Ready-made agents, Office integration, payment rails: what the second wave of agentic platforms means for the German Mittelstand
In the second week of May 2026 Anthropic released ten ready-to-run finance agent templates and Claude add-ins for Excel, PowerPoint and Word. AWS put Bedrock AgentCore Payments into preview — the first productive payment layer for autonomous agents. Microsoft Agent 365 became generally available with a cross-cloud registry. What Mittelstand stacks should take from this operationally and on the data-protection side.
The 90-second summary
The second week of May 2026 unlocked a second layer of the agent stack. Anthropic released ten ready-to-run finance agent templates and Claude add-ins for Excel, PowerPoint and Word, completing the move from horizontal toolkits to industry-specific building blocks. Amazon Web Services put Bedrock AgentCore Payments into preview — a payment layer where agents can pay autonomously in USDC via Coinbase x402 and Stripe Privy. In parallel, Microsoft Agent 365 became generally available with a cross-cloud registry to AWS and Google Cloud, and Anthropic secured 300 megawatts of additional capacity through the SpaceX Colossus 1 data centre. For you as a Mittelstand decision-maker, the question shifts: less “are we building an agent?” — more “which ready-made building blocks do we combine, where do we process which data and who pays at the end?”. Data-protection topics therefore move earlier into the procurement process — the legal assessment stays with your data-protection and legal functions, but procurement should make room for these topics before the contract is signed.
What is actually new this week
Three movements shape the past 48 hours, and all three concern the layers above model selection.
First, agents are being delivered industry- and tool-ready. Anthropic published ten finance agent templates on GitHub — from pitch builder via KYC screener to month-end closer — and made them runnable as plugins in Claude Cowork and Claude Code. There are also add-ins for Microsoft Excel, PowerPoint and Word, with Outlook coming soon. One agent builds the model in Excel, hands the context to PowerPoint and drops the cover note in Outlook — without the user having to re-explain the state between applications.
Second, the agent economy gets a real payment layer. Amazon Web Services put Bedrock AgentCore Payments into preview. Developers choose between a Coinbase wallet or a Stripe Privy wallet as the payment connection. Transactions run via Coinbase’s x402 protocol and settle in roughly 200 milliseconds in USDC on Base and Solana. The first use case is micropayments for APIs, data feeds, MCP servers and paywalled content — the next expansion targets bookings and merchant payments.
Third, the hyperscalers consolidate the governance layer. Microsoft Agent 365 has been generally available since 1 May and in the second week of May moved its cross-cloud registry with AWS Bedrock and Google Cloud into public preview. Local agent discovery via Defender and Intune is new, and Windows 365 for Agents is available as a US-only preview. In parallel, Anthropic contracted the entire SpaceX Colossus 1 data centre — over 220,000 GPUs and 300 megawatts of capacity flow into higher Claude limits within the month.
What it means for the Mittelstand
The consequence of these three movements for a Mittelstand stack is more concrete than the headlines suggest.
First, the make-or-buy decision shifts clearly toward buy. If you scheduled a KYC-screening or a month-end workflow project with two weeks of external engineering back in February, you should revisit that plan now. If a template like the KYC Screener or the Month End Closer already covers 80 percent of the workflow, the economically sensible path is forking, adapting and hardening — not rebuilding from scratch. The templates are visible on GitHub, which lowers the lock-in risk and makes code review possible during procurement.
Second, you need to prepare for Office integration as a new interface. When Claude builds models directly in Excel and runs sensitivity analyses, part of the manual spreadsheet workflow moves into the agent. In the Mittelstand this is the first use case with an immediate effect on hours — for controllers, accounting and sales steering. At the same time, business logic that was previously hidden in Excel moves into the model context — the question “which sheet is the source of truth?” will get louder over the next quarters.
Third, you will need a position on agent payments sooner than expected. The idea that an agent autonomously pays for an API call or a data feed sounds exotic from a German accounting perspective. Over the next twelve months it will land in every serious MCP marketplace. Mittelstand organisations should clarify two things: a spending limit per agent and per use case, and a booking path that channels USDC micropayments into the main ledger in an auditable way. This is not a theoretical point — it is concrete preparation for a procurement workflow that did not previously exist.
Fourth, data protection moves earlier into the procurement process. As soon as Anthropic add-ins run directly in Excel and Outlook, prompt content may flow out of your house — and with it, potentially, personal data from payroll, customer records and HR. We see three topics with clients that get raised early in preparation — and they belong with your data-protection and legal functions, not in a technical recommendation. First, the contractual classification of every vendor that processes data: several parties are in play, from Microsoft via Anthropic to whichever MCP server operator you use, and how each one is positioned contractually is a standard test for data-protection officers. Second, the international data flow: the Anthropic-SpaceX capacity sits in the US, and the cross-cloud extension of Microsoft Agent 365 shifts where inference actually happens — how that constellation is held contractually is a question for your legal function. Third, the EU AI Act, whose rules are phasing in across 2025 and 2026. Where your specific use case fits and which obligations follow is a legal assessment — we can only point out that this assessment should sensibly happen before roll-out, not after. From a procurement perspective that means: data-protection impact assessment and legal classification belong in the pilot plan as a work package, with clear ownership and a timeline.
Technical developments in detail
At the stack level, four building blocks are the most relevant moves of the week.
Anthropic Finance Agents on GitHub. Ten templates — Pitch Builder, Meeting Preparer, Earnings Reviewer, Model Builder, Market Researcher, KYC Screener, Valuation Reviewer, General Ledger Reconciler, Month End Closer, Statement Auditor — are visible as running code in the anthropics/financial-services repository. The architecture follows a lean pattern: each agent is a plugin with a clearly scoped tool set, a prompt definition and a test suite. Anyone reading the code can immediately tell which workflows are realistic and which run up against the current limits of model capability.
AWS Bedrock AgentCore Payments. The implementation rests on two open building blocks worth knowing. Coinbase x402 is an HTTP-native payment protocol — the server answers with HTTP 402 “Payment Required” and a machine-readable payment hint, the agent triggers payment and re-calls the endpoint. Stripe Privy attaches wallet capabilities to the Stripe infrastructure and makes USDC funding and withdrawals via debit card practical. For the architecture this means: you can equip MCP servers with paid endpoints without locking the agent into a proprietary marketplace.
Microsoft Agent 365 with cross-cloud registry. The public preview synchronises agents between Microsoft, AWS Bedrock and Google Cloud in a shared registry. Defender discovers local agents, Intune manages them, and Windows 365 for Agents provides a managed desktop context for autonomous workloads. Mittelstand organisations already living in the Microsoft 365 world get the control and identity bracket that was missing in April. Architecturally important: cross-cloud calls into AWS Bedrock or Google Cloud Vertex may land in a different region than the original Microsoft 365 data processing, depending on routing. What consequences that has contractually is a question for your data-protection and legal functions — on the technical side, we recommend logging the actual inference routing per agent so that this assessment is even possible.
MCP roadmap progress. In the second week of May, the official Kotlin, C# and TypeScript SDKs were updated. The stateless HTTP transport variant is in review, as is experimental skills discovery via MCP primitives. For productive setups, this means the gaps around audit trails, SSO integration and configuration portability will be addressed in the next releases — and horizontal scaling behind standard load balancers without persistent sessions moves into architectural reach.
On the compute side, the Anthropic-SpaceX agreement noticeably changes availability for Claude users in the DACH region. The five-hour limits for paid plans have been doubled, and peak-time throttling for Pro and Max is gone. Anyone running agents productively on Claude should measure the effect in May — it is real and plannable.
What we actually observe
Across our TYPO3 and DevSecOps platforms, we see that this week’s moves shift two operational questions. The first is “which templates fit which workflow?” — we recommend reviewing the Anthropic repositories before any custom build, and only building yourself where the templates do not fit. The second is “how do we allocate agent spend correctly to cost centres?” — once payments and token consumption come together, the accounting workflow needs a clear allocation logic, ideally before the first productive run.
Frequently asked questions about the second wave of agentic platforms
Are the Anthropic finance templates really usable for the German Mittelstand?+
In their current form they are tailored to typical financial-industry workflows — banking, asset management and audit. The logic behind them is generic enough, however, for manufacturing Mittelstand companies to use the Month End Closer or the Statement Auditor as a template for their own booking workflows. The KYC Screener is unlikely to go directly into production without a financial-supervision angle, but the pattern ‘structured document check against a rule set’ is transferable to supplier vetting or GDPR reviews.
How does the Microsoft 365 integration from Anthropic differ from Copilot?+
Copilot is Microsoft’s own agent layer inside the Office suite. The Claude add-ins sit alongside it — so you can run an Anthropic agent in Excel without it going through the Copilot sandbox. For the Mittelstand this means more choice on the model side and, at the same time, an obligation to take a position: which agent runs in which app, and which one takes precedence in which workflow. If you operate both in parallel, you should be able to distinguish that in your permissions model and in the audit trail.
How risky is it to let agents pay autonomously in USDC?+
Technically the setup is manageable, because x402 and Stripe Privy provide limits, allowances and audit logs. From a regulatory point of view the situation in Germany is less clear-cut — stablecoin, anti-money-laundering and accounting topics come together and need to be aligned with your compliance, tax and legal advisors. From an operational view, you should set at least three technical guardrails before going productive: a hard spending cap per agent and use case, a wallet protection strategy (who holds the key and how is it rotated), and a booking and audit log that documents micropayments in an auditable way. Which regulatory obligations follow in your specific business model is not a question we answer in your place.
Which data-protection topics surface in procurement around the new agent building blocks?+
Based on our observations across active projects, three topic areas typically end up on the table — and all three belong with your data-protection and legal advisors. First, the contractual classification of every vendor that processes data — Microsoft, Anthropic, AWS, Google Cloud and every single MCP server operator. Second, the question of which region inference actually happens in, because cross-cloud routing and the Anthropic-SpaceX capacity in the US have shifted things since last year. Third, the classification of the specific use case under the EU AI Act, whose rules are phasing in across 2025 and 2026. We build the technical side so that these assessments are even possible: logging, inference region per call, clear permissions layers, documented data flows. The legal assessment itself is done by your legal and data-protection functions.
Can we put the Anthropic add-ins for Microsoft 365 into production without further checks?+
From a technical view yes — from a procurement view rarely without a data-protection check first. We recommend kicking off three things in parallel as soon as a productive pilot is in sight. First, bring the data-protection function and legal advisors into the loop early so contractual situation, data flows and use-case classification are assessed before anything is plugged in. Second, keep the data scope tight — models with anonymised test data are uncritical, an Outlook mailbox with customer communication has a different risk profile. Third, set up prompt and output logging in an audit sink under your own control, so that the data-protection assessment can work on real data. Which of these are strictly required and where there is flexibility belongs in your data-protection function’s assessment, not in our recommendation.
Is the Microsoft Agent 365 licence worthwhile for Mittelstand stacks?+
Agent 365 is part of the new Microsoft 365 E7 subscription. For pure Office use with occasional agents the jump rarely pays off — the licence targets organisations that manage ten or more agents in parallel. If you have fewer than five agents in production today, the leaner path is your own MCP layer with a central audit sink and an identity anchor in Entra ID. If you can foresee growth into double-digit agent counts, Agent 365 is the most plausible answer on the market.
How do we wire the new templates into our DevSecOps pipeline?+
Like any other open-source building block. Fork it, move it into an internal repository, run it in the CI pipeline build against your policy, and connect tool calls to your permissions layer. It is important that the template’s test suite does not run directly against productive data — build a sandbox data set you can reproduce regressions against for each Anthropic release. If you run the template as a plugin in Claude Code, also configure an audit sink outside the agent itself.
We sort with you which agent building block fits your real workflow.
If you want to pilot one of the new building blocks over the coming weeks, we will spend 30 minutes with you sorting which Anthropic template fits your real workflow best, which Office integration pays off and whether Bedrock AgentCore Payments is already a topic for your stack today. No pitch — a working session with a concrete outcome.
Conclusion
The second week of May 2026 lifted the agent discussion onto a new plane. The question is no longer whether a model can run an agent — that is settled. The question is which ready-made building blocks exist, in which application they run, who books the small amounts and which data-protection topics have been assessed by your data-protection and legal functions before roll-out. For Mittelstand stacks this means: less in-house building, more integration, an earlier position than expected on agent permissions, Office integration and payment workflows — and data-protection topics as a fixed column in the pilot plan. The next twelve months will be decided less by model releases and more by clean procurement, booking and data-governance logic.