Integrating AI into your tools without breaking anything

The most common mistake when adopting AI isn't technical, it's about location: buying a chatbot that lives in a separate tab and hoping the team remembers to visit it. The AI that works is the AI you don't see — the one that lives inside the ERP, the CRM, the email, and the spreadsheet where the work already happens. The goal isn't to add one more tool: it's to get the ones you already have to start working on their own.

The principle: AI goes to the data, not the other way around

When the workflow forces people to copy and paste information over to the AI, the workflow dies within two weeks. Proper integration reverses the direction: the model connects to your systems through their APIs, reads what it needs with scoped permissions, and returns the result to the same place where the team works. If your system has an API or a database — and almost all of them do — it can be integrated.

The three levels of integration

• Level 1 — Assistants over your data: a search tool that answers questions using your documents, manuals, and historical records (technically known as RAG). Low risk, immediate value: company knowledge stops living in the heads of three people.
• Level 2 — Automations across tools: workflows that connect systems — an email comes in, it's classified, the data is extracted, and the record is created in the ERP — with the AI deciding where a person used to be copying by hand.
• Level 3 — Agents with scoped permissions: systems that carry out multi-step tasks on their own. Here the design of permissions and limits stops being a detail: it is the project.

Maturity isn't about jumping to level 3 in the first month. It's about each level working so well that the next one feels inevitable.

The rules for not breaking anything

Integrating AI into systems that generate revenue demands the same discipline as any other change in production, plus an extra layer of caution because the component is probabilistic. Five rules we always apply: test in a staging environment with synthetic data before touching production; give each integration the minimum permissions it needs and not one more; log every AI action so you can audit what it did and why; keep a person in the loop for every irreversible action — payments, deletions, mass sends; and measure the before and after of each workflow, because an automation that isn't measured is just an opinion.

A realistic example

A distributor receives orders by email in different formats: PDF, a photo of the delivery note, loose text. Before, two people typed them into the ERP by hand. After the integration, the email is read automatically, the data is extracted and validated against the catalog, the order is created in the ERP, and only the doubtful cases — around 8% — reach human review. Nobody switched tools: the same old ones learned to talk to each other.

That's the bar for a good integration: a month later nobody remembers how it used to be done, and the system can be explained — and audited — piece by piece.