HUB and Baldwin Deploy Claude to 20,000+ Broker Staff, Reporting 85% Productivity Gains

HUB International put Anthropic's Claude in front of more than 20,000 employees in February 2026 and disclosed what most enterprise AI announcements withhold: 85% productivity gains in targeted use cases, 2.5 hours saved per employee per week, and satisfaction above 90%. Two months later Baldwin Group went firm-wide as the first named insurance customer of Anthropic's $1.5 billion financial services joint venture.

Key Takeaways

• 2.5 hours per employee per week across 20,000-plus staff is roughly 2.6 million labour hours a year, over $100 million of recovered capacity at a conservative $40 fully loaded hourly cost.
• $1.61 billion of trailing revenue and about 5,000 employees at Baldwin against HUB's roughly $5 billion and 20,000-plus, which is the same platform bought at two very different scales.
• 42% of P&C insurers never measured AI outcomes at all, in Capgemini's finding, which is what makes HUB publishing a number the notable part.
• Over 500,000 insurance professionals now have Claude access across the named carrier, broker and consultancy deployments alone.

Two Brokers, One Platform, Different Positions

HUB began deploying in late Q4 2025 and announced on February 25, 2026, across Claude Enterprise for knowledge workers, Claude Code for technology teams and the API for custom agentic work. It did not start there: the firm piloted robotic process automation in 2020 and moved into generative AI in 2022, so the Claude rollout landed on four years of tooling rather than on a standing start.

Baldwin announced on May 4, 2026 after several months of piloting, covering roughly 5,000 employees against $1.61 billion of trailing revenue, and described "measurable improvements in client-facing insights, productivity, and workflow efficiency" without publishing figures. Its announcement landed the same day as Anthropic's $1.5 billion joint venture backed by Blackstone, Hellman & Friedman, Goldman Sachs, General Atlantic and Sequoia Capital, which targets mid-sized and private equity-backed firms.

The capital relationships are worth stating plainly. HUB is backed by Hellman & Friedman, one of the investors in that joint venture, and Baldwin is its first named insurance customer.

Why Claude rather than the alternative reduces to three properties a broker cares about more than a carrier does. Models calibrated to refuse an uncertain answer matter when a hallucinated exclusion becomes errors and omissions exposure. Enterprise data isolation matters when the corpus is policyholder loss histories and benefits compensation data. And document reasoning matters because the input is not a prompt but a 50-page renewal package or a stack of competing carrier quotes.

The Effect Lands in the Carrier's Expense Ratio, Not the Broker's

Brokers generate the submissions that carriers underwrite, so a change in broker workflow arrives upstream of carrier pricing whether or not the carrier paid for it.

The friction is well documented: incomplete ACORD applications, missing loss runs, inconsistent exposure data, and submissions sent to carriers with no appetite for the risk. Each of those consumes underwriting time on business that will not bind.

The disclosed use cases point straight at it. Account managers synthesising client information and comparing policy forms produce more complete submission packages. Producers using AI-assisted appetite matching direct submissions to carriers more likely to write them. Client-facing applications that gather and validate exposure data before the submission leaves the broker remove the back-and-forth entirely.

Size the capacity that frees up. 2.5 hours per employee per week across 20,000-plus staff is about 2.6 million hours a year, over $100 million at a conservative $40 fully loaded rate. Whether that becomes headcount reduction, revenue per producer, or service levels is HUB's decision, but the underwriting hours it saves on the other side of the submission are not HUB's to keep.

That creates an attribution problem in a rate filing. Carrier expense loading for submission processing is modelled as a carrier expense, and a reduction in it originating at the broker is not under the carrier's control. It can be reversed by the broker switching platforms, cutting technology spend, or being acquired. An expense ratio improvement that depends on a third party's technology budget is a different assumption from one that depends on the filer's own operations, and the filing does not currently distinguish them.

The economics also explain why brokers were last. Carriers have dedicated technology budgets and clear ROI paths through underwriting and claims. A mid-market brokerage on $200 million of revenue at 8% technology spend has $16 million for everything, agency management system through cybersecurity. HUB's numbers are the case that clears that bar, which is why publishing them matters against Capgemini's finding that 42% of P&C insurers never measured AI outcomes at all.

One Model at Every Stage of the Same Decision

Entity	Type	Deployment Scale	Announcement
Travelers	Carrier	10,000 staff; 30,000 via TravAI	January 2026
Allianz	Carrier (global)	156,000 employees	January 2026
AIG	Carrier	Multi-agent underwriting via Palantir	Q1 2026
HUB International	Broker	20,000+ employees	February 2026
PwC	Consultancy	30,000 certified; 364,000 planned	May 2026
Baldwin Group	Broker	~5,000 employees	May 2026
Deloitte	Consultancy	15,000 certified; 470,000 access	October 2025

The named deployments alone put Claude in front of more than 500,000 insurance professionals, before Accenture's 30,000 trained and KPMG's 276,000 with access. CB Insights recorded a 199% quarter-over-quarter rise in Anthropic mentions on insurance earnings calls.

The concentration matters because of where in the chain it sits, not because of its size. The same foundation model now prepares the broker's submission, assists the carrier's underwriting analysis of that submission, and supports the consulting firm's review of that underwriting. If it interprets a policy term a particular way or weights a risk factor consistently, that tendency is present at all three stages at once rather than being caught by one of them.

That is a correlation, and actuarial practice does not currently model it as one. Independence between the data pipeline, the pricing model and the review is assumed rather than tested. ASOP No. 56 makes the actuary responsible for understanding the models relied upon, and the appointed actuary signing an opinion over a book whose submissions were shaped by the same model used to analyse them has a correlation question that ordinary model validation does not raise.

The market data cuts both ways on how bad this is. The Ramp AI Index for May 2026 put Anthropic at 34.4% of businesses against OpenAI's 32.3%, close to even. An IA Capital survey the same month found OpenAI in roughly nine of ten insurance carrier technology stacks. If brokers standardise on one model and carriers on another, the chain is diversified by accident. If either side converges on the other, it is not, and nobody is currently measuring which way it is going.