Once a month, the corporate-card company Ramp publishes an index that tracks which AI vendors its 50,000-plus business customers are actually paying for. For two years, OpenAI sat on top by a comfortable margin. In the May 2026 release, that changed: 34.4% of paying businesses are now buying Anthropic, against 32.3% paying OpenAI. A year ago, Anthropic’s share was 9%. The shift matters because Ramp sees real invoices, not survey answers. It also lands in the same week Anthropic is reportedly raising up to $50 billion at a $950 billion valuation — ahead of OpenAI’s $852 billion. For a DAX40 buyer with a multi-year Microsoft Copilot or ChatGPT Enterprise contract, the question is no longer which lab is ahead. It is whether the lead is durable enough to bet the next renewal on.

Two months ago, in a memo that leaked out of OpenAI’s San Francisco headquarters, Fidji Simo — the former Instacart CEO who runs OpenAI’s applications business — told staff the company was now operating in what Sam Altman and research chief Mark Chen had internally labelled a “code red.” The trigger, she wrote, was Anthropic. Claude Code was eating into OpenAI’s developer base, and enterprise buyers were starting to ask uncomfortable questions about which lab they should standardise on. OpenAI needed, in her phrase, to put extra weight on productivity in general, and on enterprise productivity in particular. Inside the building, the message landed as an admission: the company that defined the consumer AI category was losing the part of the market that actually pays. On 13 May 2026, Ramp published the data point that turned that internal anxiety into a public scoreboard. Its monthly AI Index, drawn from corporate-card and invoice flows across more than 50,000 American businesses, showed Anthropic at 34.4% of paid business users and OpenAI at 32.3%. In April alone, Anthropic added 3.8 points; OpenAI shed 2.9. A year earlier, the gap ran the other way — OpenAI dominant, Anthropic in single digits. “Anthropic has already been in the lead amongst the high adoption groups like finance, tech, professional services,” Ramp economist Ara Kharazian told TechCrunch. “It’s across the other firms where OpenAI still has a lead, but that has been shrinking over the past couple of months.” The context made the moment heavier. The same week, Bloomberg and the New York Times reported that Anthropic was in talks to raise between $30 billion and $50 billion at a valuation of up to $950 billion, on the back of an annualised revenue run rate that has surged from roughly $9 billion at the close of 2025 to more than $30 billion today. Claude Code, the CLI-based coding agent that Allianz quietly rolled out to its workforce in January, is the fastest-growing product in the company’s history. If the round closes at the top of the range, Anthropic will, for the first time, be worth more than OpenAI’s $852 billion last mark. For Anthropic CEO Dario Amodei, this is the payoff for a deliberate strategy. Rather than chase Sam Altman into consumer chat, advertising and a Jony Ive-designed device, Anthropic spent eighteen months courting developers, regulated industries and integrators — the unglamorous middle of the enterprise stack. “What Anthropic did worked really well,” Kharazian said. “Start with a very technical customer base, focus on their needs, really succeed in execution and then start broadening out through tools like Cowork.” The Allianz deal in January, followed by AstraZeneca, Sanofi and Genmab signing for Claude in life-sciences workflows, gave the strategy proof points the average CIO recognises. Not by accident, those proof points sit precisely where DAX40 buyers — Munich Re, BASF, Bayer, Siemens — do their own evaluations.

Strip away the narrative and the index reads like a competitive blowout. Anthropic’s share of paid business users quadrupled in twelve months, from 9% in May 2025 to 34.4% in April 2026 — a 26-point swing. OpenAI’s share moved 0.3 points in the same window. To put that in perspective, AWS took roughly six years between 2010 and 2016 to build a comparable lead over its nearest cloud rival in business adoption; Anthropic compressed an equivalent re-rating into four quarters. On OpenRouter’s leaderboard, which samples a different population of model traffic, OpenAI last ranked above Anthropic in December 2025. Two independent panels, one signal. The revenue gap has narrowed in lockstep. OpenAI is reportedly running near a $20 billion annualised rate; Anthropic, by its own disclosure, has crossed $30 billion, with Claude Code representing the largest single contributor to growth. A January Menlo Ventures survey already pegged Anthropic at roughly 40% of the enterprise LLM market and over half of AI coding spend; the May Ramp data confirms that the trend has continued, not reverted. The funding round telegraphs the same conclusion. Google has committed up to $40 billion of additional capital, Amazon up to $25 billion, and an IPO is being discussed for as early as October 2026. But the same Ramp report carries its own counter-thesis, and Kharazian wrote it himself. The first risk is structural: Anthropic earns more when customers spend more tokens. That is not a neutral incentive. Uber’s CTO disclosed earlier this spring that the company burned through its entire 2026 AI budget in four months, much of it on Claude Code and Cursor, with individual engineers running monthly API bills of $500 to $2,000. Ramp economist Rafael Hajjar separately found that Anthropic’s most recent model update tripled token costs for any prompt containing an image — a pricing change that, in his reading, sits awkwardly with the company’s stated compute constraints. The second risk is operational. Anthropic spent much of April rate-limiting paying developers and quietly removing Claude Code access from some accounts. The Pragmatic Engineer’s Gergely Orosz argued in early May that the moves looked dev-hostile and likely masked a capacity shortage that an emergency SpaceX-Colossus deal — 220,000 NVIDIA GPUs, 300 megawatts — was rushed in to fix. The catch: enterprise procurement teams have long memories. Outages and silent quota cuts during a sales cycle are exactly what kills a multi-year framework agreement. The third risk is the one most relevant to a Großkonzern’s CFO. Some of the fastest-growing AI vendors on Ramp in April were not frontier labs at all. They were inference platforms — the Together AIs, Fireworks and Groqs of the world — selling cheap access to open-source Llama and DeepSeek derivatives. For routine classification, summarisation and customer-service workflows, good-enough AI at a fraction of frontier pricing is increasingly the rational FinOps choice. If that thesis holds, the leaderboard at the top is a fight over a shrinking premium tier.

For procurement leaders in Frankfurt, Munich and Walldorf, the May index is less a vendor recommendation than a portfolio signal. The Microsoft-OpenAI bundle that anchored most enterprise Copilot rollouts in 2024 and 2025 is no longer the unambiguous default. Allianz’s January Claude Code deployment to its global workforce, paired with the Anthropic-built compliance logging layer that records every model interaction for regulators, is the template German buyers are now studying. So is the Anthropic DACH leadership build-out under Roman Howe, who has been hiring out of Munich since the start of the year. Bayer Pharmaceuticals’ Claude deal in clinical-trial documentation, BASF’s Mistral-led production-chemistry copilot, and Munich Re’s newly seated Chief AI Officer point to the same conclusion: DAX40 vendor selection is becoming portfolio construction, not single-vendor commitment. The operational risks Kharazian flags translate directly into RfP language. Token-pricing volatility, image-prompt cost multipliers, capacity-driven rate limits and the absence of EU sovereign-cloud parity with Microsoft’s Azure footprint are all material at framework-contract scale. A Großkonzern signing a multi-year Anthropic deal in Q3 2026 should expect to negotiate hard on per-seat caps, committed throughput, regional data residency and — critically — a contractual exit ramp into open-source inference for non-frontier workloads. The strategic move is not to pick a winner. It is to build a routing layer, internally or via partners, that puts Claude on coding and high-value reasoning, ChatGPT on broad employee productivity, and a cheap inference platform on everything else. The Cohere–Aleph Alpha–STACKIT path supplies the sovereign branch of that routing tree. The May Ramp data is the cleanest evidence yet that single-vendor lock-in to either lab is now the riskier option, and procurement teams that internalise that lesson before the next renewal cycle will save material money.

PwC, one of the world’s four largest accounting and consulting firms, has expanded its alliance with Anthropic, the maker of the Claude family of AI models. Under the deal announced on May 14, 2026, PwC will train and certify 30,000 of its US professionals on Claude, roll out Anthropic’s Claude Code (a coding assistant) and Claude Cowork (an in-document assistant) across its workforce, and stand up a joint Center of Excellence. PwC is also launching a new finance practice, inside its Office of the CFO unit, built natively on Claude. The headline claim from both firms is that production engagements are delivering up to 70% faster outcomes. For European clients, the immediate question is whether their PwC engagement teams will, by default, ship Claude-shaped artefacts into their finance, deal and modernisation programmes.

Inside a glass-walled briefing room at PwC’s New York tower on Park Avenue, Paul Griggs, the firm’s US Senior Partner and chief executive, framed the announcement as a verdict, not an experiment. “The conversation around AI has shifted from possibility to execution,” he said, in remarks distributed alongside the press release. “Clients are looking for ways to apply AI that are secure, responsible, and capable of delivering measurable outcomes in complex business environments.” Standing in for the supplier side, Anthropic chief executive Dario Amodei offered the metric that will travel: “Insurance underwriting that took ten weeks now takes ten days. Security work that took hours now takes minutes. We’re excited to put Claude in the hands of hundreds of thousands of people across PwC’s workforce.” The expansion is structured around three pillars: agentic technology builds for clients, AI-native deal-making, and reinvention of the underlying enterprise function. Engineering teams inside PwC’s US firm will use Claude Code to ship production software in weeks rather than quarters. Deal teams — both corporate M&A and private-equity sponsor coverage — will use Claude agents through diligence, value creation and integration. And PwC’s finance, supply chain and HR transformation practices will work from Claude-native operating-model templates rather than slide-ware. The headline number is the workforce commitment. PwC will train and certify 30,000 US professionals, anchored by a joint Center of Excellence. The Claude Code and Claude Cowork rollout begins with US teams and expands toward PwC’s global network of more than 364,000 people across 136 countries — the largest single distribution event in Anthropic’s history. Cowork connects to client data through Anthropic’s open Model Context Protocol, the same plumbing PwC has already wired into ChatPwC, the firm’s internal assistant. The most consequential structural move is the new Claude-native finance business group inside PwC’s Office of the CFO practice. It is the first standalone PwC business unit anchored, by design, in a single AI vendor’s stack — Claude in productivity tools, Cowork in spreadsheets and decks, Claude Code for the engineering layer. Customer-zero proofs are already there: PwC has used Claude internally for journal entries, variance analysis and RFP drafting, and has helped Anthropic’s own CFO office stand up controls and international payroll. There is even a client showcase — Advocate Health, a 167,000-person US hospital system, where chief digital and AI officer Andy Crowder is positioning Claude as the foundation underneath the operating model rather than a productivity bolt-on. The pivot, for European observers, comes in the small print. PwC is positioning Claude not as one of several large language models on a procurement panel, but as the engine for its dealmaking franchise. M&A diligence — historically the most labour-intensive and highest-margin product the Big Four sell — is being explicitly re-tooled around Anthropic’s models. The strategic message to McKinsey, BCG, Bain, Deloitte, EY, KPMG and Accenture is unmistakable: match this depth of integration, or watch your dealmaking economics decay.

Read alongside the May 13 Capgemini–OpenAI announcement, this is a structural realignment of the consulting industry. Capgemini took an equity stake in OpenAI’s new Deployment Company (DeployCo), alongside Bain and McKinsey, with TPG as lead investor and Brookfield, Advent and Bain Capital as co-leads. Within seventy-two hours, PwC has chosen the opposite vendor and a different commercial shape — no equity, but the deepest workflow integration any consulting firm has yet committed to with Anthropic. The Big Three strategy houses are now publicly on the OpenAI side; the largest of the Big Four has planted a flag with Anthropic. The historical analogue is the SAP and Accenture pairing of the late 1990s, when Accenture’s willingness to retool its delivery methodology around R/3 produced two decades of compounding revenue for both firms — and locked an entire generation of finance organisations onto a single ERP backbone. The Anthropic–PwC deal echoes that pattern, with one important twist: where SAP sold software, Anthropic is selling an opinion about how white-collar work should be decomposed into agents. Every PwC deliverable shipped on Claude Code carries that opinion forward into the client’s operating model. For a German Großkonzern with a PwC engagement footprint — and that is most of the DAX40 outside the handful audited or principally advised by KPMG, EY or Deloitte — three operational consequences follow. First, finance transformation programmes already in flight will be re-scoped against the new Office-of-the-CFO playbook. The pitch will land as a productivity story; the substrate is a long-term Anthropic dependency at the controllership layer. Second, M&A diligence engagements will increasingly be Claude-mediated. CFOs and corporate-development heads should expect data-room workflows, red-flag synthesis and synergy modelling to be produced by Claude agents working alongside PwC dealmakers. Third, mainframe modernisation — a programme line item for almost every German bank and insurer still running COBOL — is now being sold with a specific Claude Code reference: a codebase four times the original scope, on time and under budget. The lock-in question is the one DAX40 procurement teams should ask first. PwC’s commercial structure — enterprise-wide Claude access, joint Center of Excellence, a finance practice built on a single vendor stack — is precisely the architecture that compounds switching costs over a four-to-six-year contract horizon. The audit-independence question follows closely behind. PwC has spent the last three years narrowing the consulting services it offers to its own audit clients, partly in response to SEC sanctions and IFRS-side scrutiny. A Claude-native finance practice that re-engineers controllership, journal entries and variance analysis sits uncomfortably close to the audit perimeter. European audit regulators — APAS in Germany, the FRC in the UK — will want to see the firewall design before they accept that ChatPwC and a Claude-native CFO practice can coexist with statutory audit work for the same group. There is also a quieter winner inside Anthropic’s own commercial strategy. The PwC alliance is the deepest single commitment within Anthropic’s $100 million Claude Partner Network — the programme it stood up earlier this year to back the services firms doing actual enterprise deployments rather than headline pilots. The bet is that the bottleneck on enterprise AI revenue is not model quality but distribution, and that owning the largest professional services workforce in the world is worth more than another point of MMLU.

Both firms lead with a 70% delivery-improvement number, drawn from a portfolio of live engagements rather than a controlled study. The supporting examples are specific enough to interrogate. Insurance underwriting cycles compressed from ten weeks to ten days — a 93% reduction — open lines of business that were not previously economically viable, particularly in specialty and small-commercial lines where underwriting cost per policy has historically gated the market. The mainframe modernisation reference — a COBOL codebase four times larger than scoped, tracking on time and under budget — is the most commercially material claim, because COBOL backlog is the single largest unresolved technical-debt item on the balance sheet of every European systemic bank. If Claude Code can credibly compress that programme by even a factor of two, the spend reallocation across DAX40 IT budgets is in the low single-digit billions of euros annually. The HR transformation reference — a stalled programme rebooted with a working prototype in one week and a full application in under two months — points to where the consulting margin pressure will land first: business-process redesign, where the deliverable has historically been a slide deck and a steering committee, not a working application. The cybersecurity claim — incident response cut from hours to minutes via agentic vulnerability operations — is the most defensible, because the time-to-contain metric is already instrumented inside most enterprise SOCs. The aggregate $2 trillion technical-debt figure both firms cite as the addressable market is necessarily soft; it is a directional anchor, not a forecast. What it does signal is the scale of the prize PwC believes it can capture by being the Anthropic-native distribution layer.

Cerebras Systems makes AI chips that look nothing like the GPUs from NVIDIA that dominate enterprise data centres. Instead of cutting a silicon wafer into hundreds of small chips, Cerebras keeps the whole wafer intact as one giant processor — the Wafer-Scale Engine. The pitch: fewer chips talking to each other means lower latency and faster answers when running already-trained AI models, a workload called inference. On 14 May 2026 the company listed on Nasdaq under the ticker CBRS, sold 30 million shares at $185 and watched the stock close 68 percent higher. It was the biggest US technology flotation since Uber in 2019, and the first signal that public markets are willing to fund AI compute beyond the incumbent.

Andrew Feldman stood under the spotlights at the Nasdaq MarketSite in Times Square, hand on the opening bell, with co-founder and chief technology officer Sean Lie a step behind him. By the time the bell stopped vibrating, the order book that bankers at Citi and Barclays had built overnight was reshaping into something nobody on the syndicate desks had quite expected. Shares priced at $185 the previous evening — already $25 above the top of a range that had itself been raised twice — opened at $350. Within minutes they touched $386. By the close, Cerebras was a $66 billion company on the listed share count and roughly $86 billion fully diluted, with an intraday peak nearer $95 billion. Feldman, who had spent the better part of two years explaining wafer-scale silicon to investors who had only ever bought NVIDIA, told CNBC the listing was simply a financing event: “We have tremendous opportunities for growth, and this was the right way to fund our growth.” The understatement was deliberate. Cerebras had filed its first S-1 in September 2024, only to pull it when the Committee on Foreign Investment in the United States opened a review into the 24 percent stake held by Abu Dhabi’s G42. The company spent eighteen months restructuring those shares into a non-voting class, refiling the prospectus, and rebuilding a customer book that had been almost entirely concentrated in a single Gulf buyer. The pricing on Wednesday night was the receipt for that work. The reception on Thursday was something else — a verdict on the broader question of whether public capital is willing to bet on AI compute beyond NVIDIA. The narrative pivot came three hours into trading. Salesforce desks that had been positioned for a tepid debut — the post-2022 IPO market has been stingy with software listings, and Snowflake’s 2020 record was the only meaningful comparable — began re-cutting their books in real time. The float, at roughly $5.5 billion, was small enough that a handful of long-only mutual funds could not get the allocations they wanted at the price; the chase upward did the rest. Cerebras finished the session at $311.07. Feldman’s personal stake was now worth about $3.2 billion on paper. Lie’s was worth $1.7 billion. Two new technology billionaires had been minted by the close. For Frankfurt and Munich risk committees that have spent twelve months trying to source NVIDIA H200 and B200 allocations through hyperscaler channels, the meaning of the day was narrower and more useful: the market had just validated a second listed pure-play AI compute supplier at a scale that makes procurement conversations possible.

Strip away the ticker-tape and the financials are more interesting than the share count suggests. Cerebras booked $510 million of revenue in fiscal 2025, up 76 percent year on year. Gross profit was $199 million, a margin of roughly 39 percent — respectable for a hardware company, embarrassing next to NVIDIA’s data-centre gross margin, which has run above 75 percent through the H100 and H200 cycle. The headline GAAP net profit of $238 million flatters the underlying business: $363 million of that came from a one-time non-cash accounting adjustment tied to the equity restructuring. Strip it out and add back stock-based compensation, and the company ran a non-GAAP net loss of about $76 million for the year. At the IPO price the fully diluted enterprise value was $56.4 billion, putting the price-to-sales ratio at roughly 110x. After the first-day pop, the multiple ran closer to 185x. NVIDIA, by comparison, trades at single-digit forward sales on a $215 billion revenue base. That premium is not a verdict on the present — it is a verdict on the inference market in 2027 and 2028. Cerebras’ customer book is the other place where the spreadsheet matters. G42 and the Mohamed bin Zayed University of Artificial Intelligence — treated as a single related-party group in the prospectus — accounted for roughly 86 percent of 2025 revenue between them. That is down from the 87 percent G42 alone represented in the first half of 2024, but it is not the diversified book that growth-equity buyers usually pay 100x sales for. The countervailing data points are the OpenAI inference deal disclosed at filing (publicly described as worth up to $10 billion in capacity over the contract term), an AWS partnership announced in March 2026 to bring WSE-3 systems into AWS Marketplace, and the build-out of six new Cerebras-operated inference data centres across North America and Europe. Those contracts are how the customer-concentration ratio narrows from here. Wafer-scale economics are the structural story underneath the ratios. A single WSE-3 packs roughly 4 trillion transistors, 900,000 cores, 44 gigabytes of on-chip SRAM and 21 petabytes per second of memory bandwidth onto one piece of silicon roughly the size of a dinner plate. Cerebras’ own benchmarks have its CS-3 system delivering more than 2,500 tokens per second per user on Llama 4 Maverick, against roughly 1,000 on NVIDIA’s DGX B200, 794 on SambaNova, and 549 on Groq. Whether that latency advantage translates into a procurement budget in Walldorf, Wolfsburg or Erlangen depends entirely on how much of the inference workload sits in the long tail of latency-sensitive use cases — trading, real-time agent loops, factory-floor vision — versus the bulk batch inference that GPUs handle competently and cheaply.

On 13 May 2026, Anthropic released Claude for Small Business, a packaged setup that puts its Claude Cowork agent inside the everyday tools a corner shop or a fifty-person agency already uses: QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace and Microsoft 365. It ships with fifteen ready-to-run workflows — payroll planning, month-end close, invoice chasing, marketing campaigns, contract review — and fifteen reusable Skills. Claude does the work and asks the owner to approve before anything is sent, posted or paid. There is no extra charge beyond a Claude licence. Anthropic is taking it on a ten-city US tour starting in Chicago, paired with a free training course built with PayPal. The point is plain: instead of buying separate software for each function, a small business can hand the back office to one agent that already speaks to all of them.

In a Brooklyn butcher shop on Prospect Avenue, the owner of Prospect Butcher Co. was filmed for Anthropic’s launch reel running payroll while the slicer hummed in the background. He did not open QuickBooks. He did not open PayPal. He typed an instruction into Claude Cowork, watched the agent settle his cash position against incoming PayPal settlements, build a thirty-day forecast, rank what was overdue, and queue the reminder emails. Then he tapped approve. That tableau — a back office collapsed into one approval gate — is the thesis Anthropic shipped to market on 13 May, when President Daniela Amodei announced Claude for Small Business as a toggle inside Claude Cowork. The bundle includes fifteen agentic workflows and fifteen task-specific Skills, plus first-party connectors to Intuit QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, Microsoft 365 and Slack. The workflow catalogue reads like the to-do list of any owner-operator: payroll planning, month-end close, business-pulse dashboards, campaign management, invoice chasing, margin analysis, lead triage, contract review, tax-season prep, employee onboarding. Approval is mandatory before Claude sends an email, posts a campaign, or moves money. Existing system permissions carry over — if an employee cannot see a ledger in QuickBooks today, Claude cannot show it to them tomorrow. Anthropic does not train on customer data on Team and Enterprise plans. The go-to-market is unusually physical for a frontier lab. Starting 14 May in Chicago, Anthropic is running a ten-city tour through Tulsa, Dallas, Hamilton Township, Baton Rouge, Birmingham, Salt Lake City, Baltimore, San Jose and Indianapolis — a half-day live AI fluency workshop for a hundred local owners per stop, run with Tenex.co. Attendees walk out with a free month of Claude Max. A free on-demand course called AI Fluency for Small Business, co-produced with PayPal, launched the same day. Lina Ochman, Anthropic’s head of SMB, told Axios the company built the workflows after surveying owners about which tasks slow them down most. Her framing was pointed: “Historically, the software industry has been built for enterprises or VC-backed startups and consumers — but not for the fifteen-person HVAC company or the thirty-person landscaper.” For anyone watching the public software tape, the timing was the message. Salesforce had entered the day down roughly 31 percent year to date; ServiceNow had shed about 36 to 40 percent through April; Adobe, Intuit, DocuSign and Box were all bruised. The fear has a name on the sell side — the SaaS-pocalypse — and a thesis: if an agent can do the work that used to require a per-seat licence, the seat itself becomes optional. Piper Sandler had cut its CRM target in late April; Dario Amodei had told audiences earlier in the year that software vendors who do not adapt could completely go bust. Anthropic just shipped the product that makes the warning concrete.

The interesting part of Claude for Small Business is not the workflow catalogue. It is the shape of the thing. For two decades, the dominant abstraction in business software has been the application — a self-contained environment with its own UI, its own data model, its own per-seat pricing. Salesforce sold the seat. ServiceNow sold the seat. HubSpot sold the seat. The unbundling of the 2000s, when Salesforce broke Siebel apart and shipped CRM in a browser, simply replaced one set of seats with another. What Anthropic is shipping has a different topology. The user does not log into an app; the user states an intent. A persistent agent — Claude Cowork — sits above the existing tool graph, reaches into QuickBooks for the ledger, into PayPal for the settlement, into HubSpot for the pipeline, into Canva for the creative, into DocuSign for the signature, and stitches the result into a single approval surface. The Skills are reusable units of judgement (close the month, chase this invoice, draft this campaign) that the agent composes on demand. The connectors are the muscle. Cowork is the spinal cord. The user is the gatekeeper. This is the first time a frontier lab has packaged that topology as a turnkey product for the segment that drives 44 percent of US GDP and employs almost half the private-sector workforce. The historical analogue is not Salesforce-versus-Siebel; it is closer to Intuit’s 1990s desktop dominance, when QuickBooks and TurboTax collapsed accounting and tax workflows that had previously required a bookkeeper or a CPA. Intuit then rode the SaaS wave and ended up owning the SMB financial stack. Anthropic is now reaching through Intuit’s API and offering to drive the car. The pivot is sharp: the moat is no longer the application, it is the agent that can use every application. Microsoft and Google are running the same play through a different door. Copilot lives inside Teams, Outlook and Excel; Gemini lives inside Workspace. Both are betting that the agent rides on top of the productivity surface their customers already own. Anthropic has chosen to ride on top of the back-office stack instead — the financial, sales, and document tools where the actual work piles up after hours. For a fifty-person business that pays for QuickBooks Online, HubSpot Starter, Canva Pro and DocuSign anyway, the marginal cost of adding Claude is one subscription. The marginal benefit is not having to swivel-chair between five tabs at 11 pm. The revenue context sharpens the thesis. Anthropic’s run rate has crossed 30 billion dollars annualised; it is preparing the ground for an IPO; and Daniela Amodei is positioning Claude for Small Business as a public-benefit story rather than a margin grab — there is no extra fee beyond the Claude licence, and the company is co-funding Community Development Financial Institutions and a Workday-Foundation solopreneur cohort. That generosity is also a wedge. Once the back office runs through Claude, the cost of switching agents is roughly the cost of rewriting every Skill and reconnecting every connector. Lock-in does not vanish under the agent paradigm; it migrates.

For a Großkonzern advisor based in Frankfurt or Munich, Claude for Small Business is a US consumer-of-IT story with two enterprise consequences. First, it is a live demonstration of what an agentic back office looks like at scale, in production, with approval gates that an auditor can actually inspect. The patterns — Skills as reusable judgement, connectors as muscle, Cowork as the orchestrator, human-in-the-loop on every send-post-pay action — translate directly upward into Konzern shared-services centres, where invoice chasing, month-end close and contract review consume armies of full-time equivalents. The fifteen workflows Anthropic shipped for the corner shop are recognisable as the same workflows a DAX40 finance organisation runs at a thousand times the volume. Second, it forces a question German Mittelstand customers will start asking out loud: where is the European equivalent? The pieces are arriving. The Cohere–Aleph Alpha merger, announced 24 April with a 600 million dollar Schwarz commitment, is explicitly pitched as a sovereign frontier stack with European-language tokenisers and Mittelstand relationships, and SAP’s Christian Klein has said Joule and Business AI will deepen integration with both. SAP itself unveiled the Autonomous Enterprise at Sapphire in early May, with Joule Studio positioned as the no-code/pro-code surface for building agentic workflows on SAP-managed infrastructure. STACKIT supplies the sovereign-residency layer that European public-sector and regulated-industry customers will demand. The bundle a Mittelstand owner could realistically buy in 2026 — Joule on SAP Business One or S/4HANA Cloud, Cohere/Aleph Alpha models hosted in STACKIT, DATEV for the books — does not yet exist as a single toggle. Anthropic just published the reference architecture. The consultancy implication is concrete. The next two quarters will sort DACH AI conversations into two piles. The first pile asks how to deploy Claude or Copilot inside an existing Microsoft 365 or Google estate, accept the data-residency trade-offs, and ship workflows now. The second asks how to assemble a sovereign equivalent — SAP plus Cohere plus STACKIT plus a German systems integrator — and accept a six-to-twelve-month delay for the assurance that customer data, model weights and inference traffic stay inside the European perimeter. The right answer for a DAX40 client will rarely be one or the other; it will be a portfolio decision keyed to data class. Procurement, legal and works-council readiness will matter more than model benchmarks.