What "AI integration" actually means
The phrase covers everything from a chatbot widget to a rebuilt operations pipeline, which makes it nearly useless as a description. In practice the work is one of four shapes.
- Reading. Something arrives — an email, a form, a document, a call transcript — and the model extracts what matters and puts it where it belongs. This is the most reliable category and where most of the early value sits.
- Writing. Drafting a reply, a summary, a description, a first version of something a person then edits. Reliable when a human approves the output, risky when nobody does.
- Deciding. Routing, prioritising, flagging. Useful with tight guardrails and dangerous without, because a confident wrong decision at scale is worse than no decision.
- Talking. A voice agent or a chatbot holding a conversation with a customer. The highest visibility, the highest risk, and the one people ask for first.
Almost every business I speak to asks for the fourth and would get more value from the first. Extraction is boring, safe, and it removes an hour a day from somebody who is being paid for their judgement.
What it costs
| Factor | Cheap end | Expensive end |
|---|---|---|
| Systems it must reach | One, with a documented API | Several, one of which has no API at all |
| What happens on failure | It asks a human | It must recover and retry correctly on its own |
| How wrong it is allowed to be | A draft somebody approves | It acts unsupervised on live records |
| Where the data lives | Already clean and in one place | Three systems that disagree about the same customer |
| Who tests it | Against real cases you already have | Against cases that have to be invented first |
Source: Cybrial’s own scoping factors, written from the builds. No build-cost figures are published here because there is no primary source for what other people charge — see the sources at the foot of this page.
The running cost surprises people in the good direction. Model usage for a small business workflow is usually tens of pounds a month, not hundreds — the cost is the build, and the build is mostly integration rather than AI. Anyone quoting a large monthly platform fee is charging you for a wrapper.
Everything is built on your own accounts and API keys, documented, and portable. If you want to take it to somebody else you can, and that is a deliberate condition rather than a concession.
What to build first
The pattern that works: pick the job that is done many times a week, follows rules, and has a checkable output. Not the impressive one.
- Enquiries arriving as unstructured text — email, forms, WhatsApp — extracted into your CRM with the fields filled in.
- Documents that have to be read and summarised. Quotes, specs, statements, tenders.
- Product data. Descriptions, attributes and categorisation across a large catalogue, which is a genuine multi-day job done by hand.
- First-draft replies to routine enquiries, queued for a person to approve rather than sent.
- Call transcripts turned into notes and next actions.
Guardrails, and why they are most of the work
A model will always produce an answer. That is the whole problem, and it is why the interesting part of this work is not the prompt.
- Scope. What it may act on and what it must pass to a person, written down before anything is built.
- Grounding. Answers drawn from your actual data rather than the model's memory, so it cannot invent a price or a policy.
- Approval. Anything customer-facing goes through a human until the evaluation says it does not need to.
- Evaluation. A fixed set of real cases, run before launch and after every change, with the results recorded. Without this you are guessing.
- Logging. Every input and output kept, so when something goes wrong you can see what happened rather than speculate.
- A kill switch. Somebody non-technical must be able to turn it off.
If a supplier has not raised most of that before quoting, they are selling a demo. The demo always works; the guardrails are what makes it survive contact with real customers.
