Verisk Q2: XactAI Licenses Grow Tenfold as AI Friction Clears

XactAI reached approximately 7,000 licensees in Verisk's second quarter, an increase of nearly tenfold since March, alongside $806 million of revenue, up 4.3%, and free cash flow of $298 million, up 57.9% (Verisk earnings release, July 2026). Three months ago the same company's earnings call was describing governance friction that stretched AI contracting timelines. One quarter later the friction is gone from the script and the adoption curve has gone vertical, and the mechanism that cleared it deserves as much actuarial attention as the number itself.

The Quarter: Steady Top Line, Softer GAAP, a Cash Machine Underneath

The headline financials from the July 29 release describe a business performing to plan. Revenue of $806 million grew 4.3% as reported and 5.8% on an organic constant-currency basis, with adjusted EBITDA of $464 million, up 4.2%, holding a margin near 57.5% (Verisk, July 2026). GAAP results ran softer: net income of $229 million fell 9.8% on a higher 24.6% effective tax rate, elevated interest expense, and litigation-related legal fees, taking diluted GAAP EPS down 3.3% to $1.75 while adjusted EPS rose 5.3% to $1.98. The cash statement is where the strength concentrates. Operating cash flow climbed 49.7% to $366 million and free cash flow reached $298 million, up 57.9%, funding a $200 million accelerated share repurchase that delivered an initial 949,190 shares, part of $1.9 billion of aggregate first-half buybacks, with $800 million of authorization remaining (Yahoo Finance, July 2026).

Full-year guidance was reaffirmed across the board: revenue of $3.19 to $3.24 billion, adjusted EBITDA of $1.79 to $1.83 billion, and adjusted EPS of $7.45 to $7.75. For actuaries reading vendor results as a proxy for carrier technology budgets, an unchanged outlook from the industry's central data utility says insurer spending on data and analytics is holding through the soft market. Chief executive Lee Shavel described "balanced growth across underwriting and claims" with results "modestly ahead of expectations" (Yahoo Finance, July 2026).

From Contracting Friction to 7,000 Licensees in One Quarter

The context that makes the XactAI number remarkable is what Verisk itself said in April. The first-quarter call introduced governance-friction language that had not appeared in prior disclosures: intellectual property, privacy, and compliance negotiations were extending contracting timelines for the company's most advanced analytics products, a structural drag actuary.info covered at the time as a leading indicator for the whole vendor ecosystem. One quarter later the same company reports its flagship claims AI growing from roughly one-tenth of its current base to about 7,000 licensees, with adoption described as a shift from experimentation to production scale on retrieval-augmented generation architectures (Investing.com transcript, July 2026).

Sequentially, the financial picture barely moved: first-quarter revenue was $783 million with 7% subscription growth, so the second quarter's $806 million continues the same mid-single-digit trajectory rather than reflecting an AI revenue surge. That is the telling detail. The tenfold licensee expansion has not yet shown up as a revenue inflection, because the adoption is riding inside existing subscriptions rather than generating new contract value. Verisk is deliberately trading near-term monetization for installed-base penetration, the classic land-first strategy, and its stated menu of future monetization options, subscription, transactional, and agentic licensing, describes how the harvest is planned. Carriers reading the quarter should register both halves: the adoption is essentially free today, and the pricing conversation arrives after the workflows are embedded.

Friction does not disappear because carriers stopped caring about governance. It disappears because the vendor re-engineered the offer so that governance review had nothing new to negotiate. The Q2 disclosures show exactly how. The two new Claude connectors, which make Verisk underwriting and claims content queryable in natural language, are included in base subscriptions, with customers separately bearing the Claude token costs (Yahoo Finance, July 2026). Bundling the AI into paper the carrier has already signed converts a new-product procurement, with its IP and privacy negotiation, into a feature release under an existing agreement. XactAI's functions, claim summarization, photo labeling, document-data extraction, and estimating recommendations, ride the Xactimate footprint the adjusters already use. The lesson for every carrier negotiating with every vendor: the contracting friction was real, and the industry's largest data provider just demonstrated that the fastest path around it is distribution through existing agreements rather than better negotiation.

The Governance Checkpoint the Bundling Skips

There is a quieter consequence inside the same mechanism, and it runs the other direction. The vendor-due-diligence expectations in the NAIC's AI model bulletin, and the vendor-review provisions carriers wrote into their own AI governance policies over the past two years, are triggered by procurement events: a new contract, a new tool, a new data flow. A capability that arrives as a feature release inside an existing subscription crosses none of those tripwires. The same engineering that dissolved the sales friction also routes the technology around the checkpoint where model-risk review was supposed to happen. An adjuster whose estimating screen began offering AI-generated recommendations this spring did not wait for the model governance committee to evaluate the recommendation engine; the committee may not know the engine is there.

The fix is a policy change, not a negotiation: carriers should define material AI feature releases from incumbent vendors as governance events, with the same inventory, validation-evidence, and monitoring expectations a new purchase would face. Verisk publishes documentation and has an established governance relationship with its regulator-facing content, which makes it the easy case. The precedent matters for the harder ones, because every vendor watching this quarter's adoption number has now seen that bundling beats negotiating, and the next embedded AI feature will come from a vendor with thinner documentation.

The Token-Cost Fine Print

The pricing structure deserves a second look before anyone celebrates the bundling. Connectors in the base subscription with token costs on the carrier is a consumption model: the marginal cost of every natural-language query, every summarized claim, every labeled photo lands on the carrier's expense line, metered by a frontier-model provider whose rates the carrier does not control. This is the same architecture actuary.info flagged in Palantir's second quarter, where usage-based AI platform economics grow with the customer's own adoption. A claims organization that rolls XactAI-style tooling across thousands of adjusters has committed to a variable cost that scales with claim volume, exactly the quantity that spikes in the catastrophe quarters when loss adjustment expense is already elevated. Expense actuaries should ask now, while the numbers are small, how AI token consumption is being allocated: to LAE, to general technology expense, or unallocated, because the answer determines whether the combined ratio sees it and whether rate indications capture it.

Verisk, for its part, told analysts its AI monetization could evolve toward subscription pricing, transactional pricing for new applications, and software licensing for agentic AI (Yahoo Finance, July 2026). Transactional pricing for agentic claims tools would put the vendor on the same side of the meter as the token provider. A top-10 carrier is already live on the Claude connectors; its expense allocation choices will quietly set the industry's precedent.

What 7,000 AI-Assisted Estimators Do to Claims Data

The second-order effect lands in the data actuaries consume. Xactimate sits underneath a large share of US property claim estimates, which makes its outputs raw material for severity benchmarks, trend selections, and reserving diagnostics across the industry. When roughly 7,000 licensed users are receiving AI-generated estimating recommendations inside that workflow, the estimates stop being fully independent observations. Recommendation engines compress variance: adjusters accept suggested line items more often than they overrule them, which narrows the distribution of estimates for similar damage and can shift its center in whichever direction the model's training data leans. None of that is visible in the aggregate severity statistics a pricing actuary downloads; the data simply becomes smoother and more internally correlated.

The practical response is not to distrust the benchmarks but to track the instrumentation. Severity trend that moderates in 2026 property lines will have at least three candidate explanations: genuine cost moderation, mix, and the arrival of recommendation-shaped estimates at scale. An actuary selecting trend for a rate filing or testing reserve adequacy against industry development patterns should know which portion of the underlying estimates ran through AI assistance, a disclosure worth requesting from vendors and TPAs directly. The profession spent a decade learning to adjust for the effects of managed care on medical severity data. AI-assisted estimating is the property-lines version of the same instrumentation change, arriving faster.

Testing Recommendation Effects in Your Own Book

Carriers do not have to wait for industry benchmarks to drift before measuring the effect; the diagnostic is available in any claims database that flags which estimates ran through AI assistance. Match AI-assisted and conventionally written estimates on peril, region, coverage, and severity band, then compare three quantities over the following twelve months: initial estimate size, supplement frequency, and reopen rate. If recommendation-assisted estimates come in lower initially but generate more supplements later, the tool is underscoping and the apparent severity improvement is a timing artifact that will unwind through development. If assisted estimates run tighter to final cost with fewer supplements, the tool is genuinely improving accuracy and the LAE spent on it is buying real signal. Either finding is worth knowing before the next reserve review, and the analysis costs a week of one analyst's time. The same cohort discipline actuary.info argued for on underwriting AI applies unchanged on the claims side: instrument the change, or inherit someone else's assumption about what it did.

Signals Worth Tracking Into Year-End

Three markers will show whether this quarter was the inflection it appears to be. The XactAI licensee count at Q3 will reveal whether tenfold-since-March was a launch spike or a curve still climbing; at the current base, growth has to come from mid-market carriers and TPAs whose governance capacity is thinner than the top-10 carrier already live. The monetization structure Verisk chooses for agentic AI, subscription versus transactional, will determine whether the vendor's incentives align with carrier efficiency or with usage growth. And the Synergy Studio migrations planned through year-end will move catastrophe modeling onto the same cloud platform, concentrating one more actuarial workflow with the same counterparty. Verisk's quarter says the AI adoption dam has broken in claims. The carriers now own the downstream consequences: metered costs in their expense ratios and recommendation-shaped estimates in everyone's benchmarks.