Short answer: Agentic AI fails in insurance for the same reason nearly every insurance transformation fails: the data foundation was never built to support autonomous decision-making. Only 22% of insurers plan to have an agentic AI solution in production by year-end 2026, and just 7% have successfully scaled any AI initiative across their organisation. The gap between those two numbers is a data engineering problem, not an AI problem, and it is solvable with the right sequencing.
Why agentic AI in insurance depends on data engineering, not model choice
Insurance is an unusually good fit for agentic AI on paper. It is a data-dense industry with structured actuarial processes, repeatable workflows, and clear rules to reason over. In practice, adoption is stalling well before the model layer. Evident’s Q4 2025 Insurance AI Use Case Tracker found that 68% of publicly disclosed insurance AI deployments were generative or agentic, with agentic AI accounting for 21% of deployments, which sounds like healthy momentum until you compare it against how few of those deployments ever leave the pilot stage.
The bottleneck is consistent across every serious industry report on the topic: legacy core systems, fragmented data environments, and a lack of in-house AI and data engineering expertise. An agent that has to reason across a policy admin system, a claims platform, and three spreadsheets, none of which agree with each other, is not an agentic AI problem to solve. It is a data integration problem wearing an AI costume.
Where UK insurers stand
The UK is further along than most markets on point deployments, but not on enterprise-wide scaling. Direct Line, Aviva, and Admiral all report claims automation rates above 60% on motor, and Aviva became the first major UK insurer to offer home cover quotes inside ChatGPT. But Earnix’s 2026 Insurance Industry Trends Report describes UK adoption as more targeted than global peers: AI is already embedded across workflows, but more often partially integrated rather than fully scaled. That is the operational divide showing up in the data itself. A 2026 AutoRek survey of UK and US insurance managers found that insurers manage an average of 17 separate data sources feeding their premium processes alone, with two-thirds handling more than 10.
Lloyd’s own written evidence to Parliament captures the market’s internal split on pace: 35% of Lloyd’s market participants expect rapid expansion of technologies like AI, while 65% expect adoption to be more gradual. Regulation is not waiting for that split to resolve. UK insurers now have to build any agentic deployment to satisfy FCA and PRA supervision, Solvency UK, UK GDPR, and Consumer Duty simultaneously, which means governance and auditability cannot be an afterthought bolted on once an agent is already live.
Where South African insurers stand
South Africa’s insurance sector is earlier in the curve and more cautious than UK peers. According to EY’s South Africa 2026 insurance outlook, AI use cases already exist across the market, chatbots, OCR, fraud detection, but remain largely functional rather than enterprise-wide, with scaling advanced underwriting models, behaviour-led pricing and agentic AI remaining limited. The gap with adjacent financial services is stark: South African banks are adopting AI at 52%, while the insurance sector sits at just 8%.
That caution is not simply conservatism. It tracks the same root cause seen in the UK, just less resourced: EY points to persistent data quality challenges, particularly in older bancassurance systems, as a genuine constraint on what agentic AI can safely do. Where South Africa is ahead is InsurTech collaboration, especially in microinsurance and embedded, mobile-driven models, which gives local insurers a distribution advantage the underlying data infrastructure has not yet caught up to. As one South African insurance executive put it in a 2026 industry roundup, the shift underway is that AI has moved from pilot projects to full enterprise-wide adoption across the businesses doing it well, which is precisely the transition most SA insurers have not yet made.
The roadmap: sequencing agentic AI the way FloJo is built to
Teraflow’s FloJo operating model exists for exactly this situation: taking a regulated enterprise from fragmented data and stalled pilots to production-grade agentic systems, in a sequence that does not require insurers to rip out core systems they cannot afford to touch. For insurance specifically, that sequence looks like this.
| Phase | What happens | Why it has to come first |
|---|---|---|
| 1. Data foundation audit | Map every data source feeding underwriting and claims, score completeness, accuracy and lineage against IFRS 17 and Solvency II auditability requirements | You cannot govern what you cannot trace, and most insurers cannot currently produce granular claims data from a single source without manual reconciliation |
| 2. Sidecar integration layer | Build an API and orchestration layer over legacy policy admin and claims systems rather than replacing them | InsTech’s analysis of underwriting orchestration engines found the strongest implementations start with a narrow, measurable use case rather than an enterprise-wide redesign, because rip-and-replace timelines kill momentum before value is proven |
| 3. Narrow-scope agent deployment | Deploy agents against one line of business or one workflow (motor claims triage, commercial submission intake) with human-in-the-loop review on every decision | Proves the data foundation actually holds under real transaction volume before governance risk is extended further |
| 4. Governance and observability | Build bias detection, decision audit trails, and drift monitoring into the agent layer itself, mapped to FCA/PRA requirements in the UK and the Draft National AI Policy in South Africa | The EU AI Act’s provisions for high-risk systems become fully operational from August 2026, and both UK and SA regulators are moving in the same direction on explainability |
| 5. Scaled orchestration | Extend from single-agent workflows to multi-agent orchestration across underwriting, claims and servicing, with each agent’s data access and escalation path explicitly defined | This is where the return on the first four phases compounds, and where most insurers who skip ahead to this stage without doing phases 1 to 4 first end up back at square one |
The point of sequencing it this way is that phase 1 is unglamorous and phase 5 is the one every board wants to talk about. Skipping to phase 5 is exactly how insurers end up among the 93% who never successfully scale an AI initiative in the first place.
What this actually buys you, once the foundation holds
The return on getting the sequencing right is well documented once insurers reach it. AI-powered claims automation is resolving claims 75% faster, with overall resolution time cut from 30 days to 7.5 days, and routine claims processing reduced from 7 to 10 days down to 24 to 48 hours. Straight-through processing rates in AI-enabled operations have jumped from a baseline of 10 to 15%, up to 70 to 90%. On the underwriting side specifically, the data engineering discipline matters as much as the model: modern MLOps practices applied to underwriting pricing models have cut time to production from nine months to six weeks for well-resourced teams, which is the single clearest illustration that the constraint was never the algorithm.
Capgemini projects that AI agents could generate up to $450 billion in economic value across the insurance sector by 2028 through revenue growth and cost savings, and hyperexponential’s benchmarking of commercial P&C carriers running agentic underwriting systems shows quote-to-bind time reductions of 60 to 99%. None of those numbers are available to an insurer that has not first done the unglamorous work in phases 1 and 2.
Frequently asked questions
What is agentic AI in insurance, specifically? It refers to AI systems that can take multi-step action autonomously within a workflow, extracting data from a submission, checking it against internal and third-party sources, triaging it by risk and complexity, and routing it for automated decisioning or human escalation, rather than simply answering a question or summarising a document the way earlier generative AI tools did.
Why do most agentic AI pilots in insurance fail to scale? The consistent pattern across UK and South African data is the same: fragmented legacy data environments, unclear data lineage, and a lack of governance built in from the start. Only 7% of insurers have successfully scaled any AI initiative across their organisation, and the gap is overwhelmingly a data and operating-model problem rather than a model-capability problem.
Do we need to replace our core policy admin system before deploying agentic AI? No, and trying to is one of the most common reasons transformation programmes stall. A sidecar integration layer, an API and orchestration layer built over the existing core system, lets insurers deploy agents against real workflows without a high-risk rip-and-replace migration.
How long does a first agentic AI deployment realistically take? Well-scoped, narrow deployments can go live in six to eight weeks once the data foundation audit is complete. Full-scale deployment across an entire claims or underwriting lifecycle, including regulatory approval and change management, typically takes 18 to 24 months.
What’s different about doing this in South Africa versus the UK? UK insurers are generally further along on point deployments but are constrained by FCA, PRA, Solvency UK and Consumer Duty governance requirements layered on top of legacy systems. South African insurers are earlier in the curve, with insurance-sector AI adoption sitting at 8% against 52% for banks, and the constraint is more fundamental data quality in older bancassurance systems. Both markets converge on the same answer: the data foundation has to be built before the agent layer, not alongside it.
Sources
- Celent: Shedding Light on Agentic AI in Insurance
- Microsoft and Cognizant: Agentic AI adoption in insurance
- Evident Q4 2025 Insurance AI Use Case Tracker, via Newton Media
- Tommaso Maria Ricci: AI for Insurance Companies, 2026 Costs, Rules and Roadmap
- Earnix: UK Insurers Scaling AI with Governance
- Insurance Business UK: AI ambition and manual reality (AutoRek research)
- Lloyd’s of London written evidence to UK Parliament
- EY: South Africa’s 2026 insurance outlook
- News24: SA insurance faces a massive AI shakeup
- Bizcommunity: Insurance industry pivots amid AI boom in 2026
- InsTech: Underwriting orchestration engines
- Datagrid: 42 Insurance AI Agent Statistics
