Written by: Claude AI.
Curator/Editor: Học Trò.
Anthropic's prospectus spends nearly a hundred pages warning investors about what its technology might do to the world. Twenty pages later, it shows what that technology is already doing for twenty-one organizations: a drug maker, two insurers, a trading firm, a browser, a refugee charity, a rotisserie restaurant in Boston and a family remodeling business in Florida. This essay sorts those case studies by industry, reads them closely, and checks each one against what newspapers, trade journals, company press releases and independent researchers have said. The aim is a fair picture of what AI can already do for a business in late 2026: the measured, the plausible and the merely asserted.
A note on authorship: this essay was researched and drafted with Claude, a model made by Anthropic, the company whose filing is being examined. That is one reason the method below leans so heavily on sources outside the filing.
1. The Brochure Inside the Risk Document
On November 10, 2026, Anthropic filed the second amendment to its registration statement with the U.S. Securities and Exchange Commission, the Form S-1/A that will take the company public. Printed out, the document runs to well over four hundred pages. Most readers who have looked at it closely have been drawn, understandably, to the risk factors. In a video essay for The New York Times, the economist Natasha Sarin pointed out that close to a hundred pages are devoted to risks, some of them commercial and some of them, in her words, "really existential." The filing states plainly that the company's models "might behave in unintended, misaligned, or autonomous ways, including by taking harmful actions that no user or third party directed." It is unusual for a company to say such things about its own product in the document it uses to sell its shares.
Turn further into the filing, however, and the tone changes completely. Between pages 229 and 248 of the printed version sits a run of glossy, magazine-style pages, each devoted to a single customer. Every page follows the same layout: a company logo, a short headline ("Reinventing commercial underwriting," "Speeding up new medicine discovery," "Finding the vulnerabilities attackers haven't"), a box of three or four large numbers labelled "Results," a "Background" paragraph, an "Impact" section, a quotation from a senior executive, and at the foot of the page a small line giving the year the company became a customer. There are twenty pages and twenty-one organizations, because the last page shares its space between two small businesses.
These pages are easy to skim past, and a sceptic might dismiss them as advertising, which in part they are. But they are also something more useful. Anthropic's own legal and financial teams have had to stand behind every word of a registration statement, because misstatements in an S-1 carry legal liability that a marketing blog post does not. And the case studies cover a striking range of work: underwriting commercial property, reviewing a nine-hundred-document legal data room, finding security holes in a web browser, validating computer chips, tracking malnourished children in East Africa, and redesigning the line in a restaurant kitchen. Taken together, they amount to one of the most concrete public inventories so far of what businesses are actually using generative AI to do.
That is the reason for this essay. The question it asks is simple: if a company manager, an engineer, a lawyer or a nonprofit director wanted a realistic sense of how AI can help an organization today, not in some predicted future, what would these twenty-one cases teach them? And because a prospectus is a selling document, the essay asks a second question in the same breath: how much of each story can be confirmed by someone other than Anthropic?
The answer to the second question turns out to be "a good deal, but not all of it." Some cases, such as the Firefox vulnerabilities and the AIG underwriting numbers, are documented in detail by newspapers, trade magazines and the customers' own public statements. Others rest largely on Anthropic's own customer stories. A handful, including the most eye-catching claim in the retail section, appear to be published for the first time in the filing itself. Readers deserve to know which is which, and so each section below sets the prospectus's figures beside what outside sources say.
2. How to Read a Case Study in a Prospectus
Before turning to the cases, it is worth being clear about what kind of evidence they are, because the limits are built into the format.
First, every number on these pages is reported by the company or its customer, not measured by an independent party. When AIG says that time-to-quote fell by 55 percent, that is AIG's internal measurement, presented on AIG's terms. That does not make it false. Large companies have every reason not to mislead their own shareholders, and several of these figures were first announced on earnings calls where executives are also legally accountable. But the definitions behind the numbers (what counts as a "submission," which lines of business were included, what the baseline period was) are rarely stated.
Second, the figures come with "as of" dates chosen by the company. Several pages carry small footnotes such as "As of March 31, 2026," or "December 2025 to May 2026." Where an outside source gives a different figure for the same thing, the explanation is often simply that the two were measured at different times. DoorDash is an example: an interview published by Anthropic speaks of 4,000 employees using the Cowork product, while the prospectus, written later, says more than 7,000 use it every month. This essay flags such differences rather than trying to settle them.
Third, some customers are also investors. The Salesforce and SAP pages each carry a footnote disclosing that the company "beneficially owns less than 1% of Anthropic's outstanding capital stock." One percent of a company valued in the trillion-dollar range is not a small holding. This does not invalidate their results, but readers should weigh their enthusiasm accordingly.
Fourth, and most importantly, these are survivors. A prospectus does not publish case studies of customers who tried the product and walked away. A widely reported MIT study from August 2025, based on 150 interviews with leaders, a survey of 350 employees and an analysis of 300 public AI deployments, concluded that about 95 percent of corporate generative-AI pilots were producing no measurable impact on profit and loss. The twenty-one companies in this filing are, almost by definition, drawn from the other five percent. That makes them useful as examples of what is possible, but they cannot tell us what is typical.
With those limits in mind, this essay uses a simple three-level label for each case:
| Label | Meaning |
|---|---|
| Confirmed independently | The core claims appear in news reporting, trade press, or the customer's own press releases or earnings remarks, not only in Anthropic material |
| Vendor-documented | The main details come from Anthropic's or a partner's published customer stories; outside sources confirm the relationship but not the numbers |
| Prospectus only | No public source outside the filing was found for the key claims; they may well be true, but the filing is the only place they appear |
The cases are grouped into ten broad sectors: medicine, insurance, finance, law, security, retail and commerce, enterprise software, services and engineering, humanitarian work, and small business. After the sector-by-sector tour, a separate section draws out the patterns that cut across industries, and a final section asks what the case studies leave unsaid.
3. Medicine and Life Sciences: Eli Lilly and Bristol Myers Squibb
Pharmaceutical companies make an instructive first stop, because their problem is less a shortage of intelligence than a surplus of information. A large drug maker sits on decades of laboratory notebooks, clinical trial results, manufacturing quality records and commercial data, much of it stored in systems that were never designed to talk to one another. Both pharmaceutical case studies in the prospectus are, at bottom, stories about getting at knowledge the company already owns.
Bristol Myers Squibb says it began using Claude in 2023, choosing Anthropic's models for "their large context windows and high-quality reasoning," and that in May 2026 it selected Claude Enterprise as its company-wide AI platform. The rollout followed a pattern that recurs throughout the filing: software developers and data scientists first (more than 3,000 of them, according to the page), then everyone else. BMS now plans to give all of its 30,000-plus employees access to Claude Cowork, Anthropic's assistant for non-programming office work. The headline number on the page is the most striking in the medical section: the time needed to answer a "real-world evidence" question, that is, a question answered by searching external medical records and insurance-claims databases the company subscribes to, is said to have fallen from as long as two months to five minutes or less. The page also says Claude is connected to 30 to 40 internal data sources through MCP, the Model Context Protocol, an open standard that lets an AI assistant query other software systems, with "thousands more under consideration."
The broad outline of this story is confirmed outside Anthropic's materials. Trade publications reported the agreement when BMS announced it on May 20, 2026; Pharmaceutical Technology noted the plan to give "more than 30,000 employees" access to Claude and described the goal as moving "beyond basic chatbot applications" toward agents that support everyday decisions, and the health-industry news service HLTH covered the same deal across research, manufacturing and commercial operations. The specific two-months-to-five-minutes figure, however, appears only in the filing.
Eli Lilly tells a parallel story at a larger scale. According to the prospectus, Lilly started with Anthropic's models in 2024 through Amazon's Bedrock cloud service, building them into its internal AI platform and its developers' coding tools. In July 2025 it gave Claude Enterprise to about 10,000 people, mainly in research; by April 2026 the tool was available across the company, which the filing describes as a first among large pharmaceutical companies. The results box claims more than 50,000 employees with access, a fivefold expansion, and says that in the five functions that use Claude most (legal and compliance, technology, finance, oncology and research) more than 85 percent of full-time staff actively engage with it. The examples are modest and believable: biologists who once relied on spreadsheet macros now build working web applications in a couple of hours; chemists visualize complex structural data in real time. The page also describes a less glamorous but important piece of work, in which Lilly's technology team helped Anthropic build the enterprise controls a regulated company needs, including a private network connection and "data loss prevention hooks" that audit what passes through the AI.
The Lilly case is the weaker of the two on independent evidence. Lilly's chief information and digital officer, Diogo Rau, has spoken publicly about AI, and the company's relationship with Anthropic is not in doubt, but no news report or Lilly press release found for this essay confirms the 50,000-employee or 85-percent figures. They should be read as prospectus only.
One outside report adds a useful note of realism to the medical section. In a March 2026 article on how Lilly and Novo Nordisk are using AI, the payments-industry publication PYMNTS reported that Novo Nordisk, which is not one of the prospectus customers, had restricted access to an internal research tool called Found Data, "powered by Anthropic's Claude," because its "running costs outpaced the value it delivered at scale." Novo's digital transformation officer, Stephanie Bova, was quoted warning against "this mad dash to have AI use cases without thinking about whether it's the best use of your time and money." The example is a fair reminder that the same model can be a success at one company and an expensive experiment at another, depending on how it is used.
| Prospectus claim | Outside confirmation | |
|---|---|---|
| BMS | 30,000+ employees; 2 months → 5 min for evidence questions; 30–40 data sources via MCP | Deal and 30,000+ figure confirmed by trade press (May 2026); time figure prospectus only |
| Lilly | 50,000+ with access; >85% active in key functions; company-wide since April 2026 | Prospectus only for figures |
What the medical cases suggest is that the first real payoff of AI in a science-heavy company is not discovering drugs but finding things: letting a scientist ask a plain-language question across many databases and get an answer the same afternoon. That is less dramatic than an AI-designed molecule, but it is the kind of improvement that compounds across thousands of employees.
4. Insurance: AIG and Travelers
Insurance is, in a sense, the business of reading documents. Commercial underwriters must take in submissions from brokers in every format imaginable (spreadsheets, PDFs, emails, loss histories) and turn them into a judgment about risk and price. The two insurance cases in the prospectus show the same technology being used for that core work at one company and for the software that supports it at another.
AIG has the strongest independent record of any case in the filing. The prospectus says AIG began working with Anthropic in 2023, starting with the Claude 2.0 model, to build a tool called Underwriting by AIG Assist. Its purpose was to process and analyze every submission received in the lines of business where it is deployed, not just the ones underwriters had time to look at. Underwriters who once spent weeks gathering data before starting a risk summary now begin each review with a curated summary and "full audit trail" generated by Claude. In AIG's Lexington Middle Market Property business, one of its largest North American units, the filing reports that the tool let underwriters quote over 30 percent more submissions and quote them 55 percent faster, with roughly 40 percent more submissions bound (converted into policies), as of March 31, 2026.
Those figures were not invented for the prospectus. AIG's chairman and chief executive, Peter Zaffino, gave the same three numbers on an earnings call, as reported by Reinsurance News: "a 30% increase in quoted submissions, reduced time to quote for underwriters by 55%, and increased binding of submissions by approximately 40%," adding that AIG was working "in close collaboration with Palantir and Anthropic" on the next, agent-based phase. An earlier Insurance Journal report explains why coverage matters so much in this business: submissions to AIG's Lexington unit had grown to about 300,000 a year in 2024 from 30,000 in 2018, yet the unit bound only about 2 percent of them. When the pile grows tenfold and staff cannot, a tool that reads all of it changes the economics.
The prospectus also describes where AIG is heading. Its claims organization has piloted a similar tool that cut the time to respond to a client after a claim is filed "from days to hours." And it is exploring teams of specialized agents: one extracts data from submissions, another checks risk against underwriting guidelines, a third benchmarks pricing against portfolio targets, and a fourth combines their work, "with underwriters overseeing every decision." The last phrase matters. In none of the insurance examples does the AI make the final decision.
Travelers is a quieter story, centered on software engineering rather than underwriting itself. According to the prospectus, in January 2026 Travelers gave Claude assistants to nearly 10,000 engineers, data scientists, analysts and product owners, and access has since expanded to nearly 25,000 employees, with a reported 50 percent reduction in the time it takes to merge code. The January announcement is well documented: Insurance Journal reported that "nearly 10,000 engineers, data scientists, analysts and product owners at Travelers" were receiving personalized Claude assistants. The 25,000 and 50-percent figures appear only in the filing.
There is a revealing detail in the outside coverage. A Fortune profile of Travelers' technology chief, Mojgan Lefebvre, notes that one month after the Anthropic announcement, Travelers launched an AI claim assistant developed with OpenAI. Large companies, in other words, are not choosing one AI supplier for everything. They are matching tools to tasks, and a prospectus naturally shows only the tasks that went to its own company.
| Prospectus claim | Outside confirmation | |
|---|---|---|
| AIG | +30% quoted, −55% time-to-quote, ~+40% bound (Lexington Middle Market Property); 100% of submissions reviewed | Confirmed independently: same figures given by AIG's CEO on an earnings call |
| Travelers | ~10,000 users (Jan 2026) → ~25,000; −50% time to merge code | 10,000 confirmed independently; later figures prospectus only |
5. Finance, Payments and Trading: PayPal and Jane Street
The two finance cases could hardly be more different. One is a consumer payments company using AI to make thousands of ordinary employees more productive; the other is a privately held trading firm using it to multiply the output of a small number of highly paid researchers.
PayPal's page describes a broad, practical rollout. More than 12,000 PayPal employees are said to use Claude Enterprise monthly, across product, technology and sales. Product teams built their own MCP connection to company data so they can create analyses and dashboards themselves instead of waiting for engineers or data analysts. Merchant underwriting and compliance reviews, which once required pulling information together from several separate tools, now take "a fraction of the time." Since August 2025 the sales team has used a custom plugin for pipeline management and account-risk review, saving an average of 45 minutes per person per week. Account managers use an internal system built on Claude to get merchant health diagnostics in minutes, cutting report preparation time to five to ten minutes. None of these numbers is large on its own. What is notable is how ordinary the uses are: reports, reviews, dashboards and sales administration.
The relationship between PayPal and Anthropic is well documented in PayPal's own announcements, though not the internal productivity numbers. In May 2025, PayPal's developer blog announced that its remote MCP server supports Anthropic's API, allowing AI assistants to carry out payment tasks. A year later, in May 2026, PayPal's newsroom announced a partnership with Anthropic aimed at small businesses, citing its own survey in which 82 percent of small businesses said adopting AI is essential to staying competitive while 73 percent said they lacked the time or know-how. The partnership pairs a free nine-lesson AI fluency course with a plugin that lets Claude handle tasks such as invoicing and refunds through PayPal. The 12,000-user and 45-minute figures are prospectus only.
Jane Street is the most intriguing case in the section and the hardest to check. The firm is a large quantitative trading house whose advantage depends on the speed and quality of its research. According to the prospectus, its developers spent years trying to build AI agents that could carry out research tasks on their own, and early attempts were "frustrating and inconsistent." That changed with Claude 3.7 Sonnet, when agents began completing long, multi-step tasks without stalling. Today, the page says, Claude is deployed "nearly wall-to-wall," researchers supervise teams of agents that analyze code, monitor model training runs and process results, and experiments that once needed constant attention now run unattended over a weekend. One researcher estimates completing about 50 percent more analysis than a year earlier. Another used the newest model, Claude Fable, to reorganize a sprawling Python application and then migrate it to a more robust language in about a week, work the page says would normally take a team months. On Jane Street's own harder internal coding test, Claude Fable 5.1 is said to have scored 32 percent better than its predecessor.
Two cautions apply. First, nothing in the public record found for this essay describes Jane Street's internal use of Claude; the story is prospectus only. Second, Jane Street is famous for writing most of its software in OCaml, an unusual programming language. A 2025 eFinancialCareers report quoted a claim that there is "more OCaml code in Jane Street than there is in the rest of the world combined," which helps explain why off-the-shelf coding assistants trained mostly on popular languages long struggled there. If the prospectus is right that recent models have overcome that, it is a meaningful data point. But it is the company's account, not an outside measurement.
| Prospectus claim | Outside confirmation | |
|---|---|---|
| PayPal | 12,000+ monthly users; 45 min/week saved per salesperson; report prep cut to 5–10 min | Partnership confirmed independently (PayPal newsroom, developer blog); figures prospectus only |
| Jane Street | ~50% more analysis per researcher; codebase migration in ~1 week; +32% on internal eval | Prospectus only |
Read together, the finance cases show two quite different models of AI value. At PayPal it is spread thinly across many people, saving minutes each week on routine work. At Jane Street it is concentrated in a few experts whose output is multiplied. Most organizations will find their own situation somewhere between the two.
6. Law: Legora
Legal work is a natural fit for language models: it is made of text, governed by precedent and sharply divided between high-value judgment and many hours of reading. The prospectus includes one legal company, and it is not a law firm but a software maker that sells to law firms.
Legora describes itself as an "agentic operating system for lawyers," used by more than 1,500 law firms and in-house legal teams. According to the prospectus, Claude has powered parts of the platform since its launch in October 2024, after Legora ran its own comparison of available models and found that Claude did best on the dimensions it cared about most: prioritizing and following instructions, and adapting as client requirements changed. Inside Legora's product, Claude acts as part of the "reasoning layer," planning and delegating work across drafting, document review and legal research, and, importantly for a profession that lives on citations, "grounding every answer in a source before a lawyer sees it." The filing singles out transactional drafting as the area where the newest models brought the largest gains.
Two examples on the page give a sense of scale. On one merger, Legora's team reviewed a 900-document data room spanning four jurisdictions, a first pass that a conventional team would spend a week on, in half a day. And across its client base, lawyers in a study of 31 firms in 14 countries reported recapturing an average of 4.3 non-billable hours a week, which the page translates into "$6.9 million in additional annual billing per 100 lawyers." Legora also uses Claude to build its own product: an engineering team of about 80 supports more than 1,500 client organizations, AI tools generate well over half of its code, some engineers run up to ten background coding agents at once, and Claude Code use inside the company grew fourteen-fold, from about 50 seats to more than 700.
Most of this is vendor-documented. Anthropic's Legora customer story confirms that Claude powers Legora's "assistant tools, document review, and intelligent workflows." One detail deserves a closer look, however. The prospectus calls the 31-firm study "independent." Legora's own report on the research says that the legal-industry analyst Ari Kaplan "partnered with Legora to conduct a comprehensive qualitative survey," interviewing 31 firms over three months, and the legal-technology site Artificial Lawyer reported the findings as "based on 31 law firm interviews." An outside analyst ran it, but it was a qualitative study made in partnership with the vendor, built on interviews rather than measured time records. "Independent" is a generous word for that.
There is also a competitive twist worth knowing about. On May 12, 2026, Anthropic launched its own legal offering, Claude for Legal, which Legal Business described as a shift from "primarily powering legal AI products such as Harvey and Legora, towards potentially becoming a legal platform in its own right." Legora's page in the prospectus is therefore a case study of a customer that is also, increasingly, a competitor. That is common in the AI industry and worth remembering whenever a model maker publishes a success story about a company building on top of it.
| Prospectus claim | Outside confirmation | |
|---|---|---|
| Legora | 1,500+ organizations; 4.3 hrs/week recaptured (31 firms); 900-doc data room in half a day; Claude Code 50 → 700 seats | Use of Claude vendor-documented; the 31-firm study was run with Legora, not independently of it |
7. Security and Critical Infrastructure: Mozilla and Verizon
If any section of the prospectus moves beyond productivity into something like public interest, it is this one. Software vulnerabilities are the raw material of cyberattacks, and they have always favored the attacker: a defender must find every flaw, an attacker needs only one.
Mozilla, maker of the Firefox browser, is the best-documented case in the entire filing. According to the prospectus, Anthropic's internal "Frontier Red Team" approached Mozilla to test Claude on defensive security work. In a two-week evaluation, Claude Opus 4.6 found 22 previously unknown vulnerabilities in Firefox's most critical code, 14 of them rated high-severity; Mozilla's engineers began landing fixes within hours and shipped them in Firefox 148. Two months later, Mozilla shipped fixes for a further 271 vulnerabilities identified by Claude Mythos Preview, a more capable model Anthropic has not released to the public. The page quotes Firefox's chief technology officer, Bobby Holley: "So far we've found no category or complexity of vulnerability that humans can find that this model can't."
Both rounds were covered widely. TechCrunch and Computerworld reported the 22 bugs, 14 high-severity, and their fixes in Firefox 148. The second round received closer, and more critical, scrutiny. SecurityWeek noted that while more than 40 vulnerabilities were addressed in Firefox 150's official advisory, "only three are credited to Claude," which suggested that "many of the 271 bugs are likely lower-severity issues or flaws that don't meet the threshold" for a public listing. Expert Insights made the same point under the headline "Critics Still Aren't Convinced." SecurityWeek also reported Mozilla's own measured view: it had not seen any bugs that "couldn't have been found by an elite human researcher." And The Decoder added that some of the flaws were up to 20 years old, and that Mozilla's pipeline lets the model write and run its own test cases to verify what it finds.
Put together, the record supports a careful version of the prospectus's claim. AI did find real, previously unknown bugs in one of the most heavily tested codebases in the world, quickly and at scale. Most of them were not dramatic, and none was beyond what the best human experts could have found with enough time. The value lies in the "enough time": bugs that would have sat unnoticed for years were found in weeks.
Verizon is a much thinner page. It says Verizon chose Claude Code for its technology and product organizations after evaluating coding tools for "security, reliability and scale," and that more than 33,000 engineers now build with it. It then notes that Verizon became one of the first telecom companies to join Project Glasswing, Anthropic's program giving vetted organizations early access to Claude Mythos for defensive security. Verizon's own newsroom confirms the Glasswing membership and says Verizon "is testing Claude Mythos Preview to gain insights to maintain the network's security for its customers." The 33,000-engineer figure was not found outside the filing.
Project Glasswing itself is documented on Anthropic's site, where the company says launch partners are using Mythos Preview "as part of their defensive security work." In an update announcing an expansion to about 150 additional organizations in more than 15 countries, Anthropic said partners had so far found "more than 10,000 high- or critical-severity security flaws." That figure is Anthropic's own.
| Prospectus claim | Outside confirmation | |
|---|---|---|
| Mozilla | 22 vulnerabilities (14 high) by Opus 4.6; 271 more by Mythos Preview | Confirmed independently, with the caveat that only 3 of the 271 received individual CVE credit and Mozilla says none was beyond elite human ability |
| Verizon | 33,000+ engineers on Claude Code; Glasswing member | Glasswing confirmed independently; engineer count prospectus only |
8. Retail and Commerce: Walmart, Shopify, Mercado Libre, Rakuten and DoorDash
The largest group in the prospectus is made up of companies that sell things, online, in stores or through delivery. Their stories divide into two kinds of AI use: assistants that face customers, and coding agents that change how the companies build their own software.
Walmart is the most important case in the group and the least confirmed. The prospectus says Anthropic "partnered with Walmart on Sparky," Walmart's AI shopping assistant, and that Anthropic's models, "like Claude Sonnet 4.6," help Walmart create personalized experiences as customers describe their needs in their own words. The results box reports more than 110 percent quarter-over-quarter growth in Sparky's weekly active users, an increase of about 35 percent in average order value for orders placed through Sparky compared with those that were not, and more than twice as many units sold through Sparky quarter over quarter. It also mentions that Claude is one of the models behind WIBEY, Walmart's internal engineering productivity agent.
Sparky's commercial results are partly confirmed: eWeek reported in February 2026 that customers who engage with Sparky "have an average order value 35% higher than those who do not." But Walmart has never publicly said which models power Sparky. Its own engineering blog post introducing WIBEY describes the agent's architecture, built on Walmart's in-house platform called Element and using open protocols such as MCP and A2A, without naming any model supplier. Walmart has had publicized shopping partnerships with other AI companies as well. As far as this essay could find, the prospectus is the first public statement that Claude plays a role in Sparky, which makes it newsworthy in itself, and also makes it a claim that rests on Anthropic's word alone.
Shopify is the clearest example of AI built into a product that millions of people use without knowing which model is underneath. Its Sidekick assistant helps merchants analyze their business, plan promotions and set up automations, work that once required an analyst. The prospectus says daily active merchants using Sidekick more than tripled year over year, that nearly half of all Shopify Flow automations created in a recent quarter were built with Sidekick's help, and that new merchants reach their first sale "in days instead of weeks." Inside the company, Claude Code is available to more than 7,500 employees. Both Anthropic's and Google Cloud's customer pages confirm that Shopify runs Claude on Google's Vertex AI service to power Sidekick, so the relationship is well vendor-documented from two independent vendors, though the growth figures come from the filing.
Mercado Libre, Latin America's largest e-commerce and fintech company, offers the most dramatic numbers in the filing, and also the most consequential claims about jobs. According to the prospectus, more than 23,000 developers use Claude Code; over the past year, agents with human oversight reviewed more than half a million code submissions and migrated roughly 9,000 services to newer platforms; in the second quarter of 2026, code submissions rose 110 percent year over year, merged code more than doubled and deployments rose nearly 75 percent, while rollbacks fell. In 2025 about 40 percent of the code Mercado Libre shipped was written by Claude and reviewed by a human; by 2026, the page says, "human-written code has become the exception." Outside engineering, more than 26,500 people in finance, marketing, HR, legal, compliance and commercial teams use Claude Cowork, 78 percent of them weekly, and finance has shortened its monthly close by a third.
The engineering figures match what Mercado Libre told investors. A Zacks summary of the company's second-quarter results reports management saying that about 20,000 developers use AI, "with human-written code becoming the exception," that code submissions rose 110 percent, merged code more than doubled and rollbacks declined, and that AI agents "independently reviewed more than half a million code submissions and migrated roughly 9,000 services." What the investor material does not do is name a supplier, and the company is known to use more than one tool: a marketing-analytics case study from Amplitude mentions employees working through "Claude, Cursor, or their internally developed AI chatbot." So the scale is confirmed and the Claude-specific attribution is the prospectus's. The page also says Mercado Libre plans to fold separate backend, frontend and database roles into a single "builder" role. Early in 2026 Bloomberg Línea reported layoffs in the company's user-experience design teams, describing them as part of a shift toward more generalist profiles; the company said the cuts affected fewer than 0.09 percent of its more than 125,000 employees.
Rakuten, the Japanese internet and e-commerce group, supplies the filing's best-known coding anecdote: Claude Code working autonomously for seven hours on a complex, multi-language codebase with 99.9 percent accuracy. The page adds that time to market for new features fell by more than 75 percent, that Rakuten now ships major releases every two weeks instead of quarterly, and that it deployed specialized Claude agents across product, sales, marketing and finance, each within a week. Anthropic's own Rakuten customer story gives the details: the seven-hour session was an open-source refactoring task, the 99.9 percent was "numerical accuracy compared to the reference method," and time to market fell 79 percent, from 24 days to 5. It also quotes the engineer involved: "I didn't write any code during those seven hours. I just provided occasional guidance." A critical analysis by the software consultancy Wavect makes the obvious but useful point that this was one well-scoped task, not a controlled study, and that "autonomous" should be read as "the publisher's description of the session, not proof of zero human intervention." The case is vendor-documented, and its headline is accurate but narrower than it sounds.
DoorDash rounds out the group with a story about spreading AI tools to every employee. The prospectus says Claude Code and Claude Cowork are available to every corporate employee, more than 7,000 people use Cowork each month and almost half of them daily, launch timelines for major products have been cut by more than half, and one engineer finished a code migration in three weeks that would once have taken four engineers more than a quarter. Designers and product managers now ship production code alongside engineers. An interview published by Anthropic with DoorDash's co-founder and chief technology officer, Andy Fang, confirms the broad picture, though it gives a smaller Cowork figure ("all 4,000 employees have Cowork"), probably measured earlier. Fang's own account is perhaps the most human detail in the retail section: years after moving into management, he says, "it wasn't until I was using Claude Code that I actually was shipping production code in our code base again."
| Prospectus claim | Outside confirmation | |
|---|---|---|
| Walmart | Claude in Sparky; +35% order value; WAU +110% QoQ | +35% order value confirmed; Claude's role prospectus only (Walmart has not named Sparky's models) |
| Shopify | Sidekick on Claude; daily active merchants ×3; 7,500+ on Claude Code | Claude-in-Sidekick vendor-documented (Anthropic and Google Cloud) |
| Mercado Libre | 23,000+ devs; 500k reviews; 9,000 migrations; human code "the exception" | Scale confirmed in investor remarks; vendor not named there |
| Rakuten | 7 hrs autonomous; 99.9%; 24 → 5 days | Vendor-documented; "occasional guidance" during the 7 hours |
| DoorDash | 7,000+ monthly Cowork users; launches cut by half | Vendor-documented (4,000 in an earlier interview) |
9. Enterprise Software Platforms: Salesforce and SAP
The two enterprise-software cases matter for a different reason from the rest. Salesforce and SAP sell the systems on which a large share of the world's companies run their sales, finance, procurement and human resources. When they build Claude into those systems, AI reaches millions of workers who never chose it, never signed up for it and may not know it is there.
Salesforce reports gains both in building its own software and in its products. According to the prospectus, more than 60 percent of its 83,000 employees use Claude. Over the past year pull requests merged per developer rose 79 percent, lifting "effective output per developer" by 151 percent, while incidents fell 5 percent. One product team used a Claude-based framework to migrate 33 API endpoints in 13 days, against an estimate of 231 person-days. Claude also powers Slackbot, the assistant built into Slack, which nearly every Salesforce employee uses and which the company estimates is generating 8.1 million hours of annualized productivity gains. On the product side, Salesforce named Claude a preferred model for its Agentforce 360 platform in October 2025, and that announcement is confirmed in Salesforce's investor release, which describes Anthropic as "the first LLM provider fully integrated within the Salesforce trust boundary, with all of Claude's traffic contained in the Salesforce virtual private cloud." That arrangement, keeping the AI's data traffic inside Salesforce's own security perimeter, is what lets banks, hospitals and drug companies use it, and it is a good example of how much of enterprise AI adoption is really about governance rather than intelligence.
SAP frames its use in the language of the "Autonomous Enterprise," a future in which "people set direction and policy, and AI executes work across the organization." Claude is one of several models SAP routes queries to through its Joule assistant and Business AI Platform, and SAP says its more than 50 domain-specific Joule assistants and 200-plus specialized agents use Claude for some of their reasoning: closing the books at quarter-end, answering employees' leave questions, rerouting supplier orders mid-shipment. The most concrete example is Joule for Consultants, a tool for the consultants who install SAP systems, which the page says saves up to 1.5 hours a day and lets teams deliver projects up to 14 percent faster. Both figures appear in an AWS blog post written with SAP, which says the tool uses "Anthropic's Claude models via Amazon Bedrock." In May 2026 The Register reported SAP's announcement of a partnership with Anthropic to bring Claude into the Business AI Platform, quoting one commentator who observed that SAP was offering Claude "as something within a 'walled garden' of SAP."
Both companies disclose on their prospectus pages that they own shares in Anthropic. That does not make their figures wrong, and several are confirmed in their own public filings and releases, but it does mean they are not neutral witnesses.
| Prospectus claim | Outside confirmation | |
|---|---|---|
| Salesforce | 60%+ of 83,000 staff; +151% effective output/dev; 231 person-days → 13 days; Claude in Slackbot | Preferred-model status and trust-boundary design confirmed independently; productivity figures prospectus only; Salesforce is a shareholder |
| SAP | Joule for Consultants: up to 1.5 hrs/day saved, projects up to 14% faster | Confirmed in an AWS/SAP technical blog; SAP is a shareholder |
10. Professional, IT and Engineering Services: Accenture and UST
Consulting and outsourcing firms occupy an unusual position in the AI economy. They are customers, using AI on their own work, and they are also distributors, carrying it into thousands of client companies that will never deal with a model maker directly. Both service firms in the prospectus describe that double role.
Accenture, with more than 800,000 employees, is the larger of the two. In December 2025 it and Anthropic announced what Accenture's own press release called a new business group, under which Accenture would "train approximately 30,000 professionals" on Claude and make Claude Code available to "tens of thousands of its developers," along with a jointly funded Claude Center of Excellence. Anthropic's announcement of the same deal described it as creating "one of the largest ecosystems of Claude practitioners in the world." By the time of the prospectus, the results box reports 45,000 Accenture professionals trained, more than the original target, and over 18,000 certified Claude practitioners.
The page's operational examples are specific. Accenture's security product Cyber.AI, which uses Claude as its reasoning engine, was deployed across Accenture's own global IT estate to help secure 1,600 applications and more than 500,000 APIs (the interfaces through which software systems exchange data). Scan turnaround fell from days to under an hour, and test coverage rose from about 10 percent to more than 80 percent. For a large U.S. broadband provider served jointly by Accenture and Anthropic, the client reported a sevenfold first-year return on its AI investment after rebuilding how it markets and sells to business customers: speed to market rose 55 percent, the cost of validating a sales lead fell 86 percent, and lead-validation accuracy rose from 13 percent to 97 percent. The partnership and training figures are confirmed independently; the client results are prospectus only, and the broadband client is not named.
UST is less well known but supplies what may be the most technically interesting case in the filing, because it reaches into the physical world. UST builds and runs technology for banks, telecom companies, healthcare organizations and manufacturers. Its showcase is a hardware validation platform called iDEC, used to prove that a computer chip behaves the way its designers intended before it goes into production. According to the prospectus, Claude Code reads chip schematics directly, then generates and runs regression tests to check what a design change affects downstream. It also compares live equipment data against a "digital twin," a virtual replica of the hardware, flagging signal-integrity problems before they reach production. UST reports that chip validation cycle time has fallen by 50 to 70 percent and that standard validation turnaround has dropped from four days to two.
UST then embeds the same approach in platforms it runs for clients: CarePath, which turns scattered insurance-claims and care data into recommended next steps for health insurers, "with human approval before any member-facing recommendation"; IntelliOps, which predicts cellular network failures and reduces alert noise for telecom operations teams; and FinX, which puts agents into bank operations. And it is training 20,000 engineers, architects, consultants and industry specialists on Claude. Anthropic's own announcement of the partnership in July 2026 describes iDEC, CarePath and IntelliOps in the same terms and confirms the 20,000 training figure; the trade publication Channel Insider covered the announcement independently. The 50–70 percent reduction in chip validation time is the company's own figure.
For readers who work in engineering, UST is the case to study. It shows a language model doing something closer to engineering than to office work: reading technical drawings, reasoning about what a change will break, writing the tests and checking the results against a model of the physical system. It is also the case where the human approval step is spelled out most clearly.
| Prospectus claim | Outside confirmation | |
|---|---|---|
| Accenture | 45,000 trained; Cyber.AI test coverage 10% → 80%+; broadband client 7× ROI | Partnership and ~30,000 training target confirmed independently; client results prospectus only |
| UST | Chip validation cycle −50–70%; 4 → 2 days; 20,000 to be trained | Partnership, platforms and 20,000 confirmed (Anthropic release, trade press); time figures prospectus only |
11. Humanitarian Work: The International Rescue Committee
The case that most clearly departs from the commercial pattern is that of the International Rescue Committee, one of the world's largest humanitarian organizations. The IRC works in more than 40 countries affected by conflict and disaster; the prospectus says it helped more than 20.5 million people in 2025. Its problem is the opposite of a pharmaceutical company's. Where BMS has too much data in too many systems, the IRC often has too little usable data, scattered across field offices, paper forms and spreadsheets, in places where every staff hour is precious.
The IRC joined Anthropic's Claude for Nonprofits program in 2025, which provides discounted licenses. According to the prospectus, use grew from a 300-person pilot to about 2,400 seats, an eightfold increase between December 2025 and May 2026, on the way to all 10,000 staff who work at computers. Half of the pilot users reported saving four or more hours a month, and the page estimates 464 staff hours returned per week. A country-wide funding landscape analysis that once took a week now takes a couple of hours, and a portfolio synthesis that took a month of manual spreadsheet work is available immediately.
The more important story is about frontline care. Health teams in Africa used Claude to organize scattered field data, and the share of key health indicators actually used in strategic planning rose from 20 percent to 100 percent. With that visibility, staff could spot problems that had not been readily detectable before, such as seasonal spikes in infection or gaps in mothers' use of maternal care, and adjust their services. In a malnutrition project, field teams photograph paper treatment records and Claude aggregates them into a picture of each child's treatment status. The IRC's senior director for policy and planning, Katherine Rodrigues, is quoted describing the goal as "shifting from a reactive system to anticipating which children are at risk of dropping out of malnutrition treatment and intervening earlier."
The outlines are confirmed. When Anthropic launched Claude for Nonprofits on GivingTuesday 2025, offering discounts of up to 75 percent, it named the IRC as an early user that was "using Claude to communicate with local partners and analyze field data faster in time-sensitive humanitarian settings," and NBC News reported the IRC as "one of the early beneficiaries" of the program. Anthropic's separate IRC customer story gives a different-looking headline number, "16.8 million people annually through more than 2,700 health facilities in 40 countries," against the prospectus's 20.5 million. The two are probably not in conflict: one describes people reached through the IRC's health facilities, the other everyone the organization helped in 2025. But the difference is a reminder to check what a number is counting before comparing it with another. The case is vendor-documented, with the relationship confirmed by independent press.
The IRC page is the clearest example in the filing of AI being used not to save money but to make new work possible. In the IRC's own framing, the point was not that staff did the old work faster but that they could now see things, children slipping out of treatment, clinics missing mothers, that they could not see before.
12. Small Business: Shy Bird and KBF Design Gallery, and a Note on Construction
The prospectus ends its case studies with two businesses that could hardly be smaller beside the corporations before them. They share a single page, and they make, in some ways, the most persuasive argument in the section, because neither had an IT department, a data science team or a consultant.
Shy Bird is a rotisserie and café in the Boston area with three locations. Its brunch service near Fenway Park suffered long waits, so co-owner Eli Feldman, who by the prospectus's account "had never written a line of code," photographed the kitchen, pulled sales and ticket data, collected his cooks' observations, and gave it all to Claude. The bottlenecks it found were "hiding in plain sight." An avocado purée that went on nearly every order sat out of reach, while a pan of overnight oats, ordered only a few times a morning, held the best spot on the line. Feldman and his cooks redesigned the line together, and brunch ticket times fell by more than 60 percent, to 12 minutes. The experience convinced the owners that the same approach could help other restaurants, and their in-house tool became a product called Colette, which the prospectus says is now used by 60 restaurants. It also reports that one Shy Bird server became a full-time back-end engineer at Colette within six months.
The Shy Bird story is confirmed independently in its essentials. NBC Boston profiled Feldman and Colette in June 2026, reporting that "about three dozen restaurants across the country currently use Colette," fewer than the prospectus's later figure of 60 but consistent with a young product that is growing. Anthropic's own Shy Bird story adds that the company encourages every manager to use Claude under an internal philosophy Feldman calls "BITE: Bring It To Everything."
KBF Design Gallery is a family-owned interior design and remodeling firm in the Orlando area specializing in custom kitchens, bathrooms and whole-home renovations. When siblings Adam and Ashley Vellequette took over from their parents, they faced what the prospectus calls a common problem in the industry: talented designers buried in administrative work. Using Claude Code and Claude Cowork, the firm built its own tools for scheduling, for validating the handoff from design to construction, and for writing contract scopes. The results box says the tools recovered 490 hours a year for a single employee, the lead designer, "nearly a quarter of a working year." Writing the contract scope now takes seven fewer hours per project, and a team of only three designers now handles 70 designed projects a year, with sales up 32 percent since the firm adopted Claude in 2025. No public source outside the filing was found for the KBF story, so it is prospectus only, though the firm itself is real and has won local design awards.
A note on construction
KBF is the only company in the prospectus that touches construction, and it does so at the scale of residential remodeling. There is no general contractor, no civil or structural engineering firm and no large engineering-procurement-construction company among the twenty-one cases. That absence is itself informative, and it matches what surveys of the construction industry have found in 2026: real but uneven adoption, concentrated in office work rather than on the jobsite.
The evidence points in somewhat different directions. The ServiceTitan 2026 commercial contractor report, based on a survey of more than 1,000 construction leaders, found that 38 percent of contractors now report measurable business impact from AI, up from 17 percent a year earlier. A survey by the construction software company CMiC and the data firm Dodge, covering more than 6,000 companies and reported by Construction Executive, found that 85 percent of contractors expect to spend less time on repetitive tasks and over 70 percent believe AI will help them make better decisions. In contrast, a study by the tool maker DeWalt, reported by Construction Equipment, reported that only 8 percent of the roughly 3,400 construction professionals it surveyed worldwide use AI today, and identified lack of practical training as the main obstacle. All three surveys were sponsored by companies that sell to the industry, and they define "use" differently.
Read alongside the prospectus, the surveys suggest where the first gains for construction and engineering firms are most likely: not in the field but in the document-heavy work around it. That is exactly where KBF found them: scheduling, scope writing and checking that what was designed is what gets built. It is also where UST's chip-validation work offers a closer analogy for engineers than any of the office-productivity stories. Reading drawings, tracing the downstream effects of a change and generating checks against a model of the physical system are tasks that structural, civil and mechanical engineers will recognize.
13. Seven Patterns Across the Sectors
Reading the twenty-one cases by industry shows who is using AI. Reading across them shows what kind of work it is doing. These patterns may be more useful to a reader whose own company does not appear in the filing.
First, writing and maintaining code at industrial scale. Almost every company in the filing uses Claude Code, and the most measurable gains are here: Mercado Libre's 9,000 migrated services, Salesforce's 231 person-days compressed into 13, DoorDash's single engineer doing in three weeks what four once needed a quarter for, and Rakuten's release cycle shrinking from quarterly to every two weeks. Software is a natural first target because results can be tested automatically. A migration either passes its tests or it does not. Much of this work is maintenance that engineers have always disliked: version upgrades, dependency patches, test coverage. The gain is less that AI writes brilliant new code than that it takes on the backlog nobody had time for.
Second, reading the pile nobody had time to read. AIG's underwriters, Legora's data-room reviews, PayPal's compliance checks and the IRC's field records share a structure: a volume of documents larger than the staff available to read them. In several cases the important change is not speed but coverage. AIG says it now reviews 100 percent of applicable submissions, not just the ones underwriters could reach. When an organization has been making decisions on a sample because it could not afford to look at everything, AI changes what it can know, not just how fast it knows it.
Third, breaking down data silos. BMS's 30 to 40 connected data sources, PayPal's custom data connector, Lilly's private network connection and Salesforce's Slack integration are all applications of the same idea: let one assistant query many systems in plain language. The open Model Context Protocol appears again and again for this reason. In practice, much of the hard work in these cases was not choosing a model but connecting it, securely, to the company's existing data.
Fourth, assistants inside someone else's product. Shopify's Sidekick, Walmart's Sparky, Salesforce's Slackbot and SAP's Joule put AI in front of millions of merchants, shoppers and office workers who did not choose a model and mostly do not know which one they are talking to. This is probably where AI's broadest reach is today, and also where the prospectus's claims are hardest to confirm, because the companies selling those products often prefer not to name their suppliers.
Fifth, defence. Mozilla's Firefox work, Verizon's network security testing and Accenture's Cyber.AI point to a use whose value goes beyond the companies involved. If AI helps defenders find old vulnerabilities faster than attackers can exploit them, everyone who uses that software benefits. The same capability, of course, could also help attackers, which is why Anthropic restricts its most capable security model to vetted partners.
Sixth, non-programmers building their own tools. A restaurateur who had never coded redesigning his kitchen from photographs and sales data; a remodeler's family building scheduling and contract tools; Lilly's biologists building web applications to replace spreadsheet macros; DoorDash's designers shipping production code; Mercado Libre procurement staff building agents to answer supplier questions without involving IT. This may be the most far-reaching pattern in the filing, because it widens the circle of people who can make software at all.
Seventh, jobs being redrawn. Several cases describe changes to roles rather than just to tasks. Mercado Libre plans to merge backend, frontend and database roles into a single "builder" role. Rakuten says roles once split across engineering, design and reliability have converged. Jane Street's and DoorDash's designers write production code; DoorDash's CTO and 30 percent of Mercado Libre's managers are coding again. These are the most consequential claims in the section and the ones a prospectus is least likely to examine critically. The only concrete outside evidence found for this essay, Bloomberg Línea's report of design-team layoffs at Mercado Libre, involved a small number of people. But the direction the companies describe is clear.
14. What the Case Studies Do Not Say
A fair reading of the prospectus also needs to note what its case studies leave out.
They do not include failures. The MIT study cited earlier, which found that about 95 percent of corporate generative-AI pilots produced no measurable effect on profit and loss, also found that purchasing tools from specialized vendors and forming partnerships succeeded about 67 percent of the time, while internal builds succeeded only a third as often, and that the largest returns came from "back-office automation." That is consistent with the prospectus cases, most of which are partnerships focused on internal operations, but it also means the filing shows the winning end of a very uneven distribution. Novo Nordisk's decision to restrict a Claude-powered research tool because of cost, reported by PYMNTS, is exactly the kind of story a prospectus would not include.
They rely heavily on self-reported productivity, and that is a known weak spot. In a randomized controlled trial published in July 2025, the research group METR assigned 246 real tasks to 16 experienced open-source developers, allowing AI tools on some tasks and not others. With AI, the developers took 19 percent longer. Yet, as METR's own summary notes, they had expected a 24 percent speed-up and, even after the experiment, "still believed AI had sped them up by 20%." The study used early-2025 tools on large, mature codebases that the developers knew intimately, and METR itself cautioned against generalizing from it; the models in the prospectus are considerably newer. But it is a useful warning that when people estimate their own time savings, as in several of the prospectus figures (Legora's 4.3 hours, the IRC's four hours a month, PayPal's 45 minutes), perception can differ from measurement. The figures most worth trusting are the ones counted by systems rather than estimated by people: AIG's quote counts, Mercado Libre's merged code and rollbacks, Mozilla's patched bugs.
They mostly describe augmentation, not replacement. Anthropic's own research tracks how Claude is used. Its Economic Index report of January 2026 found that on the consumer Claude.ai service, "augmentation is once again more common than automation," while business use through the API is "automation-dominant" and increasingly used "to automate routine back-office workflows." The prospectus cases fit that split. Nearly every page includes a phrase like "with underwriters overseeing every decision," "reviewed by a human engineer before release" or "with human approval before any member-facing recommendation." Whether that human role remains as large as models improve is a question the case studies raise but do not answer.
Finally, the numbers shift between sources, usually in ways that are easy to explain but worth noticing: DoorDash's 4,000 or 7,000 Cowork users; Colette's three dozen or 60 restaurants; the IRC's 16.8 or 20.5 million people; Legora's "independent" study conducted in partnership with Legora; Rakuten's "autonomous" session with "occasional guidance"; Mozilla's 271 vulnerabilities, of which three received individual public credit. None of these discrepancies suggests that a case is false. Together they show why a prospectus case study is best read as the company's best honest account of itself, and checked wherever possible.
15. Conclusion: What a Company Can Reasonably Expect Today
Set side by side, the twenty-one case studies in Anthropic's prospectus support a picture that is more modest than the most excited predictions about AI and more substantial than the most dismissive ones.
What is well established, by evidence outside the filing, is that AI is already doing real work at scale in a handful of areas. It is writing and maintaining large amounts of production software under human review. It is reading document volumes that staff could never cover: insurance submissions, legal data rooms, field health records. It is finding real security flaws in heavily tested code. And it is serving as the engine of assistants used by millions of merchants and shoppers. The AIG and Mozilla cases in particular are about as well documented as corporate technology claims get.
What is plausible but rests mainly on the companies' own word are the large productivity multipliers: 151 percent more output per developer, months of work in a week, two months of research in five minutes. They may be true; several come from companies that also report them to their own investors. But they are measured by the people who benefit from them, using definitions that are not disclosed.
What remains open is the human question. Nearly every case keeps a person in the loop, and nearly every case also describes that person's role changing, toward supervising agents, reviewing their work and doing the parts that require judgment. A few cases describe roles merging or disappearing. The prospectus presents this as reinvention. It may be that, and it may also be disruption for the people whose roles are being reinvented; most likely it will be both, unevenly.
For a manager trying to decide what to do on Monday morning, the most practical lesson in these twenty pages may come from the smallest company in them. Shy Bird did not start with a strategy or a platform. Its owner started with a specific, measurable problem (brunch tickets were slow), gathered the evidence he already had (photographs, sales data, the cooks' knowledge), and asked for help reading it. The answer turned out to be a misplaced tub of avocado purée. Most of the large companies in the filing followed the same path at greater scale: a specific bottleneck, existing data, a human who decides, and a measurement afterward. That is less glamorous than the "Autonomous Enterprise," but it is what the evidence, read carefully, shows AI can already do.
References
Primary source
- Anthropic, PBC. Form S-1/A, Amendment No. 2, filed with the U.S. SEC, November 10, 2026. Case studies on printed pages 229–248.
- Sarin, Natasha. "The Flaw at the Heart of Anthropic's I.P.O." The New York Times, November 15, 2026.
Medicine
- Bristol Myers Squibb Enters into Agreement with Anthropic. Pharmaceutical Technology, May 2026.
- Anthropic and Bristol Myers Squibb Expand Enterprise AI Collaboration. HLTH, May 21, 2026.
- Lilly and Novo Show How AI Is Rewiring Big Pharma. PYMNTS, March 31, 2026.
Insurance
- AI Advancing Faster than Expected as AIG Builds Multi-Agentic Solution: CEO Zaffino. Reinsurance News.
- AIG: Turning One Human Underwriter Into Five, 'Turbocharging' E&S. Insurance Journal, April 30, 2025.
- 10,000 Travelers Employees Get AI Assistants Via Anthropic Partnership. Insurance Journal, January 20, 2026.
- Why insurance giant Travelers' CTO is placing fewer, bigger bets on AI. Fortune, April 15, 2026.
Finance
- PayPal partners with Anthropic to Close the AI Gap for Small Businesses. PayPal Newsroom, May 2026.
- PayPal Remote MCP Server Now Supports Anthropic API. PayPal Developer Blog, May 22, 2025.
- How AI is affecting careers in electronic trading and HFT. eFinancialCareers.
Law
- Legora helps lawyers work more efficiently with Claude. Anthropic customer story.
- The Legora ROI report: law firms. Legora.
- Despite The Whale, You Can See Legal AI ROI. Artificial Lawyer, March 31, 2026.
- Cox, Tom. 'One engine, configured a thousand different ways': is Claude for Legal the next big shift in legal tech? Legal Business, May 28, 2026.
Security
- Anthropic's Claude found 22 vulnerabilities in Firefox over two weeks. TechCrunch, March 6, 2026.
- Claude found 22 vulnerabilities in Firefox in two weeks. Computerworld, March 10, 2026.
- Kovacs, Eduard. Claude Mythos Finds 271 Firefox Vulnerabilities. SecurityWeek, April 22, 2026.
- Claude Mythos Finds 271 New Firefox Vulnerabilities – But Critics Still Aren't Convinced. Expert Insights.
- Schreiner, Maximilian. Mozilla's agentic AI pipeline turns Claude Mythos Preview loose and finds 271 unknown Firefox vulnerabilities. The Decoder, May 8, 2026.
- Verizon joins Anthropic's Project Glasswing. Verizon News.
- Project Glasswing and Expanding Project Glasswing. Anthropic.
Retail and commerce
- Curry, David. Walmart: 'Sparky' AI Chatbot Is Lifting Spend, Engagement. eWeek, February 23, 2026.
- From models to agents: A new era of intelligent systems at Walmart. Walmart Global Tech.
- Shopify customer story, Anthropic; Shopify customer story, Google Cloud.
- MercadoLibre's AI Investments Are Set to Drive the Next Growth Chapter. Zacks.
- How Mercado Libre Scales Decision Making with AI. Amplitude.
- Despidos en MercadoLibre ponen de nuevo a la IA en el centro del debate laboral. Bloomberg Línea.
- Rakuten accelerates development with Claude Code. Anthropic customer story.
- AI coding agents: context, not intelligence. Wavect.
- DoorDash: Office Hours with Boris Cherny. Anthropic.
Enterprise platforms
- Anthropic and Salesforce Expand Strategic Partnership to Deliver Trusted AI for Regulated Industries. Salesforce, October 14, 2025.
- Accelerating SAP consulting projects with SAP Joule using Anthropic Claude models on Amazon Bedrock. AWS for SAP blog.
- Clark, Lindsay. SAP's AI strategy: Come for the openness, stay because you have to. The Register, May 19, 2026.
Services and engineering
- Accenture and Anthropic Launch Multi-Year Partnership. Accenture Newsroom, December 9, 2025.
- Accenture and Anthropic launch partnership. Anthropic, December 9, 2025.
- UST is bringing Claude to physical AI. Anthropic, July 2026.
- UST Partners With Anthropic, Bringing Claude to Its Platform. Channel Insider.
Humanitarian work
- Introducing Claude for Nonprofits. Anthropic, December 2025.
- Anthropic offers nonprofits discounts for Giving Tuesday. NBC News.
- The IRC turns fragmented health data into faster decisions. Anthropic customer story.
Small business and construction
- Borghi, Brianna. AI for restaurants: Boston restaurateur explains Colette platform. NBC Boston, June 17, 2026.
- How Shy Bird uses Claude. Anthropic customer story.
- Contractor AI Adoption Surges in 2026, Report Finds. Walls & Ceilings, March 30, 2026.
- New Survey Reveals Overwhelmingly Optimistic Results on the Use of AI in Construction. Construction Executive.
- Gribbins, Keith. AI in Construction: Everyone's Talking About It, But Almost No One's Using It. Construction Equipment, May 4, 2026.
Research and context
- Estrada, Sheryl. MIT report: 95% of generative AI pilots at companies are failing. Fortune, August 18, 2025.
- Becker, Joel, Nate Rush, Elizabeth Barnes and David Rein. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089, July 2025; METR summary.
- Anthropic Economic Index report: Economic primitives. Anthropic, January 2026.

No comments:
Post a Comment