SME Claims, the time for reckoning is Now!

Why your data strategy is your claims strategy.

A view from the SME commercial frontline across North America, the UK & Ireland, and Europe — and the case for a 24-month workbench-led transformation.

A letter to the executive committee.

There has rarely been a moment as decisive as this one, having also rarely observed so many programmes confusing motion for progress!

Your SME book is the most underestimated profit engine on your balance sheet. It is also the most operationally fragile. The carriers I admire most have stopped treating claims as a back-office cost-to- serve and started treating it as the only daily customer interaction that genuinely shapes whether a small business renews. The carriers I worry about have responded to that realisation with a procession of pilots: a fraud model here, a chatbot there, an OCR proof-of-concept somewhere in the EMEA region — each individually defensible, collectively forgettable.

Industry data confirms the gap. Roots Automation’s 2025 study found over 90% of insurers were experimenting with AI; only 22% had it deployed at production scale. Sedgwick’s 2026 property report puts AI use across carriers at 58–82%, but only 7% claim scaled success. BCG’s most recent AI Radar shows industry AI spending tripling as a share of revenue — yet just 38% of P&C insurers are generating value at scale from AI in core workflows. The translation: capital is being spent. Compounding returns are not being earned.

The translation problem in our industry is not technical. It is strategic. Pilots scale when data does. Models compound when governance does. Workbenches transform when operating models do.

This briefing is the case for changing how you frame the question. The carriers winning the next cycle will not be those who buy the cleverest models. They will be those who treat the claims handler’s desktop as the destination, the data foundation as the engine, and a 24-month transformation arc as the disciplined commitment that gets them both.

The strategic challenge — five forces converging on SME claims.

There is no single problem to solve. There is a convergence — and like all convergences, it rewards those who see it as a system.

1. The economic floor under SME claims is rising.

Across North America, social inflation, third-party litigation funding, nuclear verdicts on commercial auto and umbrella, and severe-convective-storm losses have pushed loss costs above premium-rate growth in several casualty lines. S&P Global Market Intelligence flags Commercial Multi-Peril Liability at a 114.9% combined ratio, Commercial Auto Liability at 113.0%, and Other Liability at 110.1% — these are loss- making lines that demand structurally lower expense ratios to survive the next cycle. UK and Irish carriers face their own pressures: PolicyBee research from 2025 found 88% of UK commercial sites underinsured on building values, and the underinsurance gap is now showing up in average-clause disputes that drive complaints, voids and customer attrition. In Europe, persistent NatCat severity and the long shadow of cladding remediation continue to inflate property severity. SME claims inflation is not a temporary excursion; it is the new operating context.

2. SME customer expectations have rebased — permanently.

Your SME insureds run their businesses on Stripe, Square, Xero, Shopify and the Microsoft 365 stack. They expect a claims experience that resembles those products: instant acknowledgement, transparent status, predictable next-step communication, and self-service for the simple things. The 2025 J.D. Power U.S. Small Commercial study confirmed retention is dropping across virtually every segment, and that satisfaction tracks how well the carrier communicates — not premium movement. The implication: the customer-experience moat in SME is now built or lost in claims.

3. The data is everywhere and the data is nowhere.

Every carrier I work with has more data than it did five years ago. Few have meaningfully better data. SME claims data sits across Guidewire/Duck Creek/Majesco-class admin platforms, third-party- administrator bordereaux, broker-managed inputs, OCR’d documents that nobody trusts, and dozens of point-solution stores. Lineage is partial; definitions are inconsistent across LOBs; quality is inferred rather than measured. This is the silent tax on every model your data-science team builds — and the reason 70- plus percent of AI projects underperform their business case. The EU AI Act, fully applicable from 2 August 2026, codifies this: Article 10 demands data governance, representative training datasets and demonstrable provenance for any high-risk system. The cost of poor data is no longer just operational — it is regulatory.

4. IoT and connected-risk are no longer optional inputs.

Met Office and Met Éireann APIs, satellite-derived hail-swath confirmation, telematics in fleet, water-leak sensors in commercial property, refrigeration telemetry in restaurant equipment, and connected fire panels in larger SME premises — these are now operational claims signals, not innovation-team

experiments. They reduce notification lag, improve causation evidence, and shrink leakage. The carriers building IoT-aware claims now will, by 2027, have a structurally lower combined ratio than those who treat sensors as a future agenda item.

5. Talent is leaving — and not coming back in the same shape.

The senior-handler retirement wave is real, the apprentice pipeline is thin, and the cost of replacing a tenured Validation handler is rising faster than wage inflation. McKinsey’s Claims 2030 work makes the case that the role itself is bifurcating: high-judgment portfolio strategists at one end, and AI-supervised low-complexity processors at the other. The middle is hollowing. Your operating model needs to plan for the workforce you will have, not the one you wish you still had.

These five forces are not independent. They compound. A weak data foundation makes IoT signals unusable; an unusable IoT signal degrades triage; degraded triage drives leakage; leakage degrades the combined ratio; combined-ratio pressure forces tighter staffing; tighter staffing erodes customer experience; eroded experience accelerates SME attrition. This is the loop you must break.

The economic prize — sized honestly.

Let me put numbers in front of you that are defensible across a board-room conversation. These are not vendor figures. They are triangulated from BCG, McKinsey, Sedgwick, S&P Global and our own engagement experience.

Operating-cost reduction per dollar of premium, AI- redesigned claims (BCG, 2026)

Median time to ROI when KPIs are embedded from day one (Roots, 2025)

Combined-ratio improvement for AI leaders vs laggards (WTW 2026, McKinsey)

For an SME book of $1bn GWP across UKI and selected European territories, our reference engagements suggest a 24-month value pool in the range of £45m–£90m of recurring annual benefit, comprising loss- cost leakage recovery (35–45%), expense reduction (25–35%), retention uplift (15–20%) and fraud- savings net of investigation cost (10–15%). The numbers compound. Year-three benefits typically run 1.4– 1.6× year-two. This is not a savings story. It is a structural-margin story.

But — and this is the part most business cases skim — these benefits do not materialise from buying tools. They materialise from changing how the claims function actually works at the desktop level, every day, for every claim. Which is precisely where a workbench-led approach matters.

Why a workbench changes the equation.

A claims workbench is not a UI project. It is the architectural place where the data strategy, the AI strategy, the regulatory strategy, the operating model and the customer-experience strategy converge into the daily working pattern of the handler. If those five things converge cleanly at the desktop, your transformation will compound. If they don’t, you will end up with a portfolio of pilots that look impressive in the steering committee and disappear by quarter-end.

HELIX™ Claims AI — the workbench reference architecture we have been building with insurance clients across UKI and North America — illustrates the principle. Every claim flows through a defined pathway (Standard Simple, Standard Complex, or Delegated TPA). Each stage gates the next: a handler cannot submit FNOL until every AI-recommended action has been Accepted or Declined. Each stage has a stage- specific Fraud Console check, a per-claim audit trail, and an integration registry that shows the handler exactly which external data was pulled, when, and what it returned. Coverage validation, warranty compliance, recall monitoring, CIDRA fair-presentation, underinsurance/average, sanctions and CUE all live as native engines, not branded third-party plug-ins.

Why does this architecture matter strategically? Because it operationalises governance. Every Accept/Decline by the handler is a documented human-oversight event. Every external data pull is logged with timestamp, payload and rationale. Every model decision is traceable to the data that fed it. This is what the EU AI Act calls for under Article 14 (human oversight), Article 12 (record-keeping) and Article 13 (transparency). And it happens not because the carrier added a compliance overlay — it happens because

the workbench was designed that way from inception. Compliance becomes a by-product of good operating design, rather than a parallel programme.

Today — pilots-led Tomorrow — workbench-led

Models live in data-science notebooks, called by handlers via separate tools and Slack messages.

Models are wired into the handler’s daily workflow at the right stage, with explicit Accept/Decline.

Audit trail is reconstructed from ticketing systems and email — fragile under regulatory request.

Audit trail is the workbench. Every action logged, time-stamped, attributed.

Coverage validation is a tribal knowledge artefact held by senior handlers.

Coverage AI runs clause-by-clause, with handler override and exception-management governance.

Fraud is a separate console used reactively by SIU. Fraud check is stage-specific, embedded in every claim, accept/decline like any other recommendation.

Stage advancement is implicit, undocumented, and inconsistent across handlers.

Stage submit is gated, evidence-based, and produces an immutable trail at every transition.

Pathways are theoretical — every claim notionally moves through every stage.

Pathways are real — Standard Simple, Standard Complex, Delegated TPA. N/A stages are explicitly N/A.

Data strategy at the heart of it all.

If you change one thing in your AI agenda this year, change this: stop putting models before the data they depend on. The discipline that compounds is the inverse — Discover → Validate → Augment.

This is the framework I argue for in every claims-transformation kickoff. It sounds simple. It is. The reason most carriers don’t practise it is because the alternative — buying a model — feels faster. It isn’t. It is faster only to the first proof-of-concept. After that, the unvalidated data slows everything.

Discover — what you actually have, by claim, by stage, by source.

Most carriers cannot, today, produce on demand the answer to: ‘Which external data sources were pulled for claim X at stage Y, when, and what did they return?’ The HELIX integration registry shows this. So should yours. The Discover phase of your data strategy starts with a complete inventory: every internal claims data store, every third-party feed (CUE, MIAFTR, MID, askCUE PI, Cifas, Synectics, Met Office, OPSS, FSA, MHRA, RAPEX, NICEIC, BAFE, Gas Safe, FORS, FRA, RICS, BIBA, Mandiant, Chainalysis, Companies House, FCA Register, CBI Register, ICO/DPC, NCSC, Action Fraud, OFSI), every IoT/telemetry input. Map them to the stages where they are operationally valuable. Anything not used at a stage is a candidate for retirement.

Validate — provenance before performance. Validate means lineage: you know where every data point came from, when it was last refreshed, and whether the source is authoritative for the use case. Validate also means quality measurement:

completeness, accuracy, timeliness scored at the field level, with operational SLAs against each. The mistake we keep seeing is teams optimising model F1 scores against data of unknown provenance. EU AI Act Article 10 requires this discipline; it should also drive better outcomes regardless of regulation. A coverage-validation AI trained on inconsistently coded peril fields will silently mis-classify edge cases for years before anyone notices. A warranty-compliance check trained on document data with broken OCR confidence scores will let real warranty breaches through. Validate first.

Augment — and only then.

Augment is where AI earns its keep. With validated data, augmented intelligence becomes safe to deploy: coverage validation that genuinely understands the policy schedule; recall monitoring that cross- references OPSS, FSA, MHRA and RAPEX in real time against the insured’s product portfolio; fair- presentation (CIDRA 2015) checks that compare disclosure at inception against claim facts; underinsurance modelling that uses RICS BCIS rebuild benchmarks against declared TIVs. Each of these moves leakage, cycle time and customer experience simultaneously. And each is unsafe without the prior two phases.

Governance before platform. Depth before breadth. Golden source as policy, not technology. AI after validation. These are not abstractions — they are the four sequencing rules that separate carriers who scale from carriers who launch.

Execution roadmap — 24 months, three phases, one workbench.

Is 18-24 months realistic? Let me address this head-on, because timeline credibility is where most boards lose patience with claims transformation. The honest answer: 18 months is achievable if you start with one or two LOBs and accept that specialty/long-tail extension comes in year three. 24 months is achievable for full SME-book coverage. Anything shorter is a marketing claim, not a delivery plan. Anything longer is a sign that governance is being treated as a parallel workstream rather than embedded into delivery.

Industry evidence supports this band. Vantage Point benchmarks point to ROI within 12–18 months on phased deployments. McKinsey’s digital-insurer playbook calls for 20–100 specialists hired in the first 18 months for transformations of this scale. Five Sigma deployments are delivered in months not years on individual modules — but full enterprise rollouts always extend to 18–24 months when properly governed. The shape below is what we have seen work.

Three sequencing principles make this roadmap real, not aspirational. First, Phase 1 must produce real production value in the first six months — not just foundational data work. The handler-visible workbench v1, even covering a narrow set of LOBs, is what builds organisational belief. Second, Phase 2 is the hardest. It is where data validation discipline must hold against the temptation to bolt on more models. Resist that temptation. Third, Phase 3 is where the moat compounds. Specialty extensions, broker-channel differentiators, embedded SME claims and predict-and-prevent IoT plays are only credible once the foundation is real.

The operating model that supports it.

A workbench without an operating model is just better software. The carriers who win are explicit about three operating-model shifts that the workbench enables and amplifies.

Shift 1: Pathway-based handler economics.

Stop measuring all handlers against a single productivity benchmark. Standard-Simple claims are a volume game where the workbench’s gating, recall integration and STP-eligibility logic should drive 35– 55% straight-through processing within 18 months. Standard-Complex claims are a judgment game where the handler’s value is exception management, narrative coherence and counter-party negotiation. Delegated-TPA claims are a governance game — the carrier’s value is bordereau intelligence, leakage audit and SLA enforcement, not adjudication. Three pathways, three handler personas, three distinct economic models. One workbench, three sets of measures.

Shift 2: Stage-based governance, not function-based governance.

Most carriers govern by function — a coverage committee, a fraud committee, a complaints committee. The workbench inverts this. Each stage transition is a governance event. FNOL → Triage requires confirmed coverage status. Triage → Validation requires fraud screening. Validation → Adjudication requires evidence-pack completeness. Each gate has a documented rationale, an Accept/Decline trail, and an integration ledger. Function committees become exception-handling forums for the cases the

workbench surfaces, not the routine review bodies they are today. This is leaner, faster, and dramatically more defensible to regulators.

Shift 3: Workforce repositioning, not workforce reduction.

The cleanest carriers I know have stopped framing AI claims investment as headcount reduction. The maths doesn’t support it anyway — the senior-handler retirement wave will absorb most of the so-called ‘savings’. Instead, the workforce is repositioned: tenured handlers become portfolio strategists, complex- loss specialists, broker-relationship anchors and AI-assurance leads. Mid-career handlers shift into pathway-management, exception-handling and SIU. Junior handlers run the STP queue with workbench supervision. New roles emerge: AI assurance, model-risk, integration-bus stewardship, data-quality engineering. This is the credible workforce story for the next cycle.

The headline isn’t ‘AI replaces handlers.’ The headline is ‘AI lets your tenured handlers stop drowning in low-judgment work and start anchoring the relationships and decisions only humans can.’ Frame it that way internally. Mean it.

Risks — what keeps us awake!

No partner who has lived through three of these programmes pretends the risks are theoretical. There are five I would hold close to your CRO and CDO.

• Pilot proliferation as substitute for strategy. — The most expensive failure mode. Pilots feel safer than commitments. They aren’t — they merely defer the cost of indecision. Force one workbench thesis and the courage to push everything else through it.
• Brand and partner dependence on third-party InsurTech vendors. — Many carriers have woven Sprout.ai, Shift Technology, CLARA Analytics, Tractable and others into their claims stack. Each vendor decision is defensible; the cumulative dependence is dangerous. Build native engines for the operations you intend to differentiate on (coverage, warranty, fraud, recall, fair presentation). Buy where commodity (OCR, payments, sanctions). Avoid the middle, which is where lock-in and brand- leakage compound.
• EU AI Act timing risk. — 2 August 2026 is a hard deadline for high-risk system compliance. Insurance pricing for life/health is named explicitly in Annex III; commercial claims systems used for automated decisioning will be assessed against the same standards. Carriers running active claims AI without an Article 10/13/14/15 evidence pack by Q3 2026 face fines up to €15m or 3% of turnover. This must be tracked as a board-level risk.
• Model drift and hallucination in generative AI co-pilots. — MLOps maturity is uneven across the industry. Without continuous evaluation, drift detection, fallback controls and a clear human- override pattern at every stage of the workbench, generative components will silently degrade. Build the MLOps layer before scaling generative use.
• Talent — and the cost of getting it wrong. — The right team for an 18-24 month claims transformation is small (15-30 senior people for a $1bn book), expensive, and difficult to recruit. Underinvesting in this team is the most reliable way to extend a 24-month plan into a 48-month one. Hire the leadership team first; everything else follows.

A closing word.

Ending where I started! The strategic question for your SME claims book is not whether AI matters — it does, and the evidence is now overwhelming. The strategic question is whether you will treat the claims workbench as the integrating destination for your data, AI, regulatory and operating-model agendas, or whether you will continue to treat them as parallel programmes and hope they converge.

They won’t converge on their own. They converge through deliberate architecture. They converge at the handler’s desktop. They compound when governance, data and AI are built into the same daily working pattern. They differentiate when you back them with a 24-month commitment, sequenced honestly, sponsored visibly by the CEO, and measured ruthlessly against handler-experience and customer- experience KPIs from day one.

Your SME book is not a problem to be managed. It is the place where the next cycle of carrier differentiation will be won or lost. The carriers I expect to be quoting back to me in 2028 will be the ones who, in 2026, looked their executive committee in the eye and said: this is our 24-month commitment, this is our workbench thesis, this is our data strategy, and this is the claims experience we are building for the small businesses we exist to protect.