Adding AI to business software, safely

AI is now good enough to take real work off a team: answering repeat questions, reading documents and drafting follow-ups. It is also good enough to cause real problems if it is bolted on without care. Here is how we approach it.

Start with the boring, repetitive work

The best first AI projects are narrow and measurable:

• Answering the same customer questions from your own help docs and order data.
• Reading invoices, purchase orders or forms and turning them into structured records.
• Drafting replies, summaries and follow-ups for a person to review and send.
• Answering questions about your data in plain English, like "which customers ordered less this quarter?"

Each of these has a clear before and after: hours spent, response times, errors caught.

Ground every answer in your own data

Most "the AI made it up" stories come from asking a model to answer from memory. Instead, the assistant should look up the answer in your documents or systems, show where it came from, and say plainly when it cannot find it. That one design choice removes most invented answers.

Respect permissions

An AI assistant must never see more than the person using it. If a sales rep cannot open the finance reports, neither can the assistant they are talking to. That sounds obvious, but many AI integrations use one powerful service account, which quietly gives every user access to everything.

What MCP changes

The Model Context Protocol (MCP) is an open standard for connecting AI assistants to tools and data. Instead of pasting data into a prompt, you give the assistant a set of tools ("look up an order", "run this report") and it calls them as needed. Built properly, each tool call carries the identity of the signed-in user and is checked against that user’s permissions, the same way your app checks every other request.

Keep a person in the loop for changes

Reading is low-risk. Changing records is not. Refunds, cancellations, account changes and anything sent to a customer should go through a person, at least until you have months of evidence that the AI gets that task right.

Watch quality and cost after launch

• Log every conversation and tool call so you can review what happened.
• Track the questions it could not answer; they show which docs need writing.
• Set monthly usage limits so the bill is predictable.
• Use business API terms under which your data is not used to train models.

Done this way, AI becomes a dependable part of the software rather than an experiment on the side.

More guides: Custom software or off-the-shelf? How to decide · What to prepare before you hire an app developer