The Cost Ceiling of Adversarial AI in Insurance — And the Gating Trick That Breaks It
Self-critiquing AI agents are more accurate. They are also nearly twice as expensive to run. For an industry that measures margin in basis points, that math has to change — and it is starting to.
By Dr. Anand R Fadte
CTO – Aviantos | 26th August 2026
Every insurance executive who has spent thirty minutes with a modern AI agent comes away with the same two impressions. The first is that it is remarkable how well the system reasons. The second is that it is remarkable how confidently it is sometimes wrong.
The industry’s answer to the second problem is a technique called adversarial self-critique. It is powerful, it works, and — until recently — it was too expensive to run at insurance scale. The story of how that changed is the story of one small architectural idea with disproportionately large consequences.
Why Self-Critique Matters in the First Place
An AI agent that produces an underwriting recommendation, a claims decision, or a fraud flag can do so with any level of confidence, correctly or otherwise. The uncomfortable truth about large language models is that they do not natively know when they are guessing. A model will cite a regulatory rule that does not exist with the same tone it uses to cite one that does.
Adversarial self-critique is the standard fix. A second AI agent — usually a separate model instance with a different prompt — reviews the first agent’s recommendation, actively looking for reasons it might be wrong. It checks that the cited rules are real, that the risk factors are supported by the applicant’s data, that the premium loading is consistent with the rating table, and that no material fact has been ignored. If the critic disagrees, the recommendation is revised.
The technique works. Published results across commercial underwriting benchmarks show accuracy improvements of five to ten percentage points and hallucination reductions of seventy to eighty percent when critique is applied. In a domain where a single incorrect underwriting decision can cost a carrier more than the entire annual AI budget for that line of business, those numbers matter.
The Cost Problem Nobody Talks About
There is, however, a small detail that has quietly kept most of these systems in the research-paper stage: adversarial critique roughly doubles the cost of every decision. A widely-cited figure in the recent literature is around 90% computational overhead per case. In practice this means a critique-enabled agent runs in roughly the time and at roughly the token cost of two agents.
For high-value commercial lines — a specialty aviation policy, a large corporate property placement, a complex D&O renewal — that overhead is trivial next to the value of the decision. Doubling the cost of an AI system that is deciding on a policy with a seven-figure premium is a rounding error.
For everything else, it is disqualifying.
Most insurance decisions are not seven-figure policies. They are personal auto quotes, small-commercial renewals, and standard health applications. Per-case processing costs sit in the low single digits of dollars, sometimes lower. A 90% overhead does not add a rounding error to those decisions — it eliminates the business case entirely. This is the reliability-cost trade-off that has kept adversarial critique out of production at most global carriers, despite everyone agreeing it works.
The Gating Trick
The insight that resolves the trade-off is disarmingly simple. Most of the time, the primary agent is not wrong. On straightforward cases — the fit thirty-year-old with a clean claims history buying a standard homeowners policy — the agent’s recommendation is correct, the rules are simple, and there is no meaningful hallucination risk. Sending that case to a second agent for critique is pure waste.
The hallucination risk concentrates in a minority of cases: the complex, the ambiguous, the ones where multiple rules apply, or where the applicant profile sits at the edge of the training distribution. If you could identify those cases in advance and route only them through critique, you would capture most of the accuracy benefit at a fraction of the cost.
This is what confidence-gated critique does. The primary agent produces its recommendation together with a set of confidence scores — one for each dimension of the decision. If every score exceeds a calibrated threshold, the recommendation passes directly through. If any score falls below the threshold, the critic is invoked. The threshold itself is not a hyperparameter to guess — it is calibrated against held-out data to hit a target reliability level.
In a recent evaluation on more than thirteen hundred health-underwriting cases, this architecture activated the critic on roughly one case in five. The result was an average computational overhead of under seven percent — approximately one-thirteenth of the cost of continuous critique — while retaining most of the accuracy improvement and roughly three-quarters of the hallucination reduction.
The trade-off that had priced adversarial critique out of the mass market did not, it turned out, need to be a trade-off at all. It needed to be a routing problem.
What This Changes for Global Insurers
The immediate implication is that the reliability techniques previously reserved for high-value specialty lines can now be applied across mass-market portfolios — personal lines, small-commercial, standard health — without breaking their economics. For reinsurers, brokers, and MGAs building AI-enabled offerings, this is the shift that makes the business case work.
The second implication is subtler. Confidence gating turns the AI system from a monolithic decision engine into something more like a triage architecture — the same shape that human underwriting has always had. Simple cases go to the fast path. Complex cases receive additional scrutiny. Cases at the extremes escalate to a senior human. This is not a novel operational pattern in insurance; it is the operational pattern in insurance. The AI system is finally learning to mirror it.
The third implication is regulatory. Explicit confidence scoring and threshold-based escalation are exactly what supervisors from EIOPA to the PRA to the NAIC have been asking for in their guidance on AI in financial services. A system that says “I am confident here, and here is why I escalated there” is a system a regulator can approve. A monolithic model that emits a single number is not.
Where the Trick Runs Out
Gating is not a silver bullet. If the primary agent is systematically overconfident on cases it gets wrong — a well-documented failure mode of large language models — the gating mechanism will let those errors through. Calibration of the confidence signal is therefore not optional; it is the load-bearing element of the whole architecture. Any insurer deploying this pattern needs to invest in confidence-score calibration with the same seriousness they invest in the underlying model.
The other limit is high-value lines, where the economics are the reverse of mass-market insurance. For a corporate property placement or a specialty aviation policy, continuous critique is still worth it — the accuracy gain from critiquing everything outweighs the cost. Gating is not a replacement for full critique in high-stakes contexts; it is a way to extend the benefits of critique to contexts where full critique is uneconomic.
The Real Lesson
The wider point, for insurers thinking about how to architect their AI stack, is that the cost-versus-reliability curve in enterprise AI is not fixed. It bends when someone asks the right architectural question. The question here was not “how do we make critique cheaper” — it was “do we need to critique everything?” The answer, empirically, was no.
There will be more of these questions in the next twelve months. Which parts of the reasoning trace need to be logged? Which classes of case need the largest model? Which decisions can be handled by a smaller, specialised agent rather than a general-purpose one? Each of these questions has a monolithic answer that is expensive and a routing-based answer that is roughly right and roughly free.
The insurers that ship agentic AI at portfolio scale over the next five years will not be the ones who found the smartest model. They will be the ones who asked the right routing questions first.
— — —
About the author
Dr. Anand R. Fadte is Chief Technology Officer at Aviantos and a technology leader specializing in enterprise AI, Generative AI, agentic AI, and AI solution architecture. He writes on AI strategy, enterprise AI architecture, LLM applications, and responsible AI adoption, with a particular focus on regulated industries such as insurance, banking, and fintech. With 22+ years of global technology leadership experience and 40+ awards and recognitions, he brings a practitioner’s perspective to building and scaling AI solutions that deliver measurable business value.
Connect with him on
LinkedIn: linkedin.com/in/anandf-ai-head
Website:www.fadte.comv
