What If OpenAI & Anthropic Do It?

Every founder building in AI has heard the question at least once. Usually from an investor, near the end of the meeting.

“What if OpenAI starts doing what you do?”

Most of us answer with something about our data, our relationships, our head start. Then we go home and worry about it anyway.

So this week I took the question at its worst: the labs build the models, sell the work themselves, keep the margin, and leave nothing for the rest of us. I made that case as strong as the evidence allows, and then I tested it.

In July I wrote that the labs were moving into professional services. They've moved faster than I expected.

In May, OpenAI launched its own company to put AI to work inside other businesses. It raised more than $4 billion from 19 investors led by TPG, and OpenAI owns the majority. It has agreed to buy two firms that already do this work: Tomoro, which brings about 150 engineers, and Northslope, founded by former Palantir engineers.

Anthropic set up its own version, Ode, with Blackstone, Hellman & Friedman and Goldman Sachs. The WSJ reported about $1.5 billion behind it. Ode has bought two AI services firms and goes after mid-sized companies: community banks, manufacturers, regional health systems.

Amazon committed $1 billion to its own engineers who work inside customers. Microsoft says it's putting $2.5 billion and around 6,000 specialists into a similar unit. And in September OpenAI launched Astra for Law, a legal research product used by firms like Latham & Watkins and Sullivan & Cromwell.

The money behind it keeps growing. Anthropic told investors its quarterly revenue went from $4.7 billion in the first quarter to $16.8 billion in the third. France's competition authority found that OpenAI, Google and Anthropic held more than 84% of the market for AI agents.

Put that together and the worst case writes itself. AI does more of the work, three companies own the AI, and those three start selling the work directly. Every law firm, accountancy and consultancy ends up either reselling someone else's model or not needed at all.

For most of this year, markets priced in some of that. When Anthropic launched its legal plugins in February, about $285 billion came off software, data and services stocks, and Thomson Reuters had its worst day on record. In June, Accenture fell 18% in a day.

It's a serious argument. But when I tested it piece by piece, most of it didn't hold.

Anthropic reportedly makes around 85% gross margin selling access to its models. Accenture made 32% in its latest year.

Services revenue would drag the labs' margins down, and both have been careful about how much of it they take on. Anthropic is one investor in Ode among several: the WSJ put its share at about $300 million of the $1.5 billion. OpenAI owns most of its unit, but outside investors put up most of the money.

My read is that the labs want AI deployed everywhere far more than they want to do the deploying.

Anthropic committed $100 million to its partner network in March. In October it announced a $100 million academy to train 10,000 deployment engineers by the end of 2027, and the first cohorts come from Accenture, Deloitte, McKinsey, Bain and Capgemini. OpenAI has multi-year alliances with Accenture, BCG, Capgemini and McKinsey. Google is training up to 1,000 engineers inside Accenture. Bain is even an investor in OpenAI's own deployment company.

Cloud went the same way. In a study Microsoft commissioned, IDC found its partners earn about $8.45 in services for every dollar of Microsoft revenue. Amazon's own study puts its partners at up to $7.13. Both are vendor studies, so treat them with care, but the consultancies did well out of cloud: Accenture's cloud business grew from about $12 billion to $18 billion in a single year.

The labs can't put AI inside every company themselves. There aren't enough engineers, so they're training other people's.

In most of the US, a non-lawyer can't own a law firm, and CPAs have to own the majority of an accounting firm. Even KPMG needed a special licence, with conditions, to practise law in Arizona.

The labs aren't trying to get round this. OpenAI's usage policy prohibits tailored legal advice "without appropriate involvement by a licensed professional." Anthropic's updated policy says a qualified person must remain accountable for legal and financial uses.

This only protects regulated work, though. Consulting and IT services have no licence to hold, and that's exactly where the labs went first.

When spreadsheets spread, about 400,000 US bookkeeping and accounting clerk jobs disappeared. Over the same period, around 600,000 accounting jobs were added. ATMs did something similar: banks needed fewer tellers per branch, opened more branches, and teller jobs didn't fall.

There's plenty of unmet demand. Low-income Americans get no or inadequate help with 92% of the civil legal problems that seriously affect them, according to the Legal Services Corporation, and almost half of those who don't seek help blame the cost. The US is also short of accountants.

More demand doesn't guarantee the work goes to today's firms. But a bigger market leaves more room for everyone in it.

Only 46% of people say they're willing to trust AI systems, according to KPMG and the University of Melbourne. Most corporate legal departments want their outside firms to use AI, which means they want it from advisers they already know.

The FTC, the UK's competition authority and the European Commission have all looked at the labs' deals, and that 84% figure came from a regulator. None of it stops anything today. It does make a full takeover of a regulated industry slower and more expensive.

Parts of the fear are justified.

If you put AI to work inside companies for a living, you now compete with your suppliers. That's the business Ode, OpenAI, Amazon and Microsoft are all moving into.

Repeatable work gets squeezed hardest: document review, standard IT projects, coding billed by the hour. You saw the hourly billing numbers this week.

And the tools in the middle are in an awkward spot. OpenAI lists Harvey and Legora as partners for Astra for Law, while selling its own legal product to the same law firms.

So the labs probably won't take your job. They may well take your pricing.

In July I said a lab would launch or majority-own a professional services firm within 24 months. That's partly true already: OpenAI majority-owns a company that is buying consultancies, though there's nothing in law or accounting yet.

I also said a major vertical AI tool would launch a services arm, or be bought into one, within a year. Not yet, with nine months to go.

And I expected vague official warnings against Chinese open-weight models. Treasury did threaten sanctions over IP theft, but in August the White House reportedly told companies those models won't face the new government tests. That one looks wrong so far.

I had the direction right. I was too confident about how loud it would get.

The labs will own the intelligence and a real share of the deployment. I don't think they'll own the accountable work... the advice someone signs, the outcome a firm guarantees, the client who calls when it goes wrong.

That's where the value goes: to licensed firms that own the outcome and can switch between models whenever they like. By the end of 2027, I don't expect any frontier lab to majority-own a licensed law or accounting firm.

I'm biased. Eilla is built on that bet: AI does a lot of the manual work, experienced advisers check it and answer for it, and we're mostly paid when deals close.

So the next time someone asks "what if OpenAI does it?", the honest answer is that it's already doing part of it. What matters is which part you own.

What would a client still pay you for if AI could do everything else?