MCPs get us halfway there

There's been a deluge of MCP connectors over the last few months. Your DMS has one. Your contract tool has one. Your research platform has one. (And, yes, we have one too.)

At first, it looks like a distribution play — be where the customers are. If lawyers are already working in Claude, Harvey or whatever comes next, plug your product in through an MCP. The promise that that there is no new interface to learn, and no new workflow to adopt.

Distribution is certainly one reason, but there is something more fundamental happening.

The software industry has to answer the much bigger question: what is left of your software after removing its user interface?

The best legal tech ever invented

Something Jason Barnwell said on our podcast two years ago has stuck with me ever since.

"What's the best legal tech ever invented?"

He answered his own question, "Secretaries."

Secretaries get the job done.

You tell secretaries what you need, they work out how to get it done, and they come back with the result. AI assistants have taken that philosophy and run with it. Users love typing (or dictating) what they want and having the work done for them.

For an increasing amount of business work, the interaction is shifting from “how do I use this software?” to “can you just do this for me?”

That is a pretty fundamental change in user experience. It also creates a problem for software vendors. If users want to work through an assistant, every time a vendor sends them away to use a different application, it creates friction.

So, vendors are responding with MCPs.

A to K: halfway through the alphabet

Before Gen AI, software vendors tried to build products that took users “from A to Z”. The pitch was end-to-end functionality — bring the whole workflow into one application, and you can get the job done.

Now, with Gen AI, users can get “from A to K” with a single prompt. The assistant can handle a substantial part of the work without the user ever opening the traditional applications that used to perform those tasks. The remaining steps (the “L to Z”) still need to happen, and that is where MCPs come in. They connect the existing “A to Z” solutions to the “A to K” capabilities of AI.

The idea is that AI and traditional software will work together, joining up the alphabet. Except, they don't. Not yet, anyway.

We still feel the janky edges between the chat interface and the underlying applications. The AI assistant can call a tool, retrieve some information, and trigger an action. But that doesn't mean it understands the whole workflow, or knows how to complete it properly.

We have made this mistake before. We took broken processes and laid software over the top. We ended up doing the same wrong things, just faster.

Thin MCP wrappers risk repeating exactly that mistake. They fit existing software into a new user experience without fundamentally rethinking the work.

Imagine hiring a highly competent secretary, giving them access to a bunch of specialized applications and saying: "Figure it out."

The secretary might be brilliant. But they don't automatically know your processes, your preferences, the exceptions, the dependencies, or the things that must never be missed. Much of that knowledge has traditionally lived inside the user's head. Connecting the secretary to the specialized applications doesn't transfer that knowledge.

That's the problem with simply wrapping existing software in an MCP. You get something that looks integrated, but the underlying work has not been redesigned. There are invisible gaps, awkward handoffs, and outputs that need to be verified every time.

Note: Last month, we wrote there the only two things left to build in software: skills and data. MCPs add another dimension. They make it easier for an AI agent to access existing software, but access is not capability. A connector can expose a function; it cannot, by itself, give an agent the expertise to use it, the context to know when to use it, or the rules to know the boundaries.

Going headless for AI agents

The answer is not to bolt more connectors onto existing products. We need to rethink what the product actually needs to do, and who (or what) it needs to do it for.

Traditionally, software was built around a human user. Users navigated the application, supplied the information, made decisions, and moved through the workflow. In an AI-driven world, the AI assistants are increasingly doing those things.

For the AI to work, it means software has to give the agent the data and context to understand the task, the skills and instructions to perform it, and the rules and constraints that must be respected. Plus, even with all those rules and skills, AI agents are not deterministic. Even with all those skills, data and instructions, AI agents are still probabilistic. They might approach the same task differently, interpret something inconsistently, or produce an answer that sounds right but isn't. These are not things an MCP connector will fix. They are characteristics we have to design around.

The agents should exercise flexibility and judgment where useful, and the tools the agents use should provide certainty where essential. Most of us who build products already know this.

So, why isn’t everyone rebuilding?

Rebuilding is hard for software vendors for two reasons.

The first is technical. We have to redesign the product, not just expose a few endpoints. We have to separate the underlying functionality from the interface, structure the data for machine consumption, and make sure the agent can actually complete the work.

The second is bigger, and that is commercial. We have to be willing to take a subservient role to the AI platforms. The task still gets done, but most of the interaction with the customer may no longer happen through the existing product. The AI assistant owns the experience.

This changes the perceived value. The underlying technology might be just as important as ever, but the user no longer sees it doing the work. The AI assistant gets the credit for the outcome, even when the underlying software made it possible.

This creates a problem for vendors. If visibility is separated from usefulness, vendors may struggle to capture the value their software creates.

  • AI is changing how people want to work.

  • AI is changing who does the work.

  • AI is changing how much of the value existing software vendors can capture.


A place for specialized UI

Not everything belongs behind an AI assistant.

Some work needs specialist software, e.g. you don't want an assistant drawing the blueprint for a building, you don’t want video editing and complex financial modelling to depend on vibes. The same goes for serious data infrastructure. A court-filing system tracking deadlines across thousands of matters is probably not something you improvise in a chat window.

There will always be work that benefits from a specialized interface, particularly where people need precise controls, visual feedback, or direct oversight.

The questions for every software vendor are these:

  • which parts of the product must live in its specialized UI?

  • which parts should be exposed to AI?

  • what is the price that users are willing to pay?

Next
Next

Solving intake cuts 27% off law firm revenue