The Model Is a Commodity Now. MCP Is the Moat.

The race was never about the model. It was about MCP.

Most people missed that. For three years, everyone watched the benchmark charts — GPT-4 vs Claude vs Gemini — like they were scorecards for who'd win the AI era. Better coding scores. Longer context. Faster inference.

Thomson Reuters just showed us the actual scoreboard.

What Thomson Reuters Built on MCP

On May 12, CoCounsel — the legal AI used by 1 million professionals across 107 countries — announced it's being fully rebuilt on Anthropic's Claude Agent SDK.

Not just integrated with Claude. Rebuilt on it, from the ground up.

The new CoCounsel agent plans inquiries, selects tools, retrieves authoritative content, and adapts mid-workflow. Lawyers describe a matter in plain language. The agent pursues the right legal question, drafts with citations, and produces fiduciary-grade work product. Not a chatbot. Not autocomplete. An agent that picks up a legal problem and works it like a first-year associate — except the associate has read every case in Westlaw that has ever been filed.

What makes this different from every other "we added AI" press release is what's on the other end of the MCP connection.

175 years of curated legal data. Westlaw. Practical Law. KeyCite. Content validated by practicing attorneys and legal specialists — not scraped from the open web, not approximately recalled from somewhere in the model weights. Actually queried at runtime, from an authoritative source, with the citation chain intact.

That's the actual product. The Claude Agent SDK is the machinery. The MCP connection is what makes the work defensible.

MCP Isn't a Developer Protocol. It's a Business Model.

I've been thinking about MCP wrong since it shipped.

When it launched, I filed it as "a standard way to give AI tools access to external services." Useful for developers. Great for building integrations. A cleaner plugin system than what came before.

That's not what Thomson Reuters is doing with it.

Thomson Reuters is using MCP to turn 175 years of legal infrastructure into an AI product without rebuilding the infrastructure. The knowledge is already there. The editorial curation is already there. The trust relationships with law firms are already there. What was missing was a standard protocol for an AI agent to query that corpus at runtime, mid-workflow, in response to a specific legal question a lawyer is working on right now.

MCP is that protocol.

Any company that has spent decades building an authoritative domain corpus can now wire it into Claude and ship an agent product. You don't need to fine-tune a model on your proprietary corpus. You don't need to embed your entire database and hope RAG retrieves the right chunk. You wire your data source into an agent via MCP, and the agent retrieves what it needs, when it needs it, with the original context intact.

For Thomson Reuters, the moat is 175 years of curated legal content combined with the trust that content carries in a courtroom. MCP is the hinge that connects it to the model.

What "Fiduciary-Grade AI" Actually Means

Thomson Reuters used an interesting phrase in the announcement: "fiduciary-grade AI." Their standard for "accuracy, accountability and trust."

That phrase is doing a lot of work.

In law and finance, the stakes of a wrong answer are not "the user was annoyed." They're: malpractice suits, regulatory action, financial loss for clients. A citation that doesn't exist is not a quality problem — it's a professional liability problem. A risk rating that's off by a grade is not a UX issue — it's a decision that cascades into real money.

Generic AI products cannot meet this standard. Not because the model isn't capable of good reasoning, but because the model doesn't know what it doesn't know. It will confabulate a Westlaw citation. It will approximate a financial rating from training data that's months old. It will be confident and wrong in ways that are invisible until the brief is already filed.

Fiduciary-grade AI requires fiduciary-grade data. That data doesn't come from pre-training. It comes from runtime queries to authoritative sources via MCP.

This is a new category. And it's going to matter in every domain where being wrong has real consequences.

Finance Did the Same Thing at the Same Time

One week before the CoCounsel announcement, Anthropic launched Claude Finance: 10 ready-to-run agent templates built for financial services work. Pitchbook creation. KYC screening. Month-end close. Analyst briefs.

Same pattern.

Anthropics Moody's data partnership supplies verified financial intelligence to the agents. When a Claude Finance agent analyzes a company's risk profile, it's not reasoning from whatever the model absorbed during training. It's pulling from authoritative financial data — the kind that used to require a junior analyst and three days to compile.

Two vertical launches in ten days. Legal and finance. Both industries where a wrong answer has legal and financial consequences, not just user frustration.

That's not a coincidence. That's a repeating pattern: for high-stakes domains, MCP connects models to the authoritative data that makes agents trustworthy.

The Uncomfortable Take

Here's what nobody building "AI for X" startups wants to hear.

Most AI products right now are the model plus prompting. Take Claude or GPT-4, write a system prompt, build a clean UI, ship it. That worked in 2024. For some niches, it still works.

But Thomson Reuters isn't building "the model plus prompting." They're building the model plus 175 years of curated, legally defensible content, delivered via MCP, with trust-layer guarantees that no generic wrapper will ever replicate.

The wrapper is not the product. The data is the product.

And the data is controlled by the people who spent decades building it.

If you're building a generic "AI for legal" or "AI for finance" SaaS right now — with no proprietary data advantage, no editorial moat, no domain corpus you own — you are competing directly with companies that have been accumulating those data assets since before the internet existed. They're now wiring those assets into Claude via MCP and shipping vertical products.

Westlaw has 175 years. Moody's has over 150. What does your product have?

The model is the commodity. The data pipeline is the moat. This was always going to be how it ended.

What Developers Should Actually Build

I'm not saying don't build AI products. I'm saying: understand what you're actually competing with.

If you don't have a proprietary data moat, compete on depth and workflow specificity. Skills and subagents let you build agents that understand one workflow, one codebase, one client's operational reality — better than any general-purpose product ever will. That's still a moat. It's narrower, but it's real, and Thomson Reuters can't take it from you.

But if you do have data — if you've spent years building up an authoritative, curated, domain-specific corpus — MCP is the most important thing that happened to you in 2026. You don't need to pivot to AI. You need to wire your existing data into an agent and ship.

The infrastructure is there today.

# initialize an MCP server that exposes your proprietary data
npx @modelcontextprotocol/sdk create --name your-domain-server
# expose your data as callable tools
# connect to Claude via the Agent SDK

One MCP server that surfaces your proprietary corpus. One Claude agent that knows how to use it. A domain your users already trust you in.

That's the product.

The Clock Is Running

CoCounsel goes generally available summer 2026. Claude Finance is live now.

Every week that passes is a week the companies with authoritative data are shipping agent products while everyone else is still debating chunking strategies and vector database benchmarks.

The model race ended in commoditization. That's what always happens to infrastructure. The winners in every platform shift aren't the ones who built the infrastructure — they're the ones who used it fastest to build something nobody else could build.

In AI, that thing is a vertically specialized agent backed by data the model couldn't access any other way.

Figure out what you own. Wire it in.