Imagine sitting through a demo and the sales guy leans forward for the big reveal.
“And here’s the best part. You can log in.”
You’d wonder what the hell you’re doing there.
Of course you can log in. Access is assumed.
That’s the standard ecommerce software needs to meet with AI. An MCP connection. Access to the actual business data. The ability to do useful work with it. These belong in the baseline, alongside the login screen.
AI-native is non-negotiable. Having an MCP server is table stakes. What your customer can accomplish through it is the product.
All in on AI. Rounders (1998), © Miramax Films. Image links to Movieclips’ “All In” scene.
The Interview Gets to the Real Question
Credit to Ecomm Cowboy for the conversation that helped frame this piece. In the show’s published interview clip, the discussion gets to a question every software company should be asking: how much of an operator’s work could happen inside their preferred AI, with the underlying tools connected through MCP?
Guest Quentin Renevier, CMO of TrendTrack, describes using MCP to compress the repetitive work of competitor-ad research. He also draws a boundary: creative work can be automated, but he is uncertain about what advertising platforms will permit when it comes to automating launches.
Then he makes the point worth keeping: human experience becomes more valuable when everyone can generate similar-looking work.
Our take on that conversation is the argument of this article: the connection is becoming the baseline. Judgment and execution are where the competition moves.
Watch the original interview excerpt, published by Ecomm Cowboy, or find the podcast on Apple Podcasts and the show on X.
Your Customer Shouldn’t Be the Integration
Think about the workflow we’ve trained operators to accept.
Open the billing platform. Export the transactions. Open the subscription platform. Export the cancellations. Pull the campaign report. Clean the column names. Upload everything to an AI. Explain what every field means. Ask the question. Take the answer back into three dashboards and make the changes yourself.
Then repeat it tomorrow, because today’s files are already yesterday’s business.
Somewhere in that process, the customer became the integration between the products they pay for.
Software companies should be embarrassed by how much of that work still falls on the operator. Selling somebody a system and then making them carry its data around in a spreadsheet is a hell of a definition of productivity.
The operator should be able to ask a business question, let the AI retrieve the relevant records, inspect the answer, and authorize the next step. The underlying software should make that possible without a daily export ritual.
MCP Is the Door
Model Context Protocol gives AI applications a standard way to connect to external systems. Those systems can expose data and tools the AI can use.
That’s why the login analogy works. Your login gives a human a way into the product. An MCP connection gives an AI client a structured way to work with the capabilities the product exposes, subject to its access controls.
But a login screen doesn’t tell you whether the software behind it is any good.
Neither does an MCP badge.
A connector that retrieves one summary report can be useful. It still leaves a lot of work on your desk. The question is how much of the real operation the connection supports, how clearly it represents that operation, and whether you can trust the result.
MCP is a protocol. It doesn’t magically fix contradictory records, invent missing business rules, or make an unreliable action reliable. A platform has to do that work.
Make AI-Native Mean Something
Connected Doesn’t Mean Understood
Connectors vary in the data they expose, the actions they support, and the business context they carry. A connection can work perfectly while leaving your AI to guess what your business means by “active customer” or “recovered revenue.”
That’s the job of a semantic layer: consistent definitions of business concepts, metrics, and the relationships between records. Which transaction belongs to which renewal? Does “canceled” mean the customer chose to leave, or that billing gave up? Does “revenue” mean money collected before refunds or after them?
Those definitions change the answer. Get them wrong and you can produce a technically valid report that tells the operator the wrong story.
Now spread that problem across a stitched-together stack. Your CRM, subscription app, payment gateway, and email platform can each hold a different version of the same customer’s story. A subscriber’s payment fails, their subscription gets suspended, and your email platform still labels them an active customer. Those records can all be accurate within their own systems. They still need to be reconciled before an AI can answer, “Did we lose this customer, and what should we do next?”
Giving every app an MCP connector doesn’t settle that question. Someone still has to link the customer records, define which system owns each fact, and distinguish a failed payment from an intentional cancellation. Without those shared definitions, you’ve connected the tools while leaving the interpretation work on your desk.
A stitched-together stack needs a shared understanding of the business. More connectors alone won’t supply it.
MCP supports tool descriptions and structured schemas, so a well-designed implementation can convey meaningful context. But connecting an API doesn’t automatically supply a coherent business model. The platform and its integration have to make those definitions available and apply them consistently.
If you have to explain what every field means every time you ask a question, you’re still doing the semantic layer’s job.
Ask It to Prove It
If a vendor calls itself AI-native, ask for a demonstration with a real operating question.
For a subscription business, try this: “Show me the subscriptions with failed payments this week. Separate them from customers who actually canceled. Break the failures down by reason, and show me which recovered payments led to successful rebills afterward.”
That’s an acceptance test, not a promise that every connected platform can do it today.
Watch what happens next. Does the system understand the records? Can you inspect the underlying transactions? Does it distinguish a fact from an estimate? Can it explain what data it couldn’t get?
Then ask it to propose a change. Before anything happens, you should be able to see the scope, control the permission, and verify the result afterward.
The work has to survive contact with the business. A polished answer that nobody can trace back to an actual customer, order, or transaction hasn’t earned your trust.
All In Means Building Around It
For a software company, going all in on AI means committing product and engineering resources to making the core operation accessible and useful to agents. It means treating incomplete tool coverage, confusing data, and unverifiable actions as product defects.
It also means operators keep control. A reporting agent doesn’t need permission to issue refunds. An agent reviewing subscription performance doesn’t automatically need permission to change subscriptions. Serious software makes those distinctions possible.
RevCent’s MCP exposes more than 250 operations, with access limited through OAuth client permissions. Its documented uses include natural-language reporting, data updates, and agents that monitor account activity. Those are concrete capabilities an operator can evaluate.
RevCent’s AI tools also include background assistants and voice agents, with system tools that let them work inside the account. That’s the direction: AI participating in the operation, with access appropriate to the job.
We should be judged by how well that works. So should every platform asking for a place in your stack.
Raise the Minimum
At your next software demo, ask three questions:
- Can my AI connect to this product and work with current business data?
- Which meaningful tasks can it complete, and which still require me to move data or click through screens?
- How do I limit its access and verify what it did?
Ask the salesperson to show you. Put an actual workflow on the screen. Follow it through to the result.
A vendor can have an excellent answer without promising to automate your entire company. What it can’t do is wave at a chatbot and expect you to confuse that with an operating capability.
The product, the offer, the customer relationship, the judgment about where to spend the next dollar: there’s still plenty to compete on. Making the customer manually shuttle information between paid tools shouldn’t be part of the bargain.
You already expect a login.
Expect your AI to have a way in, too.
Table stakes. Deal the next hand.