AI & Product6 min read

“I rarely go to your dashboard anymore.”

Some customers now use software products through an AI assistant instead of the dashboard. For product teams, that makes the MCP server a user interface, and it needs the same design care as one.

We hear this from users more and more. They are not using the product less; some use it more. They ask an AI assistant to find an account, pull its history, update the record, and prepare the next step. The work still runs through the product, but the browser tab never opens.

For a software company this can read as lost engagement. Logins, sessions, and clicks have been the standard health metrics for years, and when they fall the usual conclusion is that something is wrong. With an MCP server in place, that conclusion can be wrong.

After an MCP launch, some customers may stop logging in to the user interface regularly while using the product in more of their work.

§ 01

The text-only objection

A common objection is that few people want to work entirely in text. That is probably true, and it does not change much.

People do not need to prefer text over visual software. They only need to prefer finishing a task where they started it. If someone is planning a project with an AI assistant, leaving the conversation to open another product, find the right page, copy several fields, and come back is slow.

An MCP connection lets the assistant reach the product on the user’s behalf. The user asks for the outcome, and the product supplies the data or performs the approved action. For product teams, that is the reason to care: it removes steps between a request and a result.

§ 02

Usage and logins are separating

Dashboard activity used to be a reasonable proxy for product activity. The two are starting to diverge.

Take a customer who used to log in every morning to review new accounts. Now they ask an assistant for the five accounts that need attention and why. Your product still holds the records, permissions, and business rules behind that answer, and it may handle more requests than before. Counted by page views, the same customer looks less active.

A falling login rate can sit beside rising product use.

Product teams need to track both. Dashboard sessions show how people use the visual interface. MCP activity shows how often the product contributes to work done elsewhere. Neither number is complete on its own.

§ Fig. 01Two entry points, one product

Through the dashboard

CustomerOpens your product
DashboardExplores and configures
Product coreData, permissions, rules

Through an assistant

CustomerStarts in an AI assistant
AI assistantInterprets the request
MCP serverCalls the right tool
Product coreData, permissions, rules
§ 03

MCP is now an expected integration

Anthropic introduced the Model Context Protocol in November 2024 as an open standard for connecting AI assistants to data and software. Within a year it had been adopted across the major AI products and moved under the Agentic AI Foundation, part of the Linux Foundation.

Customers care less about the protocol than about whether their assistant can use your product. That question now sits next to the familiar ones about integrations, exports, and APIs. A missing MCP server means more manual work for the customer; a poorly built one can make a good product look unreliable. It needs to be treated as a product surface, not a side project.

§ 04

Two groups of users

Users are splitting into two groups.

One group will keep using the dashboard. They explore, compare, configure, and check the state of the system, and visual software is better for that work.

The other group will stay inside an AI assistant for most routine tasks, and open the dashboard only to connect an account, change permissions, or look into something unusual.

Both groups need to be served, and they should not be pushed through the same interface. The dashboard should make complex information easy to see and control. The MCP server should make common tasks easy to request and safe to complete. Both should rely on the same permissions and business rules, so a customer gets the same answer whichever way they come in.

§ Fig. 02What each interface is best at

Dashboard

Visual control

  • Explore and compare
  • Configure the system
  • Review history
  • Handle exceptions

MCP server

Task interface

  • Answer a focused question
  • Complete a routine task
  • Work across several products
  • Act from inside the conversation

Both use the same permission model and the same business rules.

§ 05

An MCP server is not a copy of your API

An API is built for developers and usually mirrors how data is stored: accounts, contacts, invoices, events. Users think in terms of tasks instead. They want to know which customers need attention, whether an invoice is unusual, or what changed since yesterday.

If an MCP server simply exposes every API endpoint under a new name, the assistant has to rebuild your product logic from scratch. It may make too many calls, pick the wrong tool, or return an answer that is technically correct and still misses what the user asked.

A better starting point is the set of tasks customers already come to your product for. Those should shape the first version. Tool names should be plain, descriptions should say when a tool is appropriate, and responses should carry enough context for the user to decide what to do next.

This is product writing. A vague tool name is the machine-readable equivalent of a button labeled “Submit,” and an unhelpful error message is still unhelpful when an assistant is the one reading it.

§ 06

People are right to hesitate when an assistant can change business data, send a message, approve a payment, or publish something under their name.

The MCP experience has to make the boundaries clear: what the assistant can see, what it can change, and which actions need approval. Administrators need a log of what happened, and revoking access should be quick. The MCP specification calls for explicit user consent, control over data sharing, and care around tools that take actions. Those are product requirements as much as security ones.

A practical order is to start with actions that are easy to reverse, let users preview work before it is submitted, require confirmation when the consequence is significant, and keep sensitive data out of the conversation when it is not needed.

§ 07

Learning from usage after launch

The first version will be wrong in ways that are hard to predict in planning.

Users will ask for the same thing in different words. The assistant will pick a tool that sounded right and was not. A response will include too much detail and bury the one fact the user needed. Someone will go back to the dashboard because the MCP server stopped one step short of finishing the job.

Those cases should drive the next iterations. Track which requests succeed, which need correcting, and where users leave the assistant. Repeated workarounds usually point to a missing tool. A tool that exists but is rarely chosen may have a naming or description problem rather than a demand problem.

It also helps to ask users to walk through the last task they completed, from the first message or click. Where a task starts is now as informative as which page gets the most visits.

§ 08

What the dashboard is for now

The dashboard is not going away, but its role is changing. It is where people set up the product, manage access, inspect details, handle exceptions, and review what an assistant has done on their behalf. For many customers it becomes a control and review surface rather than the place where routine work happens.

That has roadmap consequences. A team may simplify routine dashboard flows and invest more in permissions, history, and review tools. The MCP server, meanwhile, deserves the attention that used to go to the most-used screens.

Product teams spent years refining the dashboard because that was where users worked. The useful question now is where they actually start their work.

§ Work with us

If you are planning an MCP server for your product, or reviewing one that is already live, we can help with tool design, permissions, and usage tracking.

Let’s talk