Azure OpenAI on internal product and compliance documents
A family-owned manufacturer of aluminium components in western Sydney was spending roughly twelve staff hours a week searching for product specifications. Their catalogue of engineering PDFs, test reports and compliance certificates had grown over two decades and lived in a shared drive organised by decade rather than by product line.
Sales staff were forwarding emails to the engineering team asking for dimensions; engineers were interrupting design work to answer them. When a new estimator joined, it took six weeks before they could find anything without asking a colleague.
We deployed Azure OpenAI inside the customer's existing Azure tenant. No data left their Microsoft boundary. We ingested the PDF catalogue into Azure AI Search, ran OCR on the scanned files, and built a retrieval-augmented generation pipeline so the model cited its source documents rather than making up answers.
The copilot surfaced in Microsoft Teams, which the business already used. Staff asked questions in plain English: "What is the breaking load on the 40mm section used in the 2022 catalogue?" The answer returned with a link to the original PDF and the page number. A simple governance layer blocked queries that referenced pricing or customer names, and every interaction was logged for review.
We did not train a model. We did not build a foundation model. We integrated the customer's existing documents into a retrieval system on Azure, which is the approach we take with every AI engagement.
The customer is now extending the copilot to include supplier correspondence and warranty records in a second phase.
The manufacturer had product specifications, compliance certificates and test reports scattered across a network drive, a SharePoint site and email attachments. Estimators spent an average of twenty minutes per quote looking up the right document. We indexed the document set into Azure AI Search inside the company's existing Azure tenant, deployed Azure OpenAI with retrieval augmentation, and surfaced the answer in a Microsoft Teams tab.
When an estimator asks "What is the load rating on the 120mm beam?" the system retrieves the relevant PDF, extracts the specification, and returns the answer with a citation link to the original document. If the answer cannot be found, the system says so. It does not invent a number. This was a deliberate design choice: in a manufacturing context, a hallucinated specification is a safety problem, not an inconvenience.
No documents leave the company's Microsoft boundary. No prompts are used to train the base model. We logged prompts and responses for review, and we set a rule that anything the AI drafts for a customer quote must be approved by a human before it goes out.
In the first three months, estimators reported average lookup time dropped from twenty minutes to two. The system was used by eight staff, including two estimators, three sales people and the quality manager. We did not roll it out to the whole company in week one. We ran the pilot on one document set, measured the time saving, and then expanded to the next set.