-data-dont-bank-on-it.html
Make that knowledge usable. How that knowledge is consumed by the agent matters, a lot. This concept is known as “context engineering”, what gets retrieved, how it’s filtered, how it’s attributed. Treat this as a first-class design task. In fact, this is one of the key gates that prevent pilots transitioning into production.
https://www.moodys.com/web/en/us/creditview/blog/beyond-prompts-why-enterprise-ai-demands-context-e…
- Use the right technique. Needs internal knowledge that is up to date? GenAI + RAG (Retrieval Augmented Generation).
- Repeatable, formulaic task (e.g. Underwriting memo)? String multiple agents and tools in a deterministic workflow (schema checks, approvals).
- Pattern scoring (e.g. fraud detection)? Often classic Machine Learning, not GenAI.
- Some problems may be just process fixes—no model required, use a workflow solution instead.
- Not every problem is an AI problem, not every AI problem is a GenAI problem. For example there may be workflow solutions that can be automated with classic form filling technology; there are climate risk models that needs machine learning, not language models. Understand the nuances and deploy the right technique for the right task.
- Design for “boring” reliability. Think plumbing: role-based access, evaluation harnesses, monitoring, rollback, incident playbooks and human-in-the-loop. Success looks uneventful because the work simply gets done behind the scenes.
- Change the work, not just the tool. Rethink the definition of the role. Move human involvement from document production into evaluation/analytics; refresh Standard Operating Procedures; measure adoption and outcomes on a recurring basis. If AI can automate a large part of the production of the mundane, then that leaves more time for people to deliberate on the truly strategic and challenging cases.
- Prevent AI sprawl. Stand up a lightweight AI governance office or centre of excellence (sets policy, guardrails, model/prompt registry, approved GenAI stack, risk review) and federate delivery (domains build on shared connectors, retrieval patterns, and security baselines).
Lastly: To build, buy or partner? Unlike traditional software, GenAI is modular. You don’t need to build the entire stack. Consider build/buy/partner to piece together the best of breed components. Your focus should be on solving your enterprise’s most pressing challenges, not to reinvent the wheel on agentic retrievals or sourcing proprietary research.
So, pick a valuable problem, ground it in relevant data, choose the right technique and ship something boring but useful without letting sprawl take the helm.
