Adding AI to existing business software, safely

Start with the task, not the model

Look for work that is repetitive, high-volume and text-heavy: answering the same customer questions, summarising records, drafting replies, or finding information spread across documents. If a task happens a few times a month, automation rarely pays off.

Ground answers in your own data

A general model does not know your policies, prices or customers. Retrieval-augmented generation (RAG) fixes that: the system first finds the relevant records or documents, then asks the model to answer only from them, and shows the sources it used.

Respect the permissions you already have

An assistant must never become a way around your access rules. Run every lookup and action through the application’s existing permission checks, as the signed-in user, so people can only see and change what they already could.

• Filter retrieved documents by the user’s permissions before they reach the model.
• Let the assistant suggest actions, and require confirmation for anything that changes data.
• Log what the assistant was asked, what it used and what it answered.

Decide what data may leave your systems

Agree up front which data can be sent to an AI provider, and under which terms. Sensitive fields can often be removed or masked before a request, without hurting answer quality.

Test with real questions before launch

Collect a set of real questions with known good answers, and measure the assistant against them before release and after every change. Keep reviewing a sample of live answers once it is in use. That is how you catch wrong or invented answers early, instead of hearing about them from customers.