Why ‘Lots of Pilots, Limited Production’ Is the AI Problem Killing Underwriting ROI
A recent industry report on AI adoption in insurance landed on a sentence that should be pinned to the wall of every technology team in the sector: ‘The pattern is consistent across carriers. Lots of pilots, limited production, minimal P&L impact.’
It’s a damning summary of where the industry actually is versus where the marketing materials suggest it should be. The gap is not a technology problem. The tools exist. The models work. The gap is operational — and it’s costing carriers real money every quarter that AI sits in a proof-of-concept stage rather than embedded in a live workflow.
This article explains why the pilot trap is so persistent, what workflow-first deployment actually requires, and how stop-loss carriers specifically can move from demo to production.
Why Pilots Don’t Become Production
The typical AI pilot in underwriting goes like this: a vendor demonstrates an impressive document extraction tool or a risk scoring model. The demo uses clean data, a favourable use case, and a controlled environment. Results are good. The carrier signs a pilot agreement.
Three months later, the pilot is technically ‘successful’ — the tool did what it was supposed to do in the test environment. But it hasn’t been connected to the submission intake system. It hasn’t been integrated with the rating engine. It hasn’t been trained on the carrier’s proprietary data. The governance framework for how underwriters are supposed to interact with its outputs hasn’t been written. And because none of that infrastructure exists, the AI tool runs in parallel with the existing workflow rather than replacing it.
The underwriters still manually re-key data. Still manually apply rating logic. The AI produces outputs that get reviewed, occasionally used, and frequently set aside. The P&L impact is zero. The vendor gets paid. The cycle repeats.
< 50%
Share of insurance businesses that have deployed AI in even a single function — despite years of ‘AI transformation’ investment. (Simplifai, 2025)
The Model-First vs Workflow-First Distinction
The companies that have escaped the pilot trap share a common characteristic: they started with the workflow, not the model.
Model-first deployment asks: ‘What can AI do?’ and then looks for places to apply it. The result is impressive demos that solve isolated problems without connecting to adjacent steps in the process.
Workflow-first deployment asks: ‘Where does time and accuracy get lost in our underwriting process?’ and then builds AI to address those specific points. The result looks less glamorous but produces measurable operational change.
For a stop-loss carrier, the workflow-first question has a fairly consistent answer: time is lost at census extraction, at data normalisation, at rule application, and at proposal generation. Those four steps account for 60–70% of underwriter time on a typical submission. An AI system that addresses all four — connected end-to-end rather than as standalone modules — produces the throughput improvement that shows up in quote volume and turnaround time.
The Governance Requirement That Nobody Talks About in Demos
The second reason pilots stall is governance. Regulators, reinsurers, and internal risk teams are increasingly asking the same question about AI-assisted underwriting decisions: can you explain why the system priced this group this way?
A black-box model that produces a risk score without traceable logic is not deployable in a regulated underwriting environment. It creates liability exposure if a pricing decision is challenged. It creates reinsurance audit problems. And — critically — it creates resistance from underwriters who don’t trust outputs they can’t interrogate.
The AI systems that are actually getting deployed in production are explainable. Every rule, every adjustment factor, every data input is logged and attributable. When an underwriter or a regulator asks ‘why was this group rated 15% above standard?’, the system can produce an answer at the rule level.
This is not a complex technical requirement. It is an operational discipline requirement — and it’s one that has to be built into the system architecture from the start, not retrofitted after the model is running.
What Production-Scale Deployment Actually Looks Like
A stop-loss carrier operating AI at production scale in 2026 looks like this:
A broker submits a census file. It arrives in any format — Excel, PDF, CSV, whatever the broker uses.
An automated extraction layer pulls employee demographics, plan design, and prior claims in under 60 seconds. No manual re-keying.
A rules engine applies the carrier’s underwriting logic — industry class, age band, geographic adjustments, GLP-1 load, laser criteria — automatically and consistently.
A risk scoring model flags submissions that require human review versus those that can proceed to proposal generation.
For clean submissions, a proposal is generated automatically. For flagged submissions, an underwriter reviews only the specific risk factors that triggered the flag.
Every step is logged, auditable, and attributable.
The underwriter’s role in this model is not eliminated. It is elevated. Manual work on routine submissions is removed. Judgment is applied to the cases that actually require it — complex risk factors, unusual plan designs, high-acuity clinical flags.
The business outcome: carriers running this model are quoting in 36 hours or less on standard submissions. They are processing 30–40% more quote volume with the same headcount. Their competitive win rate with time-sensitive brokers improves because they are consistently first to respond.
The Starting Point
If your organization is in the pilot trap, the exit is not a bigger AI investment. It’s a workflow audit. Map every step from submission receipt to proposal delivery. Identify where time is consumed by manual processes that don’t require underwriter judgment. Build automation for those steps first, connected end-to-end. That’s the infrastructure on which production-scale AI runs. The models follow naturally once the workflow is instrumented.
DataHub was built workflow-first for stop-loss carriers. SmartExtractor™, SmartRules, and SmartProposal are connected modules in a single platform — not standalone point solutions. Deployment from discovery to production takes 16 weeks.
Contact us to schedule a workflow audit call.