Your most-asked question got answered by a chatbot last week. The asker didn't pay you. They didn't credit you. They didn't even know your name was in the training data.
That's a $4,000 consult that became a free reply. Multiply by the number of times your judgment shows up in a model output this month and you're staring at a real number. A six-figure number for most senior operators I talk to.
So here's the question worth sitting with: if agents are about to route the work, who owns the layer they route through? You, or someone monetizing your expertise without you in the room?
I'm Matt Cretzman. I build agent systems for experts, and I've spent the last eighteen months watching one shift coalesce. Microsoft has now named it out loud — 2026 is the year AI moves from chat to colleague. From tool to true partner. From answering your prompts to taking your meetings, owning your workflows, and routing your customers.
That shift is not neutral. It forces a choice every expert has been able to defer until now. Defer it past Q3 and the choice gets made for you.
The Three Lanes — And Why There's No Fourth
In the book I'm writing, On Whose Terms, Chapter 4 is called The Three Extractions. The whole thesis fits in one line: your expert knowledge is already being extracted. The only variable is whose terms it happens on.
Lane one is the LLM lane. Anthropic, OpenAI, Google, Meta — they ingested the open web, plus a lot that wasn't supposed to be open, and trained models on it. In the UK copyright fight, 88% of creators surveyed wanted protection. The government punted. The models shipped. Your blog posts, your podcast transcripts, your conference talks, your case studies — all of it is now answering questions for free, in someone else's product, under someone else's brand.
That's extraction on their terms.
Lane two is the platform lane. Mercor is the clearest example. Their pitch to experts is honest in a way I respect: come work hourly, dump your judgment into structured training data, get paid by the hour. The Foody interview at Redpoint last summer laid it out plainly — they're paying doctors, lawyers, engineers, and senior operators to convert tacit knowledge into model weights. The expert gets an hourly rate. The model gets the expert's career. Forever.
That's Jacob selling his birthright for a bowl of stew. It's a real check. It's also a one-time check against a permanent transfer.
Lane three is the expert-owned lane. You build the agent. You hold the IP. You set the price. You own the distribution. The customer pays you. The agent runs on your terms because the protocol underneath it — MCP — is now an open standard, not a vendor lock-in.
There is not a fourth lane. "I'll wait and see" is lane one by default, because the extraction is already happening to your published work. Silence is consent in this market.
What Microsoft Just Made Official
When the largest enterprise software company on the planet tells its customers that 2026 is the year AI becomes a colleague, they are telling their sales force how to sell. They are telling procurement how to budget. They are telling your clients what to expect.
Here's what "colleague" means in operational terms. Agents will hold context across sessions. They'll initiate work, not just respond to it. They'll join meetings, send follow-ups, draft proposals, and route inbound based on intent. They'll have memory, calendar access, tool access, and judgment delegated to them.
That last word is the one that matters. Judgment. The thing experts sell.
When a buyer's agent is doing the first three rounds of vetting on your behalf — or your competitor's behalf — the question of which knowledge base that agent is querying becomes the entire game. If it queries a general model, you're invisible. If it queries Mercor-trained weights, your competitor who sold their hours is the default answer. If it queries your published, owned, MCP-exposed expert system, you're the colleague the colleague calls.
Why MCP Changed The Math
When Anthropic launched the Model Context Protocol in November 2024, it looked like a developer convenience. A way to plug tools into Claude. Useful, niche, vendor-specific.
That's not what it is anymore. In December 2025 Anthropic donated MCP to the Linux Foundation, under the new Agentic AI Foundation. OpenAI adopted it. Google adopted it. Microsoft adopted it. Every serious agent framework now speaks MCP natively.
What that means for you: there is now an open, neutral, standard way for any agent — anywhere, on any platform, owned by any company — to call your expert knowledge as a tool. Not scrape it. Not train on it. Call it. Pay for it. Credit it.
This is the infrastructure shift I've been waiting on. Custom GPTs were a prototype. They lived inside one vendor's walled garden. MCP is the protocol layer that makes expert-owned extraction actually portable. Your agent answers questions inside ChatGPT, inside Claude, inside Copilot, inside whatever ships next — and the revenue and attribution route to you.
This is what I built Skill Refinery to do. It's a Knowledge Delivery System — a structured way to take a senior operator's IP, convert it into an MCP-exposed expert agent, and put it in the path of the agents that are about to do the routing. The expert owns the cards, the revenue, and the distribution.
The Forced Choice, Stated Plainly
Every expert reading this is in one of these three economies right now. Let me make the math obvious.
Lane one — LLM extraction. Your cost: every public asset you've ever produced, ingested, uncredited. Your revenue: zero. Your defense: nothing, because the training already happened.
Lane two — Mercor-style platforms. Your cost: your judgment converted to permanent weights in someone else's model. Your revenue: an hourly rate, capped, with no residual. Your defense: a paycheck that ends when the project ends.
Lane three — expert-owned agents. Your cost: the work of structuring your IP into a system that an agent can call. Your revenue: per-query, per-seat, per-engagement, with attribution. Your defense: you own the cards. You can pull them. You can price them. You can pass them down.
That last word matters to me. I think expertise is a gift that gets stewarded, not strip-mined. You spent twenty years earning judgment. The question of who inherits it — your kids, your firm, your clients, or a model trained by people who never met you — is not a small question. It's a stewardship question.
I don't think I'm building in this lane by accident.
What To Do This Quarter
Three moves, in order.
One — inventory your IP. Every framework, checklist, decision tree, and judgment pattern you've built. Most senior operators have between forty and two hundred distinct "how I think about X" assets sitting in their head, their email, or their notes. They are invisible until structured.
Two — pick a lane on purpose. If you're going to take the Mercor check, take it knowing what you're trading. If you're going to do nothing, do nothing knowing you're in lane one. Don't drift.
Three — if lane three is the answer, get an MCP-exposed expert agent into market before agents become the default routing layer in your industry. For most professional services, that window closes inside twelve months. For some, it's already closing.
I'm writing a book about all of this — On Whose Terms: The New Expert Economy and the Fight for What You Know. If the thesis is landing, join the launch list. I'll send the first chapter and the framework I use to audit which lane a given expert is actually in.
If you want to see the full system — the Knowledge Delivery System, the agent stack, the case studies — that's at mattcretzman.com.
The chatbots are about to become colleagues. The question is whose colleague, on whose terms, with whose judgment, paying whom.
Pick the lane.
Keep Building,
— Matt