We built an AI solution for a multi-branch building materials retailer — a tool layer that multiple AI agents plug into — so staff can ask plain-English questions and get grounded answers, pulled live from the ERP and searched from indexed company documents, supplier records, training content, and product knowledge. Here's how it works, what it does for each part of the business, and how we prove it keeps getting better.
Like most established retailers, our client's knowledge lived in five different places: an ERP holding customers, orders, stock and pricing; a vault of supplier records and documents; a SharePoint estate of policies, procedures and working documents; a learning platform full of training courses; and a public website with the product catalogue. Every question a counter person, buyer or manager asked meant knowing which system to open, where to look inside it, and often a phone call to whoever knew it best.
The fix wasn't another system. It was a foundation that lets any AI agent know all of them.
Not all business data should be treated the same. Stock levels change by the minute — they must be fetched live, never cached. Supplier documents and training content change slowly — they're indexed for meaning, so an agent can search them like a well-read colleague.
Then comes the part that makes the whole thing repeatable: every capability — live query or knowledge search — is exposed as a tool through MCP (Model Context Protocol), the open standard for connecting AI agents to systems. Each system is wrapped as MCP tools once. After that, any agent front end plugs into the same socket: the staff assistant, the Webex customer agent, IT automation — and whatever comes next. Access is scoped at the tool layer, so the public-facing agent simply never has the customer-data tools to call.
And access is identity-first. The staff assistant is surfaced as a bot in Microsoft Teams — staff use it where they already work, signed in with their existing Microsoft 365 account, no new logins and no shared credentials. Every question is asked as a named person, answers are scoped to that person's permissions, and every access is logged.
HOW: the agent calls the ERP's APIs at the moment of asking. Nothing is copied or cached, so the answer is as current as the till.
HOW: content is broken into chunks, embedded, and stored in a search index. Queries run hybrid semantic + keyword search, so "something to stop deck boards cupping" finds the right answer even when no page uses those words.
One foundation, many jobs. Each capability below is live in the client's business today — the questions are the kind people actually ask.
Look up a customer, see what they usually buy, check live stock across every branch and current pricing — without leaving the conversation at the counter.
Search the supplier directory, pull contact details, and ask questions of supplier documents — terms, lead times, product specs — instead of digging through folders.
Pull full detail on any order by its number, check what's on a customer's open jobs, and see stock positions branch by branch in seconds.
Find the right course by topic, then ask questions of the course content itself — turning a library people forget about into answers people get.
Total sales by customer, recent notes and history, and buying patterns — the context a manager needs before a call, gathered in one ask.
Natural-language search over the full catalogue and website guides, so new starters answer product questions like twenty-year veterans.
Department heads connect a high-end reasoning model — like Claude — to the very same tools, and ask the broad, analytical questions no single lookup answers: what the training catalogue actually covers, where the gaps are, and how it lines up against what the business sells.
Because every system is already exposed through the MCP tool layer, a new agent front end is an addition, not a new project — it plugs into the same tools with its own scope. Two are shipping now: one facing the customer, one facing the server room.
A Webex AI Agent on the public website, so anonymous visitors can ask product questions, get how-to guidance drawn from the blog and guides, and check practical facts like opening times and whether an item is in stock at their branch. Because it lives on the Webex platform, it hooks straight into the call centre — when a question needs a human, the conversation hands over to a real agent instead of hitting a dead end.
SCOPED: the public agent sees public knowledge and safe live facts only — never customer accounts, orders, or pricing agreements.
Staff with the right permissions can ask the agent to carry out routine, privileged IT tasks — the kind that otherwise wait for Monday or trigger a weekend call-out, when IT isn't on site but the branches are trading.
GATED: every task is permission-checked against who's asking, and limited to a defined set of approved actions.
For staff, it's a chat in Microsoft Teams — the window that's already open all day. Plain-English questions, cited answers, no new app and no new login: you're signed in as yourself, and the agent shows only what you're allowed to see.
For customers, it's a chat on the website — product questions, how-to guidance, opening hours, stock — handed to a real person in the call centre when needed.
Behind both sits the control room: one dashboard that shows how the agent is doing — and gives you the levers to tune it. Evaluation runs and head-to-head comparisons track answer quality release by release; golden-set mining turns real conversations into new test cases; index health shows every source is fresh; and an MCP explorer invokes the exact tools the agents use, no agent in between.
Most AI projects stall at "it seems to work." We built our client an evaluation pipeline — the same discipline a sawmill applies to grading timber, applied to the agent's answers. Every change to the agent is tested against real questions before it goes live.
The agent's actual usage is mined for the questions staff genuinely ask — including the ones it fumbled. These become candidate test cases.
Each candidate is grounded in evidence from the source systems, given a verified correct answer, and promoted into a growing library of test cases. Junk and duplicates are dismissed.
New instructions, a new model, a new data source — every version is run against the full golden set before it's published. Instructions are versioned like code.
An LLM judge scores each answer against the rubric, with human verdicts where it matters. Runs are compared head-to-head, case by case — so "better" is a number, not a feeling.
Swap the ERP for yours. Swap supplier documents for your contracts, spec sheets, or policies. Swap the LMS for your handbook. The architecture is the same: an MCP tool layer over your systems — live queries for fast-moving data, an indexed knowledge base for everything else — that any agent front end can plug into, with an evaluation pipeline that keeps it honest. If your team spends its day hunting answers across systems, this pattern will pay for itself.