Insurance AI use cases prioritize on five quantified metrics - loss ratio, expense ratio, combined ratio, growth lift, and time-to-value - and then resolve into three execution paths: build, buy, or partner. The 2026 evidence base is robust: McKinsey's January 2026 update measured 18.5% loss-ratio delta between top-quartile AI-deployed personal lines carriers and the median; Aite-Novarica's 2026 claims survey measured 20-25% LAE reduction at carriers running Tractable plus Five Sigma plus Shift in coordinated production; Deloitte's 2026 insurance benchmark measured 30-50% claims leakage reduction at carriers with mature AI claims discipline; AM Best's April 2026 survey measured 60-70% UW data-extraction time savings at carriers with Hyperscience or Indico in production against ACORD form intake. These are not aspirational ranges - they are observed deployments at named carriers in 2025-2026 production. The prioritization discipline anchors every business case to one of these five metrics, picks the right vendor or build path on the basis of TCO and strategic differentiation, and produces a portfolio of use cases the CRO can defend to the board and the AM Best analyst can interpret against peer benchmarks. This lesson is the use-case scoring template, the build-vs-buy-vs-partner decision tree, and the 3-year TCO model that converts industry-percentage claims into carrier-specific dollar commitments.

The Five-Metric Anchor - Loss, Expense, Combined, Growth, Time

Every AI use case must be anchored to one or more of five metrics. Use cases that cannot be anchored are deferred until a clear metric emerges. The five metrics are loss ratio (losses divided by earned premium), expense ratio (operating expenses divided by net written premium), combined ratio (the sum, the carrier's headline P&L health), growth lift (top-line written premium change driven by AI-enabled distribution or risk-selection), and time-to-value (how fast the use case ships to production and generates measurable impact).

The 18.5% loss-ratio delta number from McKinsey's January 2026 update is the headline benchmark for top-quartile AI deployment in personal lines. The delta is between carriers running mature AI in pricing, underwriting, claims, and fraud against the median peer. Decomposed, it breaks roughly into: 4-7 points from better risk selection at intake (Cytora-class triage plus appetite scoring), 3-5 points from AI-assisted pricing accuracy (Akur8 or Earnix or Guidewire Predict), 4-6 points from claims-side leakage reduction (Tractable, CCC, Shift), and 3-5 points from fraud capture (Shift, ISO ClaimSearch integration, network analysis). Mid-market commercial carriers typically achieve 8-12 points of the 18.5% delta with full deployment because commercial books have higher idiosyncratic risk and AI signal is weaker per submission.

The 20-25% LAE reduction is the second anchor. Aite-Novarica's 2026 survey measured this at carriers running Tractable for first-touch estimating, Five Sigma or Hi Marley for claims workflow orchestration, and Shift for fraud triage. The mechanism: AI handles the first-touch decision, the routine reserve adjustment, the customer-facing messaging, and the fraud-likelihood scoring - adjuster time concentrates on the 15-25% of claims that require human judgment. Adjuster headcount stable, claim volume grows, LAE per claim drops 20-25% over 18-24 months.

The 30-50% leakage reduction is the third anchor. Deloitte's 2026 benchmark measured this at carriers with full claims AI discipline - Tractable plus CCC for estimating, Shift for fraud, ISO ClaimSearch integration, and AI-assisted subrogation identification. Leakage - payment in excess of indemnity owed - drops from 6-10% of paid losses to 3-5%. The mechanism: faster fraud capture, tighter estimate calibration, better subrogation identification, fewer overpayments on first-touch close.

The 60-70% UW data-extraction time savings is the fourth anchor, the cleanest deployment math in the curriculum. Hyperscience or Indico in production against ACORD 125/140/126 intake plus loss runs plus SOV reduces submission-data-extraction time from 35-50 minutes per submission to 12-18 minutes. Mechanism: structured extraction at 95-99% first-pass accuracy on clean PDFs, 90-95% on degraded scans, with exception routing to human review for the residual.

Use-Case Scoring Template

Each candidate use case scores on six dimensions: anchor metric (which of the five it moves), expected impact (range with confidence interval), feasibility (data readiness, vendor maturity, integration complexity), regulatory exposure (NAIC §4, FCRA, MHPAEA, fair-pricing, Colorado Reg 10-1-1), time-to-value (months from kickoff to measurable lift), and strategic differentiation (does this use case build moat or close gap to peer).

A typical $1B specialty commercial carrier's prioritized use-case portfolio in 2026 contains roughly fifteen candidate use cases. The top six by composite score are: Hyperscience for ACORD intake (anchor: expense ratio; impact: 60-70% UW extraction time; feasibility: 4; exposure: 2; TTV: 4-6 months; differentiation: gap-close), Cytora for triage on commercial lines (anchor: loss ratio + expense ratio; impact: 11-18% submission throughput, 4-7 loss-ratio points on appetite book; feasibility: 3; exposure: 3; TTV: 6-9 months; differentiation: moderate moat), Federato as UW workbench (anchor: combined ratio; impact: 18-25% throughput, 15-22% loss-ratio improvement on covered books; feasibility: 3; exposure: 3; TTV: 12-18 months; differentiation: moat), Tractable for property claims estimating (anchor: loss ratio + expense ratio; impact: 20-30% LAE reduction, 8-15% leakage reduction; feasibility: 3; exposure: 2; TTV: 9-12 months; differentiation: gap-close), Akur8 for pricing on inland marine (anchor: loss ratio; impact: 2-5% loss-ratio improvement; feasibility: 4; exposure: 3; TTV: 12-18 months; differentiation: moderate moat), Shift for fraud on property and auto (anchor: loss ratio; impact: 3-5% loss-ratio improvement via fraud capture; feasibility: 4; exposure: 2; TTV: 6-9 months; differentiation: gap-close).

Build vs. Buy vs. Partner - The Decision Tree

For each prioritized use case, three execution paths are evaluated. Build means the carrier's data-science and engineering team develops the capability in-house (with or without foundation-model partnerships). Buy means the carrier licenses a vendor product (Federato, Akur8, Tractable). Partner means the carrier enters a deeper commercial arrangement than license - co-development, exclusive data-pooling, equity participation, or strategic alliance.

When to Build

Build is the right answer when three conditions are met: the use case is a strategic moat (proprietary data, proprietary risk-modeling approach, proprietary distribution insight), the carrier has the talent depth (10+ data scientists, MLOps function, strong actuarial credentialing), and the 3-year TCO of building is lower than buying. The TCO build math is typically: $1.2-2.5M Year 1 (build), $0.8-1.5M Year 2 (refine), $0.6-1.2M Year 3 (operate). 3-year TCO range: $2.6M-$5.2M. Buy math is typically: $0.6-1.2M annual license + $0.3-0.8M annual operating cost. 3-year TCO range: $2.7M-$6.0M. The two ranges overlap; the deciding factor is strategic moat, not TCO.

Build is justified for: proprietary tail-risk modeling (Berkshire Hathaway's approach), proprietary telematics-and-pricing integration (Progressive Snapshot), proprietary actuarial pricing on a high-IP LOB the carrier wins on. Build is not justified for: ACORD intake (commodity capability; Hyperscience and Indico solve this), generic submission triage (Cytora and Convr solve this), commodity fraud scoring (Shift solves this). The build-or-buy mistake most carriers make in 2026 is trying to build on commodity capabilities and buy on differentiating capabilities - backwards.

When to Buy

Buy is the right answer when the use case is gap-close (table stakes for peer parity), vendor maturity is solid (SOC 2 Type 2, NAIC §4 conformance documented, 50+ carrier deployments), and the carrier's talent depth doesn't support build. The 2026 buy menu by category:

ACORD intake and IDP: Hyperscience ($400K-$1.2M Year 1, all-in $0.5-1.5M annual TCO), Indico Data ($350K-$1M Year 1, $0.4-1.2M annual TCO). Hyperscience for enterprise-grade deployments; Indico for mid-market or specialty.

Submission triage and appetite scoring: Cytora ($1.1M-$2.4M license Year 1, $1.3-2.8M annual TCO with integration), Convr ($600K-$1.5M license Year 1, $0.8-1.8M annual TCO), Send ($400K-$1M license Year 1, $0.5-1.2M annual TCO). Cytora for mid-and-large commercial carriers; Convr for mid-market commercial and specialty; Send for surplus-lines wholesale brokers.

UW workbench: Federato RiskOps ($1.2M-$3.5M license Year 1, $1.6-4M annual TCO depending on UW seat count and LOB scope), Guidewire Cyence (bundled with Guidewire PolicyCenter, $200K-$600K incremental annual), Cytora's workbench layer (bundled with triage license, see above). Federato for specialty commercial and MGA platforms; Guidewire Cyence for Guidewire-native carriers; Cytora's workbench for mid-market commercial.

Pricing: Akur8 ($600K-$2.4M annual depending on LOB count and premium volume; $1.8-7.2M 3-year TCO), Earnix ($800K-$3M annual; $2.4-9M 3-year TCO), Guidewire Predict ($300K-$1.5M annual when bundled with PolicyCenter; $0.9-4.5M 3-year TCO). Akur8 for non-Guidewire carriers seeking AI-native pricing; Earnix for global carriers with multi-region pricing complexity; Guidewire Predict for Guidewire-native US carriers seeking embedded simplicity.

Claims estimating: Tractable ($600K-$1.8M Year 1 for property or auto, $0.8-2.2M annual TCO), CCC Intelligent Solutions ($400K-$1.5M annual for auto, bundled estimate platform), Mitchell ($350K-$1.2M annual for auto). Tractable for AI-native estimating with strongest model performance; CCC for auto-native carriers with deep estimate ecosystem; Mitchell for auto-specialty.

Fraud: Shift Technology ($500K-$2M annual depending on LOB scope and claim volume, $1.5-6M 3-year TCO), FRISS ($300K-$1M annual), Friss specialty fraud (for surplus lines). Shift for property, auto, GL fraud at mid-market and large carriers; FRISS for mid-market focused.

Claims workflow: Five Sigma ($800K-$2.4M annual; $2.4-7.2M 3-year TCO), Hi Marley ($400K-$1.2M annual primarily for messaging; $1.2-3.6M 3-year TCO). Five Sigma as comprehensive claims platform; Hi Marley as customer-messaging layer (can be combined with Five Sigma or used standalone).

When to Partner

Partner is the right answer when the use case requires shared infrastructure (data pooling), shared risk (parametric product structuring with a tech provider), or strategic alignment beyond a license relationship. Examples: Coalition for cyber underwriting (carrier partners with Coalition for embedded cyber-risk data and incident-response), parametric structuring with Descartes Underwriting or AXIS for cat exposures, broker-side data pooling for specialty LOB risk-modeling improvement (carrier partners with three-to-five major brokers for anonymized submission and loss data).

Partnership economics differ from license - typically $200K-$800K annual partnership fee plus shared revenue or shared risk component. The strategic value is in the data exclusivity and the speed-to-market on novel products. Partnership failures usually trace to misaligned commercial expectations (one side expecting a license relationship with extra service, the other expecting strategic depth and shared investment) - get the commercial terms right at signing.

Three-Year TCO - The CFO's Real Question

License pricing is the visible cost; TCO is the question the CFO actually cares about. Three-year TCO includes license, integration engineering, internal staffing (UW operations, claims operations, IT support, governance overhead), change management, training, vendor management, and exit costs (the contractual exit clause needs an annual operational test). For a $1B specialty carrier deploying Federato across three LOBs, the 3-year TCO breakdown is roughly: $4.2M license ($1.4M annual), $1.8M integration and IT, $1.5M internal staffing (UW workbench operations team, governance overlay), $400K change management and training, $300K vendor management overhead, $200K exit-clause testing. Total: $8.4M 3-year TCO. The license is 50% of total cost; the remaining 50% is the operational scaffold.

CFOs who approve license cost without TCO end up funding the operational scaffold from elsewhere, which compounds expense ratio drift. The disciplined approach surfaces full TCO in the business case, attributes ownership to specific budgets, and tracks each line item quarterly.

The Build/Buy/Partner Mistakes Most Carriers Make

Three mistakes recur in 2026 carriers' AI portfolio decisions. First, building on commodity capabilities. A $400M specialty carrier with eight data scientists tries to build proprietary ACORD intake instead of licensing Hyperscience - three years and $4M later they have a working prototype that's still less accurate than Hyperscience out of the box. Commodity capabilities should be bought; the talent should be deployed on moat-building work. Second, buying on differentiating capabilities. A carrier that wins on proprietary tail-risk modeling licenses a generic pricing tool and loses the differentiation. Buy on table stakes; build on moats. Third, partnering before talent is ready. Multi-broker data pooling fails when the carrier doesn't have data engineers to ingest, normalize, and apply the pooled data. Partnership ROI requires internal capacity to consume the partnership output.

The Portfolio View - Twelve-to-Fifteen Use Cases

A mid-market specialty carrier's prioritized AI portfolio in 2026 should contain twelve-to-fifteen use cases across the five-metric anchors, sequenced into the 12-month / 3-year / 5-year horizons from Ch1-2, with build/buy/partner decisions documented per use case. The portfolio is reviewed quarterly by the executive committee and annually by the board's risk committee. Each use case has a single named owner, a primary metric, a TCO range, an execution path (build/buy/partner), a vendor or partner selection (where applicable), and a kill criterion at defined checkpoints.

The portfolio's combined impact is what the CRO presents to the board and the AM Best analyst. For a $1B-$2B specialty carrier with a fifteen-use-case portfolio executed at industry-disciplined pace, the 3-year impact is 2.5-4 combined-ratio points and the 5-year impact is 4-7 combined-ratio points. Below 2.5 points at 3 years signals execution failure; above 4 points signals either exceptional execution or unrealistic baseline claims that won't survive an AM Best peer benchmark review.

Key Takeaways

• Five-metric anchor: loss ratio, expense ratio, combined ratio, growth lift, time-to-value. McKinsey 18.5% loss-ratio delta (top-quartile vs. median personal lines); Aite-Novarica 20-25% LAE reduction (Tractable + Five Sigma + Shift); Deloitte 30-50% leakage reduction; AM Best 60-70% UW data-extraction time savings.
• Use cases score on six dimensions: anchor metric, expected impact with CI, feasibility, regulatory exposure, time-to-value, strategic differentiation. Use cases that cannot be anchored are deferred.
• Build when use case is strategic moat, talent depth exists (10+ data scientists), and 3-year TCO is lower than buy. 3-year build TCO $2.6M-$5.2M for a typical use case; 3-year buy TCO $2.7M-$6.0M - the ranges overlap, the deciding factor is moat.
• Buy on commodity and table-stakes (ACORD intake, generic triage, generic fraud). Build on moats (proprietary tail-risk, telematics-pricing, high-IP LOB pricing). The recurring 2026 mistake is the inverse - building commodity and buying moats.
• 2026 buy menu pricing: Hyperscience $0.5-1.5M annual; Cytora $1.3-2.8M annual; Federato $1.6-4M annual; Akur8 $0.6-2.4M annual; Tractable $0.8-2.2M annual; Shift $0.5-2M annual; Five Sigma $0.8-2.4M annual. License is roughly 50% of full TCO; integration, staffing, governance overhead is the other 50%.
• Partner when shared infrastructure (data pooling), shared risk (parametric structuring), or strategic alignment beyond license is required. Partnership economics: $200K-$800K annual fee plus shared revenue or shared risk. Failures trace to misaligned commercial expectations.
• Three-year TCO is the CFO's real question; license is the visible cost. $1B specialty carrier Federato 3-year TCO: $8.4M ($4.2M license + $1.8M integration + $1.5M staffing + $400K change mgmt + $300K vendor mgmt + $200K exit testing).
• Mid-market specialty carrier's prioritized portfolio: 12-15 use cases across anchors, sequenced 12-month / 3-year / 5-year, build/buy/partner documented per use case. 3-year impact 2.5-4 combined-ratio points; 5-year impact 4-7 combined-ratio points; below or above signals execution or baseline-claim problem.