In October, on a demo company, an AI assistant connected to our new MCP server reported a client with $10.7k spent in the last 30 days. That client hadn't bought anything since March. Our own screens showed the same figure.
The assistant was an experiment. I'd built the MCP server for the platform I work on to test whether an AI agent could handle our admins' tasks, and whether that would be better than using the admin itself. Nobody had asked for it explicitly. An admin connects Claude or ChatGPT to their company's account and asks about their clients, stores and reports, and with changes switched on, the assistant can act on some of it too. For now only our own team uses it, mostly our customer success team, and it's way better than the admin pages.
The screen every customer wants
Every customer of a B2B product wants something slightly different from the admin. Say one company wants spend by store for its top clients, and another follows a KPI nobody else tracks. Each request is reasonable, and done the usual way, each becomes a filter, a column or a new report that everyone else scrolls past.
The MCP takes another road. It exposes building blocks: find clients, count and group them, read a report, set a goal, draft a task list. The agent combines them into whatever the person asked for, so the question stops being whether the admin can show something and becomes whether we have the blocks for it. That's a better answer than adding what every customer needs to the admin.
What the agent made us decide
Building the blocks turned into a product review. A screen hides a lot of decisions: a button the wrong role shouldn't press simply isn't there. An agent doesn't use your menus, so those decisions had to be made out loud. Who can see what, and what does each number actually mean?
Answering them improved the product too, and the $10.7k was the clearest case. Our per-client totals were only recalculated when a client bought something, so they froze at the last purchase, on our screens as much as in the assistant's answer. Seeing that number in an assistant's answer is how we found it. A nightly job keeps those totals current now, on the screens too.
The other big decision was about changes. A person stays in charge of them: they're off until we switch them on for a company and an admin allows them when connecting. A task list the assistant builds lands as a draft that a person reviews and shares with the team.
Hearing what nobody could do
One block doesn't touch the customer's data at all. When the assistant can't do what someone asked, it can offer to tell our product team what they wanted, which tools it tried and why those fell short. Our logs only show what people did, never what they wanted and couldn't do.
I tried the same idea on a side project, espirita.club. I had an agent act as the platform's operator and build a fictional center's site almost entirely through its MCP. It hit eleven gaps, and they became that morning's commits.
Our team felt it first
Our customer success team uses the admin server, and onboarding a customer is faster and better for it.
Myself, I mostly use a second one: a server only our staff can connect to, where any change to a real company waits for a person to approve it. With it, debugging a specific case, which could take hours or days, now takes about five minutes.
Security came along
Deciding what an agent may see and do also meant looking hard at what the app already allowed. That review turned up holes that had been in the code for years, the oldest from 2019, one of them from an upgrade I did myself last year. The MCP's own new code got a review of its own.
Isn't a screen better?
The strongest objection is that screens are predictable. Everyone sees the same report, it's quick to scan, and nobody has to phrase a question well to get an answer. A chat answer depends on how you ask, and a model can be wrong.
That's fair, and two decisions answer most of it. The report tools reuse the code behind our dashboards, so the assistant and the screen can't disagree about a number. And when a search only partly matches what someone asked, the assistant says what it couldn't match instead of passing off a partial count as the whole answer.
Where it goes
We can't wait to put it in our clients' hands, especially for creating tasks and following each company's custom KPIs.
This first server is a base of building blocks, not a finished assistant. Long term, we see specialized agents, each handling one part of the whole, built on top of it.
If you have a backlog of "can the admin also show..." requests, try giving an agent a few blocks before you build the next screen, and give it a way to tell you what it couldn't do.