Hi, it's Andreas here.

Anthropic and the buyout firm Hg are setting up a consultancy to put AI engineers inside European companies, the FT reports. The same playbook has run in the US since May.

December will take me to the UAE for some exciting work. I'll be working with a telecommunications provider on an ambitious agentic AI initiative, while also speaking at GITEX in Dubai, where I'll deliver a keynote and a panel. I still have a few slots open. If you're looking to accelerate your AI initiatives, explore a collaboration, or simply meet for coffee while I'm there, let's connect.

In this issue:

Anthropic files to go public and discloses unauthorized access by its models in tests → First-half revenue was $16.3bn, up from $1.2bn a year earlier. Four of its models reached real third-party systems when a test environment was left online.

Alphabet falls 13% in a day on negative free cash flow and an AI Mode miss → Q3 free cash flow was minus $9.07bn as capex guidance held at up to $205bn. Search grew 12% but missed; Google cited AI Mode monetization.

Microsoft offers voluntary buyouts to about 60,000 staff, a quarter of its workforce → AP reports the figure; Microsoft's announcement covers teams in the US, UK, Canada and India. A spokesperson told KUOW that AI is not driving it.

Google's Gemini 4 Carbon tops the leaderboard, and cyber defenders get it first → It ranks first on the Artificial Analysis Intelligence Index. Developers and businesses have no release date yet.

Agents now draft more than 40% of pull requests on GitHub → Up from 17% in March, per GitHub. Copilot now comes bundled with GitHub Enterprise for account-team customers, metered by usage.

OpenAI pools ChatGPT and Codex limits as Anthropic renames its top plan Ultra → From December, ChatGPT and Codex draw on one token-based allowance. Claude's renamed Ultra plan gives 10x Pro weekly usage, with no five-hour cap once verified.

Also this week:

I am convinced that the biggest opportunity in AI right now sits in the deployment layer: the work of getting AI into how companies actually run. Most people greatly underestimate the amount of work required to change workflows in enterprises; it tends to be far greater than anyone realizes or would prefer. There is so much to consider: legacy systems moving to the cloud, data access being fixed, software being connected to agents, workflows being redesigned, the human in the loop being worked out for each process, evaluations being written and kept current, and all of it being updated every time a new model is released.

That is what all of this makes this different from the software era, where you implemented a well-understood category of software, stepped back, and the customer kept it running. With agents, you're deploying work output in a process. The implementation never really ends, and neither does the enablement.

And the AI labs have seen the exact same thing. Last week Anthropic has agreed to back a new European consultancy with the buyout firm Hg, the FT reported on October 23, citing people familiar with the plan. It would place AI consultants and engineers inside companies from offices in London and Berlin, and launch before the end of the year. Neither company commented, and the plan does not appear in Anthropic's IPO filing three days later, so treat the details as unconfirmed. On the record, Hg announced "a new strategic venture" with Anthropic in September, for its own portfolio.

The idea itself is not completely new. Palantir has been following a similar approach for quite some time and has made the Forward Deployed Engineer role well-known. During my time at IBM, the company sold software and provided consultants to implement it under one roof for decades. In my experience, clients knew whose software the consultant would propose and priced that in. What's new is that the vendor is now a model lab, and the clients arrive more and more bulked through private equity. In a company owned by a buyout fund, the firm that builds your AI may now be chosen two levels above the CIO.

The U.S. typically leads by 6-9 months in AI trends. This consultancy approach, now in Europe, started in the U.S. earlier this year with labs initially renting consultancies. Accenture set up an Anthropic Business Group to train about 30,000 of its people on Claude, and Deloitte made Claude available to 470,000 staff. Then this year:

FEB  OpenAI signs multi-year alliances with McKinsey, BCG, Accenture and Capgemini.

MAY 4  Anthropic, Blackstone, Hellman & Friedman and Goldman Sachs announce an AI services firm for mid-sized companies.

MAY 11  OpenAI launches the Deployment Company, majority-owned, with more than $4bn from 19 partners led by TPG.

MAY 21  Anthropic's venture buys Fractional AI, a San Francisco boutique already working across Blackstone's and H&F's portfolios.

JUL  The venture launches as Ode, and OpenAI's venture agrees to buy Northslope.

Then came the sales. By my count, Ode's press releases have named six clients since launch, and five have Blackstone or Hellman & Friedman among their owners. The fund supports the consultancy and holds stakes in the clients, so one decision by the sponsor can open doors across dozens of companies and create a massive scaling effect. The lab gets usage, and the consultancy gets a pipeline it didn't have to find. The contest is over who owns the client relationship, and in a portfolio company that may now be the fund.

Below: why Anthropic wants this work, what the funds get out of it, and who will end up doing it.

Anthropic gave the first reason itself in May: mid-sized companies "stand to gain from AI, but lack the in-house resources."

The second reason is money, and I don't read it as diversification. About 80% of Anthropic's first-half revenue, $13.1bn of $16.3bn, was usage-based, according to its S-1. An agent that a services team puts into production keeps drawing tokens after the engineers leave. So the consultancy can price the build low, because the margin sits in the run.

The third is learning. This is the forward-deployed model Palantir made famous, and OpenAI says patterns its own forward-deployed engineers find at clients "evolve into product capabilities". Your problem becomes their roadmap.

The fourth is that Anthropic is behind in Europe. OpenAI's delivery arm has had a British base since May, when it agreed to buy Tomoro, about 150 forward-deployed engineers and deployment specialists in Edinburgh and London.In comparison, when I interviewed at Anthropic in May, they had only two people on the ground in DACH (Germany, Austria and Switzerland). Things are changing, though.

Hg's portfolio holds more than 60 European and transatlantic technology and services businesses. In its September release about Anthropic, Hg noted that companies with credible, product-embedded AI strategies are "attracting stronger buyer interest". For a buyout fund, AI in the product raises the price at exit.

The labs are right to go after this work, and wrong if they think they can do it alone. So I don't expect three lab-backed firms to own this market. I expect a wave of new services firms, organized by industry, by company size and by problem, while some of the old integrators adapt and some don't. Their edge is the answers a lab-owned firm finds hardest to give, "you don't need a model for this" or "the cheaper one is good enough", when those answers are true.

AI services firms have begun to perform well in the US and are now expanding across other markets (Europe first). It is an excellent time to be a forward-deployed engineer or to establish a firm specializing in this field. Full disclosure: I'm in the process of building a small firm of my own for this type of work.

DeepLearning.AI's short course on agent loops with LangChain Deep Agents → Eight lessons: an agent loop, a verification loop that grades output against a rubric, Slack or Jira triggers, and improvement across runs.

Claude Academy's course on rolling out Claude Enterprise → 14 lessons, in dependency order, on structure and identity, access, governance, spend and visibility. You finish with a rollout plan.

NVIDIA's open Nemotron 3.5 Transcribe turns recordings into text on your own hardware → 40 languages, recordings up to an hour, speaker labels in 13 of them. An option for teams that can't send meeting audio to a cloud service.

I take on a small number of advisory engagements each quarter for AI programs that are stuck, or for a second read of an AI proposal before you sign. Reply to this email with the one initiative you would want a second pair of eyes on. And if you want to build agents rather than read about them, the waitlist for the fourth Agentic AI Cohort is open, and you can start the Build Sprint today.

Has a vendor offered to embed an AI team with you, or are you on one? Reply with one line on how it is going. I read every reply.

— Andreas