What "AI integration" actually means for a small business
Strip away the marketing and AI integration comes down to a handful of practical patterns. Here is what they are, what they cost, and where they go wrong.
"We should be using AI" has become a board-level sentence. It rarely comes with a definition. Having built a number of these systems, we find that nearly every useful AI integration in a small or mid-sized business falls into one of five patterns. Knowing them makes the conversation concrete.
Pattern 1: Reading things so people do not have to
Invoices, application forms, contracts, emails, support tickets. A language model can read an unstructured document and produce structured data (the vendor name, the amount, the due date, the category, the sentiment) with accuracy that is now good enough for most operational use when paired with validation rules.
Where it goes wrong: treating the output as certain. Every extraction needs a confidence check and an exception queue for a human to review.
Pattern 2: Answering questions from your own documents
Staff asking "what is our policy on X?", customers asking "does this product do Y?", engineers asking "how does this system handle Z?". Retrieval-augmented generation (RAG) pulls the relevant passages from your documents and has the model answer with citations.
Where it goes wrong: poor document preparation. If the source material is contradictory, outdated or badly structured, the answers will be too. Most of the work is in the documents, not the model.
Pattern 3: Drafting for human approval
Replies to enquiries, first drafts of proposals, summaries of meetings, product descriptions. The model produces a draft; a person edits and sends. This is the safest pattern and often the highest return, because it removes the blank-page problem without removing judgement.
Where it goes wrong: removing the human step to save time. The draft quality is good; the judgement about what to send is still yours.
Pattern 4: Routing and prioritising
Incoming requests classified by type, urgency and the right team. Leads scored on fit. Tickets tagged before anyone opens them. Small models do this cheaply and quickly.
Where it goes wrong: no feedback loop. If mis-routed items are not corrected in a way the system learns from, accuracy drifts.
Pattern 5: Agents that take bounded actions
The newest pattern and the one with the most hype. An agent can look up a customer record, draft a response, schedule a follow-up and update the CRM: a multi-step task rather than a single answer.
Where it goes wrong: unbounded scope and missing audit trails. Agents should have a defined set of tools, a clear stopping condition, approval gates for consequential actions, and a log of everything they did.
What it costs
Less than most people expect for the model calls, and more than they expect for the surrounding engineering: data preparation, integration with existing systems, evaluation sets, monitoring and the human-review interfaces. A realistic first integration is a few weeks of focused work, not a year and not an afternoon.
Where to start
Pick the pattern where a mistake is cheap and the time saved is obvious. Drafting for approval and document reading are usually the right first projects. Measure the result (hours saved, errors caught, response times) and let that fund the next one.