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

EXL's $310 million definitive agreement to acquire iMerit, signed June 24, 2026 and expected to close in Q3, is structured as $170 million upfront with up to $140 million in earnouts tied to two-year performance milestones (EXL press release, June 24, 2026). The deal 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 that determines how carrier AI models actually behave before they ever reach an underwriting desk or claims queue.

Trade coverage has focused almost entirely on the price and the strategic positioning. The more consequential question sits one level deeper: when the vendor that trains an insurer's underwriting model using proprietary human feedback is also the vendor running that model in production, what does independent actuarial model validation actually examine? The training layer is where model behavior is shaped, and it is the one layer that current carrier governance frameworks do not reach.

The Training Layer: What Ango and Scholars Actually Do

Reinforcement learning from human feedback, commonly called RLHF, is the process by which a base language or prediction model is iteratively refined using human preference judgments. A human annotator reviews pairs of model outputs and signals which is better; that preference signal trains a reward model; the reward model then steers subsequent fine-tuning runs. The resulting model does not just answer questions in the abstract. It reflects the implicit loss function that the human annotators encoded, which means annotator expertise determines model quality in a way that no amount of post-deployment back-testing can fully reconstruct.

iMerit's Ango platform provides the tooling infrastructure for this process: chain-of-thought reasoning capture, red-teaming workflows, multimodal evaluation support, and automation that routes annotation tasks to qualified reviewers (iMerit, Ango Hub product documentation). The Scholars network is the human capital layer that makes the process domain-specific. Rather than generic crowdworkers rating whether one chatbot response is "better," the Scholars network comprises physicians, scientists, engineers, and linguists who bring subject-matter expertise to preference judgments in regulated, high-stakes domains. For insurance, that means underwriters, claims adjusters, and actuarial specialists rating model outputs on whether they reflect accurate risk assessment, appropriate coverage interpretation, or sound reserving logic, not whether the prose sounds natural.

This is the strategic asset EXL is actually buying. The Ango platform is replicable technology; iMerit's competitors in the annotation market include Scale AI, Surge AI, and a half-dozen well-funded alternatives. The Scholars network, with its domain-credentialed roster and documented quality processes in regulated industries, is substantially harder to assemble and significantly harder to validate for a new entrant. "Specialized high-quality data is the foundation of AI success," Radha Ramaswami Basu, iMerit's CEO and founder, noted in the acquisition announcement (GlobeNewswire, June 24, 2026). The emphasis on quality over volume is the differentiating claim.

Why EXL Placed This Bet Now

EXL reported $570.4 million in Q1 2026 revenue on April 29, with data and AI-led services now representing 60% of total company revenue, up 28% year over year. The insurance segment reached $193.9 million, a 12.6% year-over-year increase, as carriers migrated existing engagements from legacy digital operations toward higher-margin AI workflow deployments (EXL Q1 2026 earnings, April 29, 2026). Full-year 2026 revenue guidance stands at $2.30 billion to $2.33 billion. The company is growing, and AI is driving it.

The problem with building on general-purpose AI infrastructure at that revenue scale is exposure. EXL's Insurance LLM, launched in September 2024 in partnership with NVIDIA using the NeMo framework, was trained on over a decade of domain-specific insurance data. But the fine-tuning process for any model update, any new use case, or any carrier-specific customization requires ongoing labeled training data and qualified human feedback. Before this acquisition, EXL sourced that capability through a patchwork of vendors and internal annotation teams. After it closes, EXL controls the full loop from raw insurance data through labeled preference pairs through deployed model.

"Success requires industry-specific data, rigorous evaluation and reinforcement learning to deliver reliable results in business-critical workflows," Rohit Kapoor, EXL's CEO, stated in the acquisition announcement (EXL Newsroom, June 24, 2026). That framing is not incidental marketing. It is a direct signal that EXL intends to compete on the quality of its fine-tuning pipeline, not just the scope of its deployment capabilities. The earnout structure reinforces the strategic logic: $140 million in contingent consideration tied to milestones over two years suggests EXL is betting on iMerit's network expanding within insurance, healthcare, and banking rather than simply absorbing a static capability.

The Accountability Gap in Actuarial Model Validation

The actuarial model validation framework that most carriers apply to vendor-supplied AI focuses on the deployment layer: inputs, outputs, back-testing against held-out data, discrimination testing, documentation of model logic. An actuary validating an EXL underwriting model can examine which features drive predictions, whether protected characteristics proxy into the output, and how the model performs across rating segments and loss ratios. These are meaningful checks. They are also insufficient for a model shaped by RLHF, because the preference feedback loop that determined the model's implicit decision criteria is upstream of anything the validation can observe.

The Scholars feedback data, the reward model trained on those preferences, and the fine-tuning runs that steered the base model toward "good" underwriting behavior all exist inside iMerit's proprietary process. A carrier's actuary does not receive those artifacts. The carrier receives a fine-tuned model. What the reward model was optimizing for, which preference pairs it was trained on, which Scholars made which judgments, and whether the reward model itself was validated against actuarial soundness criteria: none of those questions are answerable from deployment-layer outputs alone.

This is a structural gap, not a criticism of any specific governance framework. Actuarial standards addressing predictive models are built around the deployment interface because that is where actuarial judgment has historically entered the process. The fine-tuning layer did not exist as a commercially relevant concern when those frameworks were developed. As EXL vertically integrates training with deployment for carrier workflows, the governance question shifts from "how was this model built" to "who decided what counts as a good prediction, and by what criteria." That decision now sits inside a single vertically integrated vendor.

The practical implication for actuaries certifying these models is not that certification is impossible. It is that certification requires new contractual access rights: audit rights to training data samples, documentation of Scholars qualification criteria, reward model validation reports, and change-management protocols that notify the carrier when a fine-tuning update materially shifts model behavior. Carriers that renew EXL contracts without specifying these rights in the post-acquisition environment will find that the governance gap between what they signed up to validate and what they can actually see has widened significantly.

The Data Pooling Question

Fine-tuning an insurance AI model requires labeled insurance data: claims records, underwriting submissions, policy language, loss experience, pricing decisions. The quality and specificity of that training data drives how well the resulting model performs on insurance tasks compared with a general-purpose alternative. By 2025, an estimated 70% of enterprises had adopted RLHF or direct preference optimization as their primary AI alignment method, up from 25% in 2023, and the differentiator in enterprise deployments is increasingly the domain specificity of the preference data, not the architecture of the base model (PowerToFly, 2025).

For EXL, the acquisition raises a data governance question that the press release does not address: when iMerit processes insurance training data on behalf of multiple EXL carrier clients, is each carrier's data siloed, or does iMerit build shared training corpora that improve model quality across the client base? Both answers are commercially defensible. Siloed annotation is cleaner from a competitive intelligence standpoint but does not realize the scale advantages that make a training-layer platform economically attractive. Pooled annotation builds better models faster but means that Carrier A's claims patterns, pricing signals, and loss experience could inform the fine-tuned model that Carrier B eventually runs in production.

This is not a hypothetical concern. EXL's insurance client list includes major property-casualty carriers, life insurers, and specialty lines operators across the United States and United Kingdom. If iMerit's Scholars use data from multiple carriers to train a shared insurance reward model, the resulting model encodes competitive intelligence extracted from carriers who may have no idea their proprietary experience is being pooled. The NAIC's emerging vendor registry framework, which requires carriers to document third-party AI dependencies in rate and form filings, is directly implicated: a carrier's regulatory disclosure of its AI model would not currently extend to disclosing the training data provenance of that model's fine-tuning layer.

Carriers should treat the post-acquisition contract renewal as a data governance negotiation, not just a pricing discussion. The specific questions that require explicit contractual answers are: whether training data is siloed or pooled across EXL's carrier client base, what consent the carrier has given for its proprietary data to be used in EXL model training generally, whether the carrier retains any rights to training artifacts derived from its data, and what notification obligations EXL has when fine-tuning updates materially affect a model the carrier has certified.

How This Reshapes the Insurance AI Vendor Landscape

The insurance AI value chain has three structurally distinct layers: raw data generation and aggregation, model training and fine-tuning, and deployment and workflow integration. Before June 24, no single major analytics vendor controlled all three for insurance specifically.

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

The training layer was the structural gap EXL has now closed. Verisk's competitive moat is its raw data: ISO loss cost filings, property databases, catastrophe model inputs. Guidewire's moat is its transactional core-system position: carriers that run PolicyCenter and ClaimCenter give Guidewire a privileged view of policy and claims workflow data. Neither Verisk nor Guidewire has announced a capability to fine-tune insurance-domain models using domain-credentialed human feedback. That is now EXL's claim.

The build-versus-buy calculus for carriers changes as a result. A carrier evaluating EXL for AI deployment in 2024 could in principle use EXL for deployment and retain a separate annotation vendor for fine-tuning, keeping the two functions separate and the governance accountabilities distinct. After the acquisition closes, EXL is the integrated vendor for both. A carrier that wants EXL's deployment capabilities but a different fine-tuning vendor will need to negotiate that separation explicitly, and EXL's commercial incentive runs in the opposite direction. Vertical integration favors bundling, and bundling tends to erode the separation that actuarial governance frameworks depend on.

What Carriers and Actuaries Should Do Before Q3

The acquisition closes in Q3 2026. That leaves roughly one quarter for carriers and their actuarial teams to inventory their current EXL relationships and identify where the training-layer integration creates new governance exposure.

The most urgent action is a data use audit. Carriers with active EXL engagements should pull their current master services agreements and data processing addenda and identify whether the data use rights provisions were written to contemplate AI training use of carrier data. Most agreements executed before 2024 were not written with RLHF in mind. Standard BPO data use clauses typically permit the vendor to use client data to perform the contracted service, not to use it to train models that will be deployed for other clients.

Second, carriers planning model certifications for AI-assisted underwriting or claims workflows in 2026 or 2027 should extend their validation scope upstream. The actuary certifying the model should be able to answer, at minimum: what fine-tuning was applied to the base model, when it was applied, what training data categories were used, and whether the reward model driving fine-tuning has been validated against the carrier's actuarial and regulatory requirements. If those answers are not available from EXL as a matter of standard documentation, they should be negotiated into the service agreement before the next contract renewal.

Third, carriers should watch whether EXL's post-acquisition marketing shifts toward vertically integrated pricing. If EXL bundles iMerit fine-tuning with its deployment engagements in the way SaaS vendors bundle implementation with licensing, the operational separation that currently allows independent vendor oversight will compress. The time to negotiate the right to use a third-party annotation vendor alongside EXL's deployment platform is before that bundling becomes the commercial default.

EXL has built a credible AI revenue engine at 60% of $570 million in quarterly revenue, and the iMerit acquisition extends that position upstream in a direction neither Verisk nor Guidewire has matched. The actuarial accountability question it opens is not a reason to avoid EXL's platforms. It is a reason to define, in contract and in governance documentation, exactly which parts of the AI stack the certifying actuary is accountable for and exactly which parts sit upstream of that accountability. The training layer is now one of them.