Most businesses asking how to "use AI" already have the systems they need: an ERP or accounting package, a CRM, Microsoft 365 or Google Workspace, perhaps a helpdesk. The question is not which AI platform to buy, but where AI can safely do useful work between those systems.
Below are five patterns we use repeatedly, what each is good for, and the controls that make it safe in production.
Pattern 1: Read and structure
What it does: takes unstructured input, such as an email, a PDF or a call note, and returns defined fields.
Good for: order intake, invoice capture, lead details from inquiry emails, contract key dates.
Controls: a fixed output schema, validation of every field against your data, and a confidence threshold below which the item goes to a person. This is the most reliable pattern and the best place to start. Our guide to how AI document processing works covers it in detail.
Pattern 2: Classify and route
What it does: decides which category an item belongs in, and therefore where it goes.
Good for: shared inbox triage, support ticket routing, expense categorisation, maintenance request urgency.
Controls: a clearly defined, mutually exclusive set of categories, an "unsure" category that routes to a person, and periodic sampling to check accuracy. Classification errors are usually cheap because a person sees the item next anyway, which makes this pattern suitable for running unattended early on.
Pattern 3: Draft for approval
What it does: writes a first version of something a person will review: a customer reply, a quote cover note, a summary of a case.
Good for: routine customer questions, status updates, internal handover notes.
Controls: drafts grounded in your data (the actual order status, the actual policy), never in the model's general knowledge; a person sends; and tracking of how often drafts are edited heavily, which tells you where the drafts are weak.
Pattern 4: Search and answer
What it does: answers questions from your own documents and data, citing where the answer came from.
Good for: staff questions about procedures, product specifications, contract terms and customer history.
Controls: answers only from approved sources, a visible citation for every answer, and an explicit "I could not find this" when sources are silent. Without these, the system will produce confident answers that are simply wrong.
Pattern 5: Bounded agent
What it does: completes a multi-step task by choosing which tools to use, such as looking up an order, checking tracking and drafting a reply.
Good for: tasks where the path depends on what is found: exception investigation, document chasing, order status inquiries.
Controls: a narrow job definition, read-only access wherever possible, approval gates for anything that changes money or commitments, a complete action log and testing against real past cases. See AI agent development for how we scope these.
Patterns to avoid
- AI doing arithmetic that matters. Let the model extract numbers; let ordinary code add them up and compare them.
- AI as the system of record. The ERP or CRM remains the source of truth. AI reads from it and proposes changes to it.
- Unbounded chat on top of everything. A general assistant with write access to several systems is hard to test and hard to trust. Start with specific jobs.
- No evaluation set. If you cannot re-run a set of known examples when the prompt or model changes, you cannot tell whether a change made things better or worse.
How the integration is actually wired
In practice, AI steps sit inside ordinary workflows. A workflow platform such as n8n, Make or Power Automate, or a small custom service, receives the trigger (a new email, a new file), calls the AI model with a defined prompt and schema, validates the response, and then calls the APIs of your existing systems. The AI is one step among many. That structure is what makes it testable, observable and replaceable when better models arrive.
How to keep AI steps reliable
AI integrations fail quietly when nobody is checking them. A few habits keep them dependable:
- Validate every output. Check extracted values against your data, such as known vendors, valid dates and totals that add up, before anything is written to a system.
- Use confidence thresholds. Low-confidence results go to a person rather than being guessed.
- Keep an evaluation set. A few hundred known examples, re-run whenever a prompt or model changes.
- Log everything. Inputs, outputs and decisions, so any result can be explained later.
- Monitor accuracy over time. Suppliers change layouts and customers change habits; accuracy that is not measured drifts.
Data and privacy
Before connecting AI to business systems, agree where data is processed and stored, whether providers may retain or train on it, and who can see logs. Business and enterprise AI services generally offer terms that prohibit training on your data; consumer tools often do not. Our security page explains how we handle this.
Where to start
Pick one high-volume task where the input is unstructured and the output is well defined. Read-and-structure or classify-and-route projects usually pay back fastest and carry the least risk. Once that is running reliably, the same infrastructure supports drafting, search and eventually bounded agents.
If you would like help choosing the first one, our AI integration service starts with exactly that conversation.