MenuStart a projectStart

AI & data

Ask your business data in plain English with MCP

This automation lets someone ask a question like “how many bookings do we have next week?” and get the figure straight from your own data. Claude answers by calling small, read-only tools on an MCP server, then shows which tool and date range it used.

An example of how the studio would build it. No live services are connected.

From a plain-English question to a figure you can check.

  1. Step 01

    Question asked

    Someone asks a plain-English question, such as how many bookings there are next week.

  2. Step 02

    Access checked

    The request reaches your MCP server, which confirms who is asking and which tools their role may use.

  3. Step 03

    Tool chosen

    Claude picks the matching read-only tool, such as one that counts bookings, and fills in next week’s dates.

  4. Step 04

    Query run

    The server validates the inputs and runs a fixed, scoped query through a read-only database account.

  5. Step 05

    Answer given

    Claude answers with the figure, the date range used and the tool it came from, so anyone can check it.

The result

Answers without hunting through dashboards.

Every tool is read-only and narrow, and each person sees only what their role allows. The assistant can’t write its own queries or change records.

Before we connect it.

Where this is useful

Useful when the same questions keep reaching the one person who knows where to look, or when dashboards exist but nobody opens them. A sensible first version offers three or four tools for questions that come up every week, such as bookings, revenue and open quotes.

What we need to connect

  1. 01Read-only access to your database, or to the API of the system that holds the data.
  2. 02A short list of real questions people ask, with the agreed definition behind each answer.
  3. 03Who should have access, what each role may see and where the assistant will be used.

What needs careful handling

The safety comes from the tools, not the prompt: each one runs a fixed, parameterised query through a read-only account and returns only the fields and rows it needs. Personal data stays out unless a role genuinely needs it. When no tool fits a question, the assistant should say so rather than estimate, and every call is logged so an answer can be traced to its source.

Questions people ask.

What is MCP, in plain terms?

MCP, the Model Context Protocol, is an open standard for connecting AI assistants to tools and data. An MCP server offers a few named tools with clear inputs, and an assistant such as Claude calls them to fetch real figures instead of guessing.

Can the assistant change or delete our data?

Not in this design. The tools only read, the database account behind them has read-only permissions, and there is no tool for free-form queries. Anything that changes data would be a separate, approved tool with its own checks.

Monday arrives → Posted to Slack
Automation 89

Weekly KPI digest in Slack

Email received → Person sends it
Automation 64

AI support inbox triage and reply drafts

All automation examples