The ideas behind our AI case study — models, prompts, MCP, search and evaluation — in plain English, no computer science required. The jargon table comes first; the pictures fill it in.
A large language model is a machine trained on an enormous amount of text until it becomes very good at one thing: predicting what text should come next. Do that at sufficient scale and something remarkable falls out — the ability to read, write, summarise, translate and reason in ordinary language.
Out of the box, a model knows only what it read during training — nothing about your business, and nothing after its training date. Two things close that gap. The prompt is the program: written instructions that set the model's role, rules, tone and boundaries — programming in English rather than code, and the single biggest lever on quality. Grounding is the evidence: handing the model your documents and live data to answer from, with citations, instead of letting it answer from memory.
One more essential: tokens. Models read and write in word-pieces called tokens, and every answer is metered in them — cost and speed both scale with token count. That meter is why the next section matters commercially, not just technically.
Fast chat models answer in a beat and cost pennies per answer. Left to their own devices they are capable but shallow — with tight instructions and good grounding, they are excellent. Reasoning models think before they answer — working step by step through hard, open-ended questions over raw sources. They give superb results, but each answer burns many times the tokens and arrives much more slowly. (Under the bonnet, frontier models use a mixture of experts: many specialist sub-networks with a router picking the right few for each step — how they scale capability without scaling every cost.)
The commercial question in every AI deployment is therefore: do you pay the premium model's price on every answer, or engineer the fast model up to the premium result? Our answer is the evaluation harness in the case study — measure the fast model against real questions, tune its instructions, and repeat until it scores alongside the expensive one. Then keep measuring so it stays there.
MCP — the Model Context Protocol — is the open standard for connecting AI to business systems. Think of it as a universal socket: each system's capabilities are wrapped as tools once, and then any AI that speaks the standard can use them — Microsoft Copilot, Claude, a website agent, and whatever comes next. Every major AI vendor has adopted it.
That is why it is the way forward: no lock-in to one AI vendor, no rebuild when a better model ships, and one place to enforce who can use which tool — the public website agent simply never has the customer-data tools to call.
Keyword search matches the letters you typed. If the document says "preventing timber warping" and you searched "stop deck boards cupping", you get nothing — same idea, different words. Semantic (vector) search turns text into numbers that capture meaning, so your question lands near the right passage even when they share no words at all. Production systems run hybrid search — meaning first, exact keywords as backup, so part codes and product names still match precisely.
Strip any AI application back and you find the same shape: anagent. The engine inside is a model — rented off the shelf, the same raw ability everyone else has. What makes it your agent is built: the instructions it follows, and the connections into your data. And because those parts are built, they can be tested — which is exactly what the tests below validate.
Which means the pattern travels far beyond chat. Wherever skilled people turn an input into a draft or a judgement, the business already holds the answer key: work it has done before, and how that work turned out. So the test is the same every time — have the AI redo work the business has already done, and compare the results side by side with what actually happened. Five examples, function by function; every diagram is deliberately the same shape:
A customer emails a question. The AI reads it, gathers the relevant facts — the order, the product page, the returns policy — and a draft reply appears for your service agent to send, edit or escalate. The answer key already exists: years of resolved tickets sit in the helpdesk, each one a real question paired with the reply that actually fixed it.
A supplier invoice arrives as a PDF. The AI reads it, matches it to the supplier and the purchase order, and proposes the posting — amounts, cost codes, VAT — for finance to approve with one click, or correct. The answer key is the ERP itself: thousands of past invoices, each paired with exactly how your finance team coded and approved it.
An enquiry comes in. The AI reads what the customer actually needs, prices it from the live price list and the catalogue, and drafts the quote for the rep to check, personalise and send — the same hour, not next week. The answer key is already in the sales system: every quote the business has ever sent, and, crucially, which of them won.
A contract lands — forty pages, needed by Friday. The AI reads every clause against your playbook — the terms you accept, the ones you never do — and flags what needs a lawyer's eye, with the reason attached. The answer key is the contract archive: what was flagged, negotiated and finally signed on every deal before this one.
A complaint comes in. The AI gathers the whole picture — the order, the delivery record, every earlier conversation — drafts a considered response, and tags the likely root cause so patterns surface across months, not anecdotes. The answer key: the resolved complaints the business already keeps, together with the outcomes that satisfied the customer.
Everything on this page is at work in a real deployment — the tool layer, the search, and the evaluation harness that makes the economics work.