AI Labs Just Entered Consulting. Small Agencies Should Be Bullish.

Anthropic just put $300 million into starting a consulting firm. The thesis is simple. Frontier AI labs need services to drive model consumption, and any agency that still treats implementation as work that is beneath them just lost the next ten years.

On May 4, 2026, the world's most valuable AI lab co-founded a $1.5B services company with Blackstone, Goldman Sachs, Hellman & Friedman, Apollo, General Atlantic, Leonard Green, GIC, and Sequoia. Read the cap table twice, then ask why a model company is paying to do delivery work.

Most takes read this as labs eating agencies. They are wrong. Labs are using services as distribution because models do not consume themselves. Below is the cap table, the segment economics, and the moves a small AI-native shop should make this quarter.

• Anthropic, Blackstone, Hellman & Friedman, Goldman, Apollo, General Atlantic, Leonard Green, GIC, and Sequoia put $1.5B into a consulting JV. Anthropic alone wrote a $300M check. The structure treats services as a first-class asset class.
• OpenAI is doing the same play at roughly twice the scale. *The Development Company* raised $4B from 19 investors at a $10B valuation in the same week. This is a category move, not a one-off.
• The JV targets PE-owned mid-size businesses. Below that line, roughly 310 million companies still have no automation and roughly 1 million people on earth can deliver it. The JV does not touch them.
• Selling tools puts you in a permanent race against the foundation model. Selling implementation puts you in the budget line that grew because the foundation model got better.
• The agency move that wins the next ten years is local, scope-led, code-first, and built on the same models the JV is selling at enterprise rates.

Read the cap table before the press release

Strip the WSJ headline and the structure is a syndicate. Anthropic, Blackstone, and Hellman & Friedman each committed roughly $300M. Goldman Sachs added about $150M. Apollo Global Management, General Atlantic, Leonard Green, GIC, and Sequoia Capital filled the rest, for total committed capital of approximately $1.5B.

The useful enterprise pitch is: embed Anthropic engineers inside portfolio companies and redesign workflows around Claude. That points at the system-integrator layer between a foundation model and a working business. webvise works in the smaller-company version of that space: scoped workflow systems, AI-assisted processes, and production software around real operations.

OpenAI announced *The Development Company* in the same week, raising $4B from 19 investors at a $10B valuation. Same shape, different brand. Both labs concluded that direct services delivery is now too important to outsource to Accenture or Deloitte.

Why labs need services to consume their own models

A model behind an API only earns when somebody points workflows at it. Most enterprises still cannot do that themselves. Their data is scattered across SAP, Salesforce, ServiceNow, and a CMS that the team has not touched since 2019.

The gap between 'Claude is impressive' and 'Claude is in the P&L' is implementation work. PE portfolio companies are now writing checks for that exact gap.

Sequoia partner Julien Bek published the math earlier this year. For every $1 of software budget, $6 to $12 of services spend sits next to it. The accountant costs more than the QuickBooks license. The lawyer costs more than Westlaw.

AI did not change that ratio. It made the services side reachable for small teams, and labs noticed before agencies did. An earlier post on copilot and autopilot agency models covered the offering-shape side of the argument; this post covers the lab-strategy side.

The segment Anthropic and Goldman cannot reach

The JV will not call your prospect. The press release is explicit: PE-owned portfolio companies, embedded Anthropic engineers, workflow redesign at scale. In practice that means revenue above roughly $100M and an operating partner who can sponsor the engagement.

Below that line lives the segment that actually represents most of the demand.

Roughly 310 million companies globally have no meaningful automation, and roughly 1 million people on earth can actually deliver it. The JV's ICP covers maybe 10,000 of those companies. The other 309.99 million are not getting a Goldman partner on the call.

The market is the companies outside the JV ICP.

What implementation as channel looks like inside a small agency

This shape has been the operating model here for a while. The stack is Next.js 16, TypeScript, tRPC, Drizzle ORM, Neon Postgres, Better Auth, Vercel AI SDK, and Mastra, deployed on Vercel through Turborepo. Every production change passes through Claude Code in the pipeline.

Scoping now runs around the system being built instead of fixed package prices. Focused builds, custom systems, and ongoing support are estimated around the workflow, users, integrations, data model, AI requirements, and launch support. The same five-stage process runs from a contained landing page improvement to a custom AI workflow build: Discovery, Planning, Execution, Optimization, Launch.

The useful lesson is practical: Claude-native delivery only matters when it is tied to a concrete workflow, a data model, and a production handoff. That is the quote shape buyers can evaluate.

The current AI engagement shape: a scoped discovery or prototype phase followed by a build sprint that wires Claude into one operational workflow the client was paying a human to do. Lead enrichment, support triage, content drafting, document extraction. These work types show up in public AI-services playbooks because they can produce verifiable ROI inside ninety days.

Decide: compete, or ride alongside

The right read is rarely either-or. Enterprise sales motions leave delivery gaps that the named-on-the-deck firm cannot fill alone. Local implementation partners get pulled in. Position now to be the partner that gets pulled in.

If you have any kind of scoped-delivery track record and a Claude-native stack, the decision tree is shorter than it feels. You compete in everything below mid-market and you ride alongside in everything above it. webvise is taking on a small number of AI consulting and AI automation engagements this quarter under exactly that read.

What to do this quarter if you run a small AI-native shop

Three moves, in priority order.

• Pick one outsourced line item inside one industry. ColdIQ proved this at $7M ARR; the JV is proving it at $1.5B. Narrow accumulates proprietary data faster than broad.
• Stop calling implementation 'integration.' Buyers do not understand 'integration.' They understand 'we will replace the work that costs you €80K a year and hand you the audit log.' Same thing, different sales conversation.
• Build inbound first, outbound second. Manual cold outbound looks best on paper and does not ship from a small team. Lean on content, X cadence, warm-network projects, and the kind of decision-tree blog posts that LLMs cite directly. Outbound makes sense once you have agentic infrastructure that drafts and queues sends behind a daily approval batch.

When this thesis breaks

The bullish read assumes you can actually deliver. If you cannot, the JV is a death sentence rather than a tailwind. Three honest tests.

• You cannot ship a Claude-native automation end-to-end without subcontracting. Then your moat against the JV is zero. The JV will outdeliver you on price and credibility before the year ends.
• Your pricing has not moved in three years. A €30 per hour rate set in 2023 means the buyer thinks of you as labor, not as a partner. The JV charges enterprise rates for the same work; you have to move.
• You cannot point to a named client and a verifiable outcome. No case study, no pricing power. The JV's pitch is a Goldman Sachs logo. Yours has to be a paid result.

The JV signals where the market is moving. Frontier labs just told the entire AI services industry that implementation is the channel, that services budgets are real, and that the next ten years of AI revenue lives inside the workflows their models cannot configure themselves. If you are still selling tools, this is the warning shot. webvise is built around the first read and is taking on a small number of AI consulting and AI automation engagements this quarter.