🌟Vasilij’s note
This week two things happened that I think will define the next 18 months of AI in professional services. KPMG announced a strategic alliance to deploy Claude across its entire global workforce of more than 276,000 people. PwC followed days later, deploying Claude to build technology, execute deals, and reinvent enterprise functions for clients. These are not pilot programmes. These are operational decisions. The Big Four are no longer evaluating AI - they are deploying it at scale, with full commercial backing, across the same client industries that mid-market consultancies serve. At the same time, I recorded and published the complete guide to Claude for Small Business - Anthropic's new plugin that puts 31 workflows inside the tools smaller firms already use. The reaction was predictable: people noticed the features. What matters more is the question underneath: before you connect a single business tool to an AI agent, do you know what your data privacy settings actually are? That is the question most firms still haven't asked. I tested it live, on camera, with real business accounts. The answer has implications for every plan tier below Team. Read the maker note and watch the video before you connect anything.
In today's edition
This week in agents | What changed
KPMG deploys Claude to 276,000 staff in strategic global alliance
Announced 19 May, this is the largest single professional services deployment of Claude to date, covering audit, tax, advisory, and consulting workflows globally. → The Big Four are no longer running pilots. For mid-market consultancies competing for the same clients: the question is no longer whether to deploy AI, it's whether you can do it faster and more specifically than a firm with 276,000 trained operators.
Anthropic acquires Stainless
The SDK automation startup whose platform was used by OpenAI, Google, Cloudflare, and dozens of AI companies to auto-generate SDKs and developer tooling. Announced 18 May. → Anthropic is building infrastructure at the developer layer, not just the product layer. This accelerates how quickly third-party integrations reach Claude's ecosystem - which directly affects the connector and plugin market consultancies are building on.
OpenAI's ChatGPT Personal Finance launches for Pro users in the US
Announced 18 May, the feature lets users connect financial accounts, view spending dashboards, and ask questions grounded in their own financial data. → The race to embed AI agents inside professional workflows - accounting, finance, legal - is accelerating beyond the enterprise tier. Small business and personal finance use cases are being contested by frontier labs directly. Consultancies with financial services clients should track this closely.
Top moves | Signal → impact
KPMG and PwC deploy Claude at enterprise scale within one week of each other
KPMG's strategic alliance announced 19 May covers its global workforce of 276,000+ across audit, tax, and advisory. PwC announced 14 May it is deploying Claude to build technology, execute deals, and reinvent enterprise functions for clients. Both firms are positioning Claude as a delivery infrastructure layer, not a productivity add-on. → For mid-market consultancies: the competitive pressure is now structural. Your clients will ask whether your AI capability matches what their Big Four advisors are already running. The answer needs to be yes, and it needs to be specific - vertical workflow depth, domain knowledge, speed of deployment. Generic AI access is no longer differentiation.
Anthropic forms $200 million partnership with the Gates Foundation
Announced 14 May, the partnership focuses on applying AI to global health and development challenges - malaria, tuberculosis, maternal health, agricultural productivity in low-income countries. → Not directly operational for most consultancies, but strategically significant: Anthropic is now embedded in regulated, high-stakes, outcome-measured environments far beyond enterprise productivity. The governance and safety patterns required in these deployments will shape how Claude's compliance and audit capabilities develop.
Microsoft is actively shopping for an OpenAI replacement
Reported 15 May, Microsoft is looking to acquire Inception - a company that builds diffusion-based language models - as the amended non-exclusive licence with OpenAI came into force. Microsoft retains its 27% stake through 2032 but is now hedging its model dependency aggressively. → The infrastructure moat is continuing to collapse. For consultancies advising clients on AI vendor strategy: lock-in risk fell materially when the exclusive licence ended in April, but Microsoft's hedging signals that even well-resourced firms don't trust single-vendor dependencies. Build model-agnostic delivery frameworks now.
Upskilling spotlight | Learn this week
How Claude Code Works in Large Codebases: Best Practices and Where to Start (Anthropic)
Published 18 May via the Anthropic engineering blog. Covers production deployment patterns for large codebases - monorepos with millions of lines, legacy systems, microservices across separate repositories. Even if your firm isn't running Claude Code at that scale yet, the rollout sequencing and adoption patterns map directly to any serious Claude deployment. Practical and directly applicable to consultancies managing complex client environments.
Google Cloud Agent Orchestration Workshop (Google / AWS, 26 May)
A live hands-on workshop covering multi-agent systems that fail without proper state sharing, coordinated approvals, and failure recovery. Builds the orchestration layer using AWS Step Functions, Amazon Bedrock Agents, and Apache Airflow - with demos of retry logic, human approval gates, and graceful failure handling. Directly applicable to consultancies moving agents from pilot to production. Registration open now.
Maker note | What I built this week
This week I recorded the complete guide to Claude for Small Business - Anthropic's new plugin that puts 31 pre-built workflows (payroll planning, cash flow snapshots, invoice chasing, month-end close, contract review) inside Claude Cowork, connected to QuickBooks, PayPal, HubSpot, Stripe, and nine other tools.
The demo works. The payroll planning skill pulls accounts receivable, cash timing, and transaction history from QuickBooks and PayPal and produces a 90-day cash forecast in minutes. What used to take three spreadsheets and a Sunday evening now takes a five-minute review.
Decision: Before you connect a single business tool, check your plan and your privacy settings. On Team and Enterprise plans, your data is not used for model training — that is covered in commercial terms. On Pro and Max plans, the default allows Anthropic to use conversations to improve models, with retention up to five years if the training toggle is on. The setting is called "Help improve Claude." Go to Settings → Privacy → Help improve Claude → toggle it off before connecting any business account. The video covers this in full, including a live walkthrough of every step.
Operator’s picks | Tools to try
Claude for Small Business Plugin (Anthropic)
Use for: end-to-end business operations workflows - payroll planning, cash flow forecasting, invoice chasing, month-end close, contract review, lead triage. Connects to QuickBooks, PayPal, HubSpot, Stripe, Canva, and others.
Install via Claude Cowork → Customize → Browse Plugins → Anthropic and Partners.
Standout: 31 pre-built skills accessible via slash commands (/plan-payroll, /cash-flow-snapshot, /invoice-chase). Customisable per-skill to match your firm's specific standards and templates.
Caveat: Pro and Max plan users must disable "Help improve Claude" in privacy settings before connecting business accounts. Team plan recommended for any deployment involving client or financial data.
Cursor Composer 2.5
Use for: coding agent work requiring multi-step reasoning, complex codebase navigation, and long-horizon task execution. Trained with targeted reinforcement learning and synthetic data - meaningful capability step up from prior versions.
Standout: new distributed training techniques improve reliability on ambiguous or multi-constraint tasks, which mirrors the kind of requirements consultancies encounter in real client work.
Pair with: Claude Code for large legacy codebase analysis alongside Cursor for active development work.
Raindrop Workshop (GitHub)
Use for: giving Claude Code the ability to read traces, write evals against codebases, and fix what's broken. Provides livestreamed traces, a self-healing eval loop, and local replay across TypeScript, Python, Go, and Rust.
Standout: closes the gap between Claude Code writing code and Claude Code understanding whether the code actually works in production.
Deep dive | Skills and Plugins: The Consulting Ops Stack
Within five days in May 2026, two of the world's four largest professional services firms announced firm-wide Claude deployments at a combined scale of roughly half a million professionals. This is not a product story. It is a structural shift in what "AI-enabled consulting" means - and it changes the competitive calculus for every mid-market firm serving overlapping client industries.
On paper
KPMG's strategic alliance with Anthropic, announced 19 May, covers its global workforce of 276,000+ people across audit, tax, advisory, and consulting. The deployment is positioned as infrastructure - not a productivity tool for individuals, but a capability layer across delivery teams.
PwC's announcement, five days earlier on 14 May, goes further in articulating intent: the firm is deploying Claude to build technology, execute deals, and reinvent enterprise functions for clients. This is client-facing, not just internal.
Both deployments follow Accenture's 30,000-person Claude Centre of Excellence and Cognizant's 350,000-associate rollout announced earlier this year. The pattern is now consistent: the largest professional services organisations in the world are deploying Claude as a delivery infrastructure decision.
Anthropic's commercial trajectory reinforces this: the Gates Foundation partnership ($200 million, 14 May), the Stainless acquisition (18 May), and higher usage limits announced on 6 May all point toward Anthropic building the infrastructure and channel coverage required to sustain enterprise-scale deployments.
In practice
The gap between "deployed to staff" and "delivering measurable value to clients" remains significant. Enterprise deployments at this scale typically face 2–4 weeks of adoption friction per team, 10–20% edge case rates requiring manual intervention, and 42% deltas between expected and actual AI usage (Battery Ventures research). KPMG and PwC are not immune to these patterns.
What they have that most mid-market firms don't is dedicated change management infrastructure, centralised governance capability, and the budget to absorb the learning curve.
The structural risk for mid-market consultancies isn't that the Big Four will immediately outperform them in AI delivery. It's that within 12–18 months, their clients will start asking questions that assume Big Four-level AI capability - and expect it from their other advisors too.
IG Group hit full ROI in three months using Claude Skills. Deloitte and Google Cloud research shows 74% of deployments achieve ROI within the first year. These numbers are achievable for mid-market firms - but only for those who start now, not those who start when the questions arrive.
Issues/backlash
Governance at scale is still unresolved. Security researchers continue to document prompt injection and data exfiltration risks in production deployments. The EU/EFTA/UK data residency constraints on Claude's Microsoft Copilot integration (Claude excluded from EU Data Boundary by default) remain live compliance questions for any firm operating across those jurisdictions.
There is also a legitimate question about client data governance when AI agents are running across client engagements at enterprise scale: who owns the outputs, who is liable for errors, and how are client confidentiality obligations maintained when AI tools can surface patterns across multiple client datasets simultaneously? The Big Four have legal teams to resolve these questions. Smaller firms need to resolve them before clients ask.
My take (what to do)
Startup (15–40 staff): The KPMG and PwC announcements are not a threat to you today - they are a signal to act this quarter. The deployments at scale prove the pattern works. Your advantage is speed and specificity. Pick the one workflow that your best people do manually 10+ times weekly: proposal assembly, discovery-to-brief, status reporting. Build one skill this week. Calculate ROI using the formula: (time saved per instance × weekly frequency × 52 × partner hourly rate) - (setup hours × hourly rate + annual subscription). If the number is positive within six months, proceed. If not, the workflow isn't ready.
SMB (50–120 staff): You are in the window. Large firms are operationalising. You can still build domain-specific AI delivery capability that is faster and more tailored than a firm deploying to 276,000 generalists. The question is whether you assign a skills owner and build a governed portfolio, or wait until client conversations force the issue. Assign one delivery lead to document your three highest-volume workflows accurately over two weeks of real usage. Time them. Calculate ROI including maintenance. One skill in production and measured is worth more than five in pilot.
Enterprise (150–250 staff): Your competitive positioning question has changed. "Do we use AI?" is no longer the question. "What is our AI delivery capability relative to firms that have deployed to 350,000 staff?" is. The answer needs to be domain depth and deployment speed - vertical skills for your specific client industries, not horizontal productivity tools. Build the governance framework now: centralised skills registry, approval workflows for anything touching client data, audit logs, rollback procedures. Then deploy aggressively within that framework.
How to try (15-minute path)
Open the Anthropic newsroom and read both the KPMG and PwC announcements. Note specifically how each firm is positioning Claude - what client-facing outcomes they are claiming, not just internal productivity. (5 min)
Map one current engagement where a competitor with AI-enabled staff could deliver the same output faster or cheaper. Write down the specific workflow step. (5 min)
Calculate the annual cost of that workflow in your firm: time per instance × weekly frequency × 52 × average loaded cost of the person doing it. If the number exceeds £8,000 annually and the workflow follows consistent steps, you have your first skills candidate. (5 min)
Success metric: A specific workflow identified, costed, and ready for a build decision - not a general intention to "explore AI."
"The conversation around AI has shifted from possibility to execution. Clients are looking for ways to apply AI that are secure, responsible, and capable of delivering measurable outcomes in complex business environments."
Spotlight tool | Google Gemini Enterprise Agent Platform
Launched at Google Cloud Next 2026, this replaces Vertex AI as Google's primary enterprise agent development environment - bundling agent building, deployment, data integration, security, and optimisation under one roof.
The Agent Runtime now supports long-running agents that maintain state for days, with a persistent Memory Bank. Payhawk reported 50%+ reduction in expense submission time using the Financial Controller Agent built on it.
→ Agent Development Kit for technical teams
→ Long-running agents with multi-day state persistence
→ Memory Bank for persistent cross-session context
→ Managed MCP infrastructure via Apigee
What did you think of today's issue?
Did you find it useful? Or have questions? Please drop me a note., I respond to all emails. Simply reply to the newsletter or email [email protected].
This issue’s sponsor
n8n
An open‑source automation platform that lets you chain tools like DeepSeek, OpenAI, Gemini and your existing SaaS into real business workflows without paying per step. Ideal as the backbone for your first serious AI automations.
Refer and win
Share this newsletter for a chance to win!