AI integration
AI integration is the engineering work of connecting an AI model to the software and data an organisation already runs, so it can act on real, current information rather than on text typed into a prompt. It covers retrieval, permissions, output validation and monitoring.
Also known as: AI implementation, LLM integration, AI enablement
Last reviewed
Why AI integration matters
A model on its own knows nothing about your business. It is a capable stranger, and everything that makes it useful in your context — your customers, your orders, your documents, your rules — has to be supplied.
This is why the value of AI is rarely the model and almost always the integration. The gap between an impressive demo and something usable in production is not model quality; it is whether the system can reach current data, act within permissions, and be checked when it is wrong. Most AI projects turn out to be system integration projects with a model at the end.
How AI integration works
Four pieces, in roughly this order of importance. Retrieval: fetching the right records and putting them in front of the model, which decides more outcomes than model choice does. Permissions: enforcing in your own code what data this user's request may reach, because a model given broad read access will summarise things the user was never entitled to see. Structured output: requiring a fixed shape your code can validate, so free text never flows into a database field. And monitoring, because accuracy drifts as your data changes and providers update models underneath you.
Increasingly the connection to business systems is made through MCP connectors rather than bespoke plumbing per system.
AI integration vs system integration
System integration makes your existing systems exchange data with each other. AI integration connects a model to those systems so it can read and act on what they hold.
The dependency runs one way and it is worth being clear about: AI integration generally presupposes system integration. If your CRM and job system do not currently talk to each other, adding a model does not join them up — it just gives you something that can only see one of them. That is why an AI-ready data ecosystem is frequently the honest first project rather than the AI feature itself.
When you need it
Whenever an AI feature has to know something specific about your business, which is nearly always. A model answering general questions needs no integration; a model answering "where is my order" needs all four pieces above.
Before commissioning any of it, check one thing: does the authoritative answer to the question you want answered actually exist, correctly, in a system that can be queried? If not, that is the project. On our AI Opportunity Planning engagements a focused scoping engagement starts around £1,500, and establishing whether the data is genuinely reachable is usually what moves the estimate most.
AI integration: common questions
Four things: retrieval, getting the right records in front of the model; permissions, enforced in your own code rather than requested in a prompt; structured output your code can validate; and monitoring, because accuracy drifts as data changes and providers update models. The model itself is the smallest part.
No, and one depends on the other. System integration makes your existing systems exchange data. AI integration connects a model to those systems. If your systems do not currently talk to each other, a model will not join them up — it will only be able to see one of them.
Not model quality, but context. Either the model was given the wrong records, or it was given data that is out of date or incorrect. Getting the right three paragraphs in front of a mid-range model beats the best available model reading the wrong document, and costs less per call.
Want to talk about your project?
Tell us what you’re trying to achieve and we’ll map the fastest credible path.