By Patrick O’Neill, President & Founder, Redhand Advisors | August 2026
I recently facilitated an AI discussion with a group of risk management leaders from large organizations across various industries. No vendors present, no sales agenda, just practitioners comparing notes on AI.
One participant said it directly: “I frankly wouldn’t have been this open” in a bigger forum. What follows is what they said, anonymized, along with my take after 30 years in RiskTech.
The ROI Isn’t Showing Up
Every broker, carrier, and system provider these leaders work with is marketing AI heavily. Yet, none of the participants could point to many tangible benefits they’ve received from it.
The frustration has teeth because these buyers carry real mandates. One was tasked with delivering significant savings across their P&L and said technology has contributed nothing toward their target.
It’s gone further than disappointment. One team now produces policy schematics in third party AI tools and catches coverage issues before their broker does. When they showed their AI-generated work to the broker, the answer was that management wouldn’t approve outside tools.
My take: This matches what I see across the market. Data from this year’s RMIS Report survey shows it clearly: only 40% of organizations are using AI capabilities within their RMIS, yet roughly 60% are using other AI tools in their department. That 20-point gap is the clearest picture of the problem I can give you. People found the value; they found it outside their core systems.
The gap exists because most AI features shipped so far are broad and generic: summarize the notes, chat with the data. One participant called them functionally identical to the general-purpose AI tools their company already licenses. Providers need to meet clients where they are, with tools built for the specific problems risk teams face every day. Generic AI features don’t solve those problems, and clients can tell.
So Clients Built It Themselves
The concrete wins in the room all came from teams using general-purpose or domain-specific AI tools on their own initiative:
- Two years of claims data analyzed to audit vendor performance. Done manually, it would have been a week of work nobody would have undertaken. The analysis uncovered findings that may not otherwise have been detected and ultimately supported meaningful cost savings.
- Limits, deductibles, and retentions quickly extracted from hundreds of policies into one consolidated table. Additionally, the AI tool has been effective in extracting more nuanced information, such as notification triggers and notice requirements.
- Summarizing excess casualty policy and identifying key exclusions.
- Preparing quarterly claims review package by gathering claims data, creating a trend analysis in Excel, drafting an executive summary, building a PowerPoint presentation, and preparing meeting materials.
- Building working claims and underwriting dashboards, plus a multi-page executive summary, built directly from the raw data.
- Training documentation generated from a narrated screen recording: the AI took the screenshots, annotated the click targets, and wrote the steps.
Notice the pattern. These are unglamorous jobs: summarization, research, cleanup, extraction, reconciliation, documentation. That’s exactly where the money is.
My take: The teams getting results skipped the “AI strategy” phase entirely. They picked one painful gap, ran an experiment, checked the output, and kept going. The RMIS Report survey backs this up too: the top blocker is no longer security and privacy (last year’s #1) but people not knowing what AI can actually do for them. That’s solved by experimenting, and by asking the tool an open question: here’s my data, what can you find?
The Question Nobody Was Asking a Year Ago
The sharpest discussion was about the need for a RMIS. More than one participant is actively weighing whether the RMIS is still the best investment.
Their reasoning deserves attention. The switching barrier that once locked buyers in, the brutal data conversion project, is largely gone; AI can do the mapping now. TPA feeds, data cleansing, and report production can all be handled with AI tools and a shared folder. A one-time build cost compares well against a subscription that increases every year.
My take: I understand the impulse, but replacing a RMIS outright isn’t realistic today for most organizations. A modern RMIS does a lot of comprehensive, hard-won work: claims administration with payments, reserves, and adjuster workflow; linked claims, policy, and exposure data; TPA and carrier feeds; audit trails and compliance support. Rebuilding all of that with general-purpose AI tools is a much bigger lift than a 2-hour dashboard demo suggests, and for teams that self-administer claims there’s no credible substitute at all.
The real near-term opportunity is AI enhancing the RMIS, and this is where providers should focus:
- Data conversion. Carrier and TPA feed setups are expensive. AI can do this work much more efficiently and effectively. Providers already have the target mappings; pointing AI at them would cut cost and turnaround dramatically.
- Reporting. Let users ask questions in plain language and get answers from live data instead of running a report.
- Insights. Profile the data proactively: surface quality gaps, duplicate records, emerging loss patterns, and the fields worth collecting that clients haven’t thought to ask for.
- Data quality. The most underrated problem in risk technology. AI is remarkably good at finding flaws in a dataset and embedding that in the platform would deliver great value.
Providers that build these tools, open their APIs and create an MCP connection, so sophisticated clients can connect their own tools, will keep the platform at the center of the risk workflow. The conversations happening in rooms like this one are the signal to start now.
What I’d Tell Any Risk Manager Right Now
The room’s advice matched what I tell clients, so I’ll give you the combined version.
Start with a specific gap your current tools don’t cover. Broad “use AI everywhere” mandates stall teams.
Keep a human reviewing everything early. Dial back the oversight only after your prompts prove reliable.
Clean your data first. AI can find the flaws, and that alone is worth the exercise, but it can’t fix a fundamentally dirty dataset.
Ask the tool what it can do with your data before prescribing the use case. The non-technical people in these organizations adopted fastest once they saw their own results.
And keep an eye on what your vendor ships, but don’t put your own experimentation on hold while you wait.
One More Thing
The group’s final agreement was to share use cases across organizations rather than guard them as an edge. I think that’s the highest value move available to this industry right now. Buyers sharing openly what they need, and what they’re building when they can’t get it, gives providers the clearest roadmap they’ll ever receive.
I keep a running list of tested AI use cases in risk management. If you want it, ask.
Patrick O’Neill is President and Founder of Redhand Advisors, the leading independent advisory firm focused exclusively on the RMIS and RiskTech market, and publisher of the annual RMIS Report. poneill@redhandadvisors.com | redhandadvisors.com