A companion piece to the latest episode of Human in the Loop, the podcast where we argue about how much of the loop AI should actually be in.

If you’ve spent any time on LinkedIn lately, you’ve seen the pitch: hook up Claude, plug it straight into Amazon, and let it run your business. Set bids. Edit listings. Rewrite copy. Optimize the campaign. Maybe sleep through the whole quarter.

It’s a great-sounding pitch. It is, also, mostly nonsense — at least if you’re an enterprise brand doing tens of millions in monthly GMV with hundreds of thousands of dollars a month in ad spend.

That doesn’t mean MCP is hype. It means MCP is being misread.

This post unpacks the conversation we had on the podcast: what Model Context Protocol actually is, where it’s already useful today, where it falls apart for serious marketplace operators, what a sane architecture looks like, what it does to headcount, which SaaS categories are most exposed, and what the right move is for a brand or seller listening to this in 2026.

First, the definition: MCP is a glorified API

Let’s strip the buzzwords. Model Context Protocol (MCP) is a standard, originally from Anthropic, that defines how applications expose data and tools to large language models in a way the model can actually use. That’s it.

It’s not an AI. It’s not a model. It’s not a strategy. It’s a specification — a contract for how, say, an Amazon-side service or a HubSpot account or your internal database describes what it knows and what it can do, so a model like Claude can query it intelligently and take actions through it.

The reason it matters isn’t that it’s magical. It’s that historically, every time you wanted an LLM to “talk to” an external system, somebody had to hand-build a brittle integration: custom prompts, custom tool definitions, custom auth, custom context windows. MCP turns that one-off integration work into something closer to plug-and-play. Build the server once, any MCP-aware model can use it.

So when you see a vendor proudly announcing “MCP support,” what they’re actually saying is: we’ve made our data and our actions available to LLMs through a standard pipe. That’s useful. It is not, by itself, “an AI that runs your business.”

That distinction is doing a lot of work in the rest of this post.

What changes when the pipe gets standardized

The reason MCP is a real shift, even though it’s “just” a protocol, is that LLMs are only as good as the inputs they’re given. You can use the most capable Claude variant on the planet — Opus for deep reasoning, Haiku for speed — and it doesn’t matter if the model is staring at a blank context window. Smart models with no inputs aren’t smart at anything except guessing.

Before MCP-style protocols were standard, “give the model the right inputs” meant a custom engineering project for every system you wanted to plug in. That’s a high tax on experimentation. With a standardized protocol, the cost of trying something — what if Claude could see our Amazon ad data? — drops dramatically.

So the right way to think about MCP is less “AI that runs my business” and more “the cabling that finally lets you wire your business into AI without a six-month integration project.” That’s still a big deal. It’s just a different big deal than the LinkedIn version.

Where it works today: analytics, querying, exploration

There’s a class of work that an LLM connected to your marketplace data through MCP is genuinely good at right now.

The clearest one is analytics. Instead of opening a BI tool, building a dashboard, or wading through a CSV export, you ask in plain English: “Why did our ACOS spike on these three SKUs last week?” or “Show me where TACOS improved month over month and what creative changed.” The model pulls the relevant slice through MCP, reasons about it, and gives you an answer with the work shown.

That’s a real productivity unlock. It compresses the gap between “I have a question” and “I have an answer” from hours to seconds, and it’s exactly the use case where the cost of being slightly wrong is low — you can always sanity-check the underlying numbers before acting.

A few other categories that work well today:

  • Cross-channel exploration. If you’ve got data for Amazon, Walmart, and TikTok Shop in a place the model can reach, asking “where am I leaving margin on the table across channels?” is a perfectly reasonable prompt.
  • Operator briefings. Generating the morning summary of what changed yesterday, what needs attention, what’s trending up.
  • Drafting and listing work. Rewriting bullets, suggesting A/B tests, generating creative variants.

These are all “model in an advisory role” use cases. The human still decides. The model just makes the human faster and more informed.

Where it breaks: full autonomy on a real ad budget

Now consider the version of the pitch where Claude isn’t advising — Claude is running the account. Claude is publishing bid changes. Claude is editing campaigns. Claude is the operator.

For a small test budget, sure — try it. For a $500,000-a-month ad spend on a real enterprise brand? It’s a stretch, and not because the model is dumb. It’s because the surrounding system isn’t there yet.

Three things specifically don’t hold up.

The data gap. What you can pull out of Amazon through MCP is not the same set of signals a serious bid-management system uses. Modern marketplace optimization tools are built on years of telemetry: bid adjustments at the placement level, hour-of-day patterns, share-of-voice data, cross-channel cannibalization signals between Amazon and TikTok or Amazon and Walmart. A lot of that simply isn’t exposed in the raw API surface an LLM is reaching through. You’re only as good as the data you can see — and through a single MCP connection, you can’t see enough.

The middle-layer problem. Going Claude → MCP → Amazon with nothing in between means you have no place to store history, no place to enforce business rules, no place to keep your own derived signals, and no place to maintain a record of decisions the model made and why. For querying, that’s fine. For operating a business, that’s a fragility you can’t accept.

The interface problem. LLMs are conversational. Marketplace operations are not. If Claude makes 412 bid changes overnight, where do you review them? Where’s the diff? Where’s the rollback button? Where’s the audit trail your finance team or your agency partner needs? Conversational UIs are extraordinary for some workflows and the wrong shape entirely for others. Ad ops is the wrong shape.

The takeaway isn’t don’t use MCP. It’s don’t expect MCP alone to be the operating system of your business. You need a layer of your own in between.

What a sane architecture actually looks like

If you were starting fresh today and you wanted to seriously bet on agentic AI inside a marketplace business, the architecture you’d build is not “Claude + MCP + Amazon.” It’s closer to this:

  • Your own database as the source of truth — pulling in Amazon, Walmart, TikTok Shop, Shopify, ad platforms, anything that matters — so your model never has to round-trip to a partner API just to remember what happened yesterday.
  • A middle application that exposes business-aware tools to the model: not raw “edit campaign,” but “raise bids on SKUs where ROAS exceeded target by X% for Y consecutive days within these guardrails.”
  • An interface layer where humans review what the model proposed, what it executed, and what it learned — diff views, audit trails, rollback, comments.
  • MCP as the cabling, not as the brain — used to connect both the partner systems (Amazon, Walmart) and the model (Claude) into your middle layer.

That architecture is not “we replaced our software with Claude.” It’s “we rebuilt our software with the model as a first-class participant.” Those are very different bets, and only one of them survives a real Q4.

The headcount question

People want a clean answer here, and there isn’t one.

Will MCP-plus-LLMs eliminate roles in a marketplace business? Almost certainly not, at least not the way the louder voices on LinkedIn suggest. The strategic work, the cross-functional judgment, the relationship management with retail account reps, the should we even be on this channel? call — none of that is on the table.

Will it consolidate roles? Yes. The clearest pattern is one operator doing what used to take three. The analyst, the campaign manager, and the reporting person can plausibly become one person plus a model. The model carries the volume; the human carries the judgment and the exception cases.

There’s a useful precedent here: Anthropic itself has written about running marketing with what was effectively a one-person-plus-AI team for a stretch. It doesn’t prove every business can do it — Anthropic is a unique company with unique distribution — but it does prove the shape is possible.

For a real marketplace brand, the practical version is more conservative: don’t fire your team, but next time you’re hiring, ask whether you’re actually hiring for a role or for throughput. If it’s throughput, you may not need to hire at all.

Which SaaS gets eaten — and which doesn’t

There’s a now-famous quote from an Nvidia executive that every software company is going to become a token factory. Strip the dramatic language and the underlying claim is reasonable: a lot of software is just a UI sitting on top of a database, and if a model can talk directly to the database and render the UI on demand, the moat of the UI itself starts looking thin.

Some categories are clearly exposed.

Legal tools are an obvious one — much of the work is reading text, summarizing text, generating text from templates, and answering questions about text. That’s exactly what LLMs are best at.

General-purpose workflow tools are exposed in interesting ways. HubSpot’s Breeze agent is a good example: instead of clicking through menus to build a multi-step form workflow, you just describe it in natural language and it gets built. Zapier is going down a similar road — type the automation you want and the model attempts to wire it up. The classical SaaS interface — screens, fields, save buttons — starts to feel optional.

Marketplace operations, on the other hand, are less exposed than they look. The reason is exactly what we covered earlier: operating a marketplace business is not “pass text around.” It’s a high-frequency optimization problem against partial, noisy data, with cross-channel side effects, and a real cost of being wrong. The model is a brilliant participant. It is not, on its own, the system.

That said, the actual competitive risk for marketplace SaaS isn’t Claude. It’s the lean, AI-native competitor that builds the next generation of optimization tooling agentic-from-day-one — fewer screens, more natural-language workflows, MCP everywhere, the model treated as a coworker rather than a feature. That’s the company that eats your lunch. And as a SaaS team, the response isn’t to fight the trend — it’s to be that company.

So what should an operator actually do?

If you run a brand or a marketplace business and you’ve made it this far, here’s the honest version of the takeaway.

Yes, dive into MCP. Set up an analytics-grade connection. Let your team query their data in plain English. You’ll get value within a week.

No, don’t hand it the budget. Not yet. Not without a middle layer you control and an interface that lets you see and reverse what it did.

Run a real test. Pick a slice of spend you can afford to be wrong about — a single campaign, a single channel, a single SKU — and let an MCP-connected model operate it within tight guardrails. Watch what it does. Compare it to what your team would have done. Decide based on evidence.

Watch the second-order shifts. The interesting question isn’t can I replace my agency with Claude? It’s what does my org look like when the model is doing 60% of the analytical work? That’s a hiring plan question, a tooling question, and a competitive-positioning question all at once.

Don’t wait for perfect. The ChatGPT analogy holds. The teams that integrated early — even when the models were rougher and the integrations more painful — got a meaningful head start. The same playbook is on offer here.

Want to listen to our full episode on MCP?

Check out Human in the Loop, with new episodes releasing every Wednesday.