This piece is by Sasha Haco, CEO & Co-Founder, Unitary
The insurance sector is under mounting pressure to modernise. According to recent industry data 63% of claims professionals point to inefficient systems as a major barrier to effectiveness. At the same time, 56% of industry leaders believe AI’s biggest impact in the next five years will be reducing manual work and admin, and 55% of insurers have already invested in AI or machine learning for claims.
The opportunity for automation is clearly enormous, with potential gains in cost, speed, accuracy and customer experience. Yet according to the MIT State of AI in Business 2025 report, 95% of generative AI pilots fail. Despite businesses pouring billions of dollars into gen AI, the vast majority of pilots never make it into production. The challenge may be particularly acute for insurance, a sector built on assessing risk. The risk is that hype, particularly around “AI agents,” begins to outpace what these tools can safely deliver in regulated, process-driven environments.
AI agents are everywhere right now: heralded as a breakthrough that can automate entire workflows simply by giving a large language model access to tools and natural-language instructions. It’s an appealing idea: no code, no integration work, and the promise of end-to-end automation. But while AI agents are powerful, they are also widely misunderstood. Unless business leaders and policymakers distinguish between where they genuinely work and where they fundamentally do not, companies risk placing big economic bets on the wrong tool for the job.
At their core, AI agents work by interpreting an instruction, predicting the most logical next step based on patterns, and executing that decision through software interfaces or APIs. This makes them well-suited to exploratory and unstructured tasks – research, data gathering, or summarisation – where the goal is to produce a helpful answer, not a perfectly consistent one.
But AI agents come with an inherent property: they are probabilistic, and therefore generate a new prediction each time they act. And while that’s a strength for creativity, it becomes a weakness for operational work. “Probably correct” is not sufficient when 88% of insurance customers say losing human oversight is their biggest concern with AI-driven processes and 70% report mistrust in AI decision-making. Accuracy, auditability and consistency underpin regulatory compliance and consumer protection.
Insurance workflows depend on reliability. Claims handling, underwriting support, policy administration and reporting all involve long chains of deterministic steps where the required action is fixed, not interpretive. A single misstep can have real consequences: an incorrectly updated policy amendment, a misrouted claim, or the wrong document attached to a case file can trigger compliance breaches, customer disputes, or costly remediation work. Even if an AI agent is 95% accurate on each step, small error rates compound across long workflows. A 20-step process with 95% accuracy at each stage has only a 36% chance of being completed correctly from end to end.
The problem isn’t that AI agents are “not ready” or need more training – it is more fundamental than that. AI agents make probabilistic predictions, not deterministic guarantees. If a task has one correct output every time, traditional software will always be the more reliable option.
This distinction between probabilistic tasks that require judgment and deterministic tasks that require consistency is crucial for any leader thinking about automation. Most business processes are a blend of both. Some steps require reasoning that humans have historically had to perform; others are straightforward and rule-based. This is why the conversation around AI agents needs to evolve. These tools will have a role in insurance, but not as a universal automation engine.
The future of automation in insurance is hybrid. The strongest solutions pair deterministic software for steps that must be executed flawlessly with generative AI for moments that require interpretation or expert-level understanding. Logging into systems, locating the right policy, downloading documents and updating records are deterministic tasks. But analysing a medical report or interpreting whether a loss falls within policy terms is far more complex and benefits from AI’s ability to process and structure information. A hybrid approach ensures AI is only used where its reasoning adds value, while software guarantees accuracy and compliance everywhere else, giving insurers the speed and efficiency they want without sacrificing reliability.
For insurers beginning their automation journey, the first step isn’t choosing technology — it’s understanding the workflow. Break each process into steps and identify which ones must follow fixed rules and could be automated, and which require interpretation, and therefore human escalation. When assessing vendors, ask two simple questions: can they clearly separate these step types, and how do they handle errors? The goal isn’t to remove human expertise, but to remove repetitive work so teams can focus on the decisions that matter.
Customers and regulators rightly require predictability, auditability and transparency if AI is to be deployed in environments where accuracy is non-negotiable. As the UK pushes to lead in AI and accelerate productivity gains, we need to ensure the tools we champion are fit for purpose. AI can unlock enormous economic value. But only if we apply it with precision, not blind optimism.
Be the first to comment