Between May 11 and 14, 2026, three things happened in four days.
On May 11, OpenAI announced The Deployment Company — a standalone subsidiary backed by over $4 billion from 19 investors, acquiring a 150-person AI consulting firm to embed engineers directly inside client organizations. On May 12, Anthropic released 12 legal practice-area Claude plugins and announced integrations with more than 20 legal tech companies including Thomson Reuters and Harvey. On May 13, Anthropic launched Claude for Small Business — 15 pre-built workflows connecting directly to QuickBooks, PayPal, HubSpot, and other tools small businesses use daily, at no extra cost, along with free training workshops across 10 US cities.
Taken individually, none of these is surprising — AI companies targeting enterprise, shipping industry solutions, sounds like normal business expansion. But pile them together, add the trajectory of Anthropic’s product evolution over the past two to three years, and a larger pattern emerges: AI products are systematically changing what they sell. Not “a model that writes code,” but “an AI that understands how your industry operates.”
Put differently: the unit of reuse is shifting.
The history of software is, in large part, a history of the unit of reuse getting progressively larger. The earliest programmers wrote machine code — zero reuse. Then came assembly, high-level languages, function libraries, frameworks, APIs. Each step packaged a larger chunk of functionality into a unit that the layer above could use directly. By the SaaS era, AWS turned server operations into API calls, Shopify turned running an online store into a subscription — the reusable unit had become an entire business capability, with all underlying technical detail abstracted away.
AI is undergoing a similar shift, except this time what’s being packaged isn’t servers. It’s cognitive labor.
When Anthropic first launched Claude in 2023, it sold the same thing every model company sold: tokens. You bought “intelligence” — what you did with it was your problem. Claude Code, released in May 2025, marked the first inflection point. It was no longer a model waiting to be called. It was an agent that could read a codebase, execute commands, and manage git on its own. The reusable unit shifted from “understanding a piece of text” to “completing a programming task.” By the end of 2025, Claude Code had surpassed $2.5 billion in annualized revenue, accounting for 20% of Anthropic’s total.
But Claude Code’s success exposed a deeper problem: the agent loop clearly worked, but only developers could use it. In January 2026, Anthropic released Cowork — the same agent architecture, but with a desktop interface replacing the terminal, and Office documents, browsers, and local apps replacing code files as the objects of action. The positioning, in Anthropic’s own words: “Claude Code for general computing.”
And then came the concentrated volley of May.
Claude for Financial Services (first launched July 2025, massively expanded May 2026) provides 10 pre-built financial agent templates — from pitch book creation and earnings analysis to month-end close and KYC screening. Each template bundles three things: domain-specific instructions (skills), pipes to industry data sources (connectors), and sub-agents for task decomposition. Data connections span FactSet, S&P Capital IQ, Moody’s, LSEG, and other industry standards.
Claude for Legal does the same thing, but for lawyers. Twelve practice-area plugins — commercial counsel, litigation associate, privacy counsel, law student — each wired into the professional tools that domain uses: Thomson Reuters’ Westlaw, iManage, NetDocuments, Harvey.
Claude for Small Business performs the same operation a third time, but for the owner of a 15-person HVAC company or a 30-person landscaping business. Workflows cover payroll, month-end close, invoice chasing, sales campaigns — each one an actual chunk of a small business owner’s life. The tools aren’t FactSet and Westlaw. They’re QuickBooks, PayPal, and HubSpot.
Three products across different industries, different scales, different complexity levels, built on exactly the same architecture. These aren’t three isolated launches. They’re one product strategy unfolding in three directions.
Spread the three launches out side by side, and three shared assumptions come into view.
First, the same agent loop can be reused across domains. Claude Code validated a pattern in programming: give an agent a goal, a set of tools, and an execution loop, and it can complete multi-step tasks. This pattern doesn’t depend on programming — it depends on the general capability of multi-step reasoning plus tool use. Drop it into financial data sources and compliance frameworks, and it becomes a financial agent. Drop it into legal databases and document systems, and it becomes a legal agent.
Second, industry know-how can be encoded into reusable packaging. What skills and plugin systems do, at their core, is write a domain’s operating procedures as machine-executable instructions. This has happened many times in software history — Salesforce standardized sales management as SaaS, QuickBooks standardized bookkeeping as software. The difference is that past standardization structured human operations; current standardization defines AI behavior. The logic is the same: turn tacit knowledge into explicit configuration.
Third, distribution cost must approach zero for this to make sense. Claude for Small Business carries no additional charge — it’s included in existing subscriptions. If Anthropic needed to station an engineer at every HVAC company, it would be indistinguishable from OpenAI’s DeployCo — both just labor-intensive consulting. Anthropic’s bet is that 15 pre-built workflows plus standardized tool connectors can cover most SMB needs without human intervention. Whether that bet holds depends on two variables: how standardized SMB data environments actually are, and whether business owners are willing to do the initial setup themselves.
If all three assumptions hold, Anthropic is building a “cognitive labor SaaS layer” — one agent architecture at the bottom, pre-built industry workflows on top. Marginal cost approaching zero, switching costs rising with depth of use. This is a high-margin business model, fundamentally different from OpenAI’s human-embedding approach.
OpenAI’s DeployCo heads in exactly the opposite direction. Its core assumption: the bottleneck for enterprise AI isn’t model capability, it’s deployment capacity. The solution: buy a 150-engineer AI consulting firm (Tomoro), raise $4 billion to acquire more, and station engineers inside client organizations. Who are the clients? The portfolio companies of the 19 investment firms in the consortium — over 2,000 businesses held by PE giants like TPG, Bain Capital, and Brookfield.
This is distribution via capital structure: rather than convincing companies one by one to adopt AI, get PE investors to mandate standardized deployment across their portfolios. The price tag is a 17.5% annualized guaranteed return — according to reporting by TNW and Yahoo Finance, a structure that looks less like venture capital and more like a high-yield credit instrument wearing a tech company’s clothes. If deployment cycles stretch out, OpenAI foots the return gap itself.
Google’s strategy leans closer to “winning by default.” It doesn’t need a consulting arm or embedded engineers, because its products already cover most SMBs’ daily tools — Gmail, Docs, Sheets, Meet, Merchant Center. What it’s doing is injecting Gemini’s capabilities into those products so users get AI features without changing their tools. Free training programs (SMB Learning Path, AI Boost Bootcamp, National Small Business Week AI workshops with the SBA) lower the adoption threshold further.
The divergence across these three paths points to the same core judgment: the next phase of AI commercialization isn’t a model capability race. It’s a distribution model competition. The model itself is becoming substitutable infrastructure, like a power grid everyone can plug into. The real question is who can deliver electricity to every household and every factory — and get them to use it.
Anthropic’s answer: standardized products — low marginal cost, but requiring some self-learning from customers. OpenAI’s answer: heavy human services — deep integration, but hard to scale. Google’s answer: ecosystem enhancement — widest reach, but shallowest depth.
“Reuse unit shifting from code to industry know-how” works as an analytical frame, but has blind spots as a prediction.
The most fundamental question: if model capability continues improving at current rates, how much residual value do pre-packaged industry workflows retain? Could a sufficiently intelligent general model, given “help me close the books this month,” automatically figure out QuickBooks’ data structure, accounting standards, and a company’s internal conventions, without Anthropic pre-writing the workflow? If the answer is yes, the value of these industry solutions gets progressively diluted by advancing model capability.
For now, this risk looks smaller than it appears. Not because models aren’t smart enough, but because the barriers in real business environments go beyond “knowing how to do something.” They include permission structures (who can touch what data), compliance frameworks (what operations need audit trails), and data pipes (how to accurately pull a specific account from QuickBooks). These are constraints of the physical and institutional world, not intelligence problems. The smarter the model, the more it needs an “already-authorized operating environment” — and that environment is precisely what industry solutions are constructing.
A separate risk is the SMB market itself. Historically, nearly every company that set out to “educate small businesses on new technology” hit a wall. SMB owners don’t follow a “best technology” decision logic — they follow a “least bad” decision logic. They have too little time, too little error tolerance, and have been burned too many times by software vendors’ promises. Anthropic’s 10-city tour (100 people per stop, 1,000 people total), set against America’s 33 million small businesses, looks more like market validation than scalable customer acquisition.
The deepest uncertainty: SMB owners are doing mental accounting between “AI saves me 3 hours” and “AI gets to look at my bank account.” These two values operate on completely different pricing logics, and they don’t touch.
Anthropic’s dense four-day release window in May 2026 is less a product launch milestone than a product strategy developing into visibility — three years of accumulation coming into focus at a single moment. From Claude API to Claude Code to Cowork to three vertical solutions, each step compounds the same logic: find a cognitive task an agent can execute, package it as a product that doesn’t require the user to understand the underlying model.
If this judgment holds, the next directions might include: more vertical industries (healthcare is already underway), finer-grained functional plugins (not just “legal” but “M&A lawyer,” “data privacy counsel”), and a push from “pre-built workflows” toward “self-adapting industry agents.”
For anyone watching the AI industry, rather than tracking which model scores what on which benchmark, look at this instead: who is turning AI from “what you can do with it” into “what it will proactively do for you” — and that “what” isn’t code, but the month-end close report your industry needs to file tomorrow. If this judgment is right, the metric for Anthropic isn’t model leaderboards. It’s how many HVAC company owners named Zhou went home at 7 PM on the last day of the month.
This article draws on Anthropic official announcements, reporting from Axios/The Decoder/Law.com, Goldman Sachs 10KSB survey data, US Chamber of Commerce SMB data, Reddit community discussions, and concurrent launches from OpenAI and Google. Key facts were cross-verified against primary sources.