Salesforce just shipped sixty-plus MCP tools so AI agents can use it. HubSpot did the same a few months earlier. Every major SaaS company is doing some version of this, wrapping their existing APIs in the protocol that lets agents talk to them, and announcing they support the AI era.
They don’t. Wrapping a twenty-five-year-old API for human developers in a new protocol doesn’t make it good for AI agents to use. The agent still has to do work the API was never designed for it to do, at a cost that’s about to start mattering. Every category of enterprise software is going to get rebuilt, and the founders rebuilding it have an advantage the incumbents can’t match.
The user of software is changing
For forty years, business software was used by humans clicking through screens. Every database, every API, every workflow assumed a person at a keyboard.
That’s breaking. The next generation of users won’t click. They’ll talk to an AI assistant, and the assistant will do the actual work. Read the email. Update the CRM. Score the candidates. Watch the pipeline. The assistant won’t even wait to be asked half the time. It’ll run agents in the background that join meetings, watch inboxes, and take action without anyone in the loop for any single step. When that’s how people work, the software underneath becomes the product. And almost no enterprise software was built for that.
What wrapped APIs mean in practice
Take a simple task. “Log this call I just had with Northpeak, apply the discount they asked for, schedule a follow-up Tuesday.”
On a typical CRM with a wrapped API, that’s nine separate calls. Find the account. Find the deal. Log the activity. Update the amount. Update the timestamp. Create a task. Check the calendar. Create the event. Send the invite. The agent reasons through each step, picks which endpoint to hit, parses each response, handles each error. Every step costs tokens. Every step is a place the agent might get something wrong.
We’ve built MCP wrappers for clients across several enterprise categories, and this pattern shows up consistently. We ended up going past the basic wrapper for most of them, adding intent layers and feedback loops that learned from how agents actually used the systems. The basic wrappers worked, but they were brittle in production, and the cost got noticed quickly by the people paying the AI bills.
When two SaaS products both support agents, the one that costs less per interaction wins. If your agent burns a million tokens on Vendor A’s CRM but does the same work in a hundred thousand tokens on Vendor B’s agent-native CRM, you’ll see it in your AI bill at the end of the month. Customers won’t argue about which architecture is better. They’ll look at their compute costs and switch.
What rebuilding looks like
Software designed for agents looks different. Three things matter.
The first is intent-based interfaces. Instead of nine separate calls, one call. “Log this call, apply this discount, schedule this follow-up.” The agent says what it wants, not how to do it. This means rebuilding the software’s logic so the workflows live inside the platform, not in the agent’s prompt.
The difference isn’t just where the workflow runs. It’s what’s doing the work. When an agent calls a wrapped API, the agent’s own LLM has to do the thinking on every step. It reads the schema, picks the endpoint, parses the response, decides what to do next. Each step is paid for in tokens. When that same agent calls an intent on an agent-native platform, the workflow runs as plain code on the server side. The agent’s LLM doesn’t reason through nine calls because it’s making one call to a workflow that already knows what to do. The cost difference isn’t a percentage. It’s a hundred times or more.
The second is software that learns from how agents use it. When agents keep stringing together the same handful of calls in the same order, that’s a signal. The software should notice and turn it into a single call. You can’t bolt this on to something that already exists.
The third is audit and control built for agents, not humans. When agents act on someone’s behalf, every action needs to be traced. Who did this? On whose authority? Was it allowed? Can it be undone? Agent-native software treats every action as a logged event with full context, with guards on anything destructive and rules that can stop an agent from going off the rails. Without this, no enterprise will let agents touch anything that matters.
None of these can be added after the fact. The incumbent CRMs can’t rebuild their object models without breaking decades of customer integrations. The incumbent project tools can’t redesign their workflow engines without disrupting every automation customers have built. The cost of rebuilding for them is paid in customer churn. The cost for a new entrant is zero.
Where the opportunity lives
Look for categories where the existing player is large, expensive, and bloated for what most customers actually use. CRM, HRMS, helpdesk, contracts, project management. Most teams use ten percent of the features and pay for one hundred percent. A lighter, agent-native version that costs less and works better with the AI tools customers are already adopting has natural pull.
Look for categories where the data model is well-understood enough that you can ship something useful in months rather than years. Invoicing, expenses, inventory, knowledge bases, forms. Not glamorous, but the same shape across most companies.
The opportunity isn’t to compete with the incumbents on their terms. They’ll keep the customers who’ve already committed. The opportunity is the customers who haven’t, who’ve outgrown their current tool, or who notice their AI agents work better and cost less against your software.
If you’re thinking about what to build, the question isn’t which AI tool to ship next. It’s which category of software you’d rebuild from scratch, knowing the user is shifting from a human at a keyboard to an agent acting on someone’s behalf. The incumbents can’t follow you there. Their architecture won’t let them.
Share your exclusive thoughts to:
jeevika@thefoundermedia.in
