EXL's $310 Million iMerit Acquisition Moves Insurance AI Into the Training Layer

EXL's $310 million agreement to acquire iMerit, signed June 24, 2026 and expected to close in Q3, is structured as $170 million upfront plus up to $140 million in earnouts tied to two-year performance milestones. It adds the Ango annotation platform and a global Scholars expert network to EXL's insurance AI stack, moving the analytics vendor upstream from model deployment into the fine-tuning layer where model behavior is shaped before anything reaches an underwriting desk.

Key Takeaways

• $140 million of the $310 million is contingent, tied to two-year milestones, which prices the deal on iMerit's network expanding inside insurance rather than on a static capability being absorbed.
• The Scholars network, not the Ango platform, is the asset. Annotation tooling has well-funded substitutes; a domain-credentialed reviewer roster with documented quality processes in regulated industries does not.
• Reinforcement learning from human feedback sits upstream of every deployment-layer check an actuary runs. The reward model, the preference pairs, and the fine-tuning runs are not artifacts the carrier receives.
• Data and AI-led services are 60% of EXL revenue, with the insurance segment at $193.9 million in Q1 2026, up 12.6% year over year, which is the scale that makes the training layer worth owning.
• No major insurance analytics vendor held all three layers before June 24. Verisk and Guidewire each hold raw data and deployment with no dedicated training capability.

What $310 Million Buys

Reinforcement learning from human feedback refines a base model using human preference judgments. An annotator reviews pairs of model outputs and signals which is better; that signal trains a reward model; the reward model steers subsequent fine-tuning. The finished model reflects the implicit loss function the annotators encoded, which is why annotator expertise determines model quality in a way post-deployment back-testing cannot fully reconstruct.

Ango supplies the tooling: chain-of-thought reasoning capture, red-teaming workflows, multimodal evaluation, and routing of annotation tasks to qualified reviewers. Scholars supplies the people. Rather than generic crowdworkers rating whether a response reads well, the network comprises physicians, scientists, engineers, and linguists judging outputs in regulated domains. For insurance that means reviewers rating whether an output reflects accurate risk assessment, sound coverage interpretation, or defensible reserving logic.

The split matters for what EXL actually acquired. Ango is replicable technology with competitors including Scale AI and Surge AI. A credentialed reviewer roster with documented quality processes is materially harder to assemble and harder for a new entrant to validate. "Specialized high-quality data is the foundation of AI success," said Radha Ramaswami Basu, iMerit's CEO and founder.

The earnout structure carries the same reading. EXL reported $570.4 million in Q1 2026 revenue with data and AI-led services at 60% of the total, up 28% year over year, and full-year guidance of $2.30 billion to $2.33 billion. At that scale the fine-tuning pipeline is a recurring input cost, previously sourced through a patchwork of vendors and internal teams.

Where Model Validation Stops

Most carrier validation of vendor AI examines the deployment layer: inputs, outputs, back-testing against held-out data, discrimination testing, documentation of model logic. An actuary can establish which features drive predictions, whether protected characteristics proxy into the output, and how performance varies across rating segments.

Those checks do not reach the reward model. The Scholars feedback data, the reward model trained on those preferences, and the fine-tuning runs that steered the base model all sit inside the vendor's process. The carrier receives a fine-tuned model. What the reward model optimized for, which preference pairs trained it, and whether it was itself validated against actuarial soundness criteria are not recoverable from deployment outputs.

Vendor	Raw Data	Training Layer	Deployment
EXL (post-iMerit)	Partial (client workflow data)	Yes, via iMerit Ango + Scholars	Yes, Insurance LLM + workflow platforms
Verisk	Yes, ISO loss cost data, analytics databases	No dedicated training capability	Yes, MCP connectors, analytics modules
Guidewire	Yes, transactional core-system data	No dedicated training capability	Yes, ProNavigator, claims AI, PricingCenter
Scale AI / Surge AI	No insurance-specific data	Yes, general annotation platforms	No insurance deployment

This is a structural gap rather than a failure of any one framework. Predictive model standards were built around the deployment interface because that is where actuarial judgment historically entered. The governance question has moved from how the model was built to who decided what counts as a good prediction, and that decision now sits inside one vertically integrated vendor.

The fix is contractual rather than technical: audit rights to training data samples, documentation of Scholars qualification criteria, reward model validation reports, and change-management notice when a fine-tuning update materially shifts behavior. The commercial pressure runs the other way. Before the acquisition a carrier could use EXL for deployment and a separate annotation vendor for fine-tuning, keeping the accountabilities distinct. Holding that separation now requires negotiating against a bundling incentive, and neither Verisk nor Guidewire offers a competing training-layer capability to negotiate with.

Whose Data Trains the Reward Model

Fine-tuning an insurance model requires labeled insurance data: claims records, underwriting submissions, policy language, loss experience, pricing decisions. By 2025 an estimated 70% of enterprises had adopted RLHF or direct preference optimization as their primary alignment method, up from 25% in 2023, and the differentiator in enterprise deployments is the domain specificity of the preference data rather than the base architecture.

That raises a question the announcement does not settle. When one vendor annotates insurance training data for multiple carrier clients, is each carrier's data siloed or pooled into shared corpora? Both are commercially defensible. Siloed annotation is cleaner competitively but forgoes the scale advantage that makes a training-layer platform attractive. Pooled annotation builds better models faster, and means one carrier's claims patterns and pricing signals can inform the model another carrier runs in production.

EXL's insurance client base spans property-casualty carriers, life insurers, and specialty operators in the United States and United Kingdom. The NAIC's emerging vendor registry framework requires carriers to document third-party AI dependencies in rate and form filings, but that disclosure describes the deployed model. It does not currently extend to the training data provenance of the fine-tuning layer, so pooling would be invisible at exactly the point regulators look.

The exposure predates the deal. Standard outsourcing data use clauses permit a vendor to use client data to perform the contracted service, not to train models deployed for other clients, and most agreements executed before 2024 were not drafted with RLHF in view. Carriers renewing on those terms are consenting to a use case their contract never described.