AI where it pays for itself.
Nowhere it doesn't.

An AI agent is software given one defined job, the tools to do it, and the ability to decide the steps itself. A chatbot answers a message; an agent might read the enquiry, check availability, ask the two qualifying questions, book the slot and write the summary into your CRM. The useful distinction is that an agent completes a task rather than producing text.

Retrieval-augmented generation means the AI answers from your own documents — price lists, policies, product specs, past conversations — rather than from what a general model happens to remember. It is what stops a chatbot confidently inventing a delivery time. For any business where a wrong answer costs money, RAG is not optional.

It varies with scope, so the honest answer is that we quote after the audit. What determines it is how many systems have to be connected, whether your data is clean enough to ground answers on, and how much human approval the flow needs. A single well-scoped agent is a small project; a connected system across sales and operations is not.

In small businesses, almost never — it removes the parts of their day nobody wanted. The realistic outcome is that the same team handles more enquiries, faster, with fewer things falling through. If someone is selling you headcount reduction at your size, they are selling the fantasy rather than the tool.

Systems are built inside your accounts, with your keys, and your data stays exportable. We are explicit about what is sent to a model provider and what never leaves your systems, and we will design around anything you are not comfortable sending. If that constraint makes a feature impossible, we say so rather than quietly doing it anyway.

We run our own. WebFlip Hub is an internal system we wrote — agents, agent memory, a skills layer, lead and campaign modules, and usage metering — and it is what we use to run this business day to day. Screenshots are on this page. It is also why we are comfortable telling a client when an off-the-shelf tool is genuinely good enough for them.

Whichever suits the job and the budget, rather than one vendor we are tied to. Model choice is driven by the task — reasoning quality, speed, cost per call and where the data is allowed to go. Being platform-agnostic matters here more than in most software, because the sensible choice changes every few months.

Every build ships with numbers attached: how often it handled a case without escalating, how often a human corrected it, what it cost to run, and what it removed. An agent that quietly escalates everything looks fine in a demo and is worthless in production — the difference only shows up if you are measuring.