This case study is anonymised. We build internal systems, so we describe the kind of system and why it was built, not who it was for.
The situation
The client is a trading company that regularly has to source materials for specific orders: work out what is needed, find suppliers, compare prices and reply to customers before a deadline.
Before, that meant reading order documents for the list of materials, searching online shops and wholesalers by hand, copying prices into a spreadsheet and keeping deadlines in someone's head or calendar. The work was repetitive, and prices copied by hand from many tabs were easy to get wrong.
What we built
A set of n8n workflows that do the legwork and leave the results in the company's existing database, where the team already works.
Material extraction. For a chosen order, a workflow reads the attached documents and uses a language model to pull out the required materials with quantities and units.
Supplier and price search. For each material, a model with web search proposes candidate offers from price comparison sites, online shops and wholesalers, each with a link to a product page. Then comes the important part, which uses no AI: the workflow opens each product page and reads the price only from the page's structured product data. A verified price always takes precedence over the model's guess, and a page that cannot be reached is marked as such instead of failing the run.
Specification check. The model is asked one narrow question per offer: does it match the specification? It does not judge price.
Ranking. The shortlist is ordered by plain code: offers that do not match are dropped, the rest are sorted by price per requested unit, and anything without a confirmed match is flagged for a human to check. Delivery costs are shown as a note to confirm with the shop, not mixed into the ranking.
Request drafts. Another branch looks up suppliers in public business registers and prepares draft requests for quotation for the team to review and send.
Reminders. A scheduled job sends one summary email before deadlines approach and skips anything already closed. When there is nothing to report, nothing is sent.
How it works day to day
The team picks an order, starts the extraction and the supplier search, and gets back a ranked shortlist with verified prices, sources and notes on what still needs checking. Reminders arrive before deadlines become urgent. A small dashboard shows the state of the workflows, and a switch can pause them without anyone opening n8n.
Nobody copies prices from browser tabs into a spreadsheet any more, and a deadline no longer depends on someone remembering it.
What we learned
Let the model judge, let code rank. Language models are useful for finding candidates and reading specifications. Prices, sorting and totals belong to ordinary code, so every position on the shortlist can be traced to a source.
Normalise units before comparing. Offers come per piece, per pack or per metre. Comparing them without converting to the requested unit produces confident nonsense. When a conversion is not possible, the system should say so rather than guess.
Be explicit about time zones. Tools often store dates in one time zone and schedule in another. Decide the convention once and read stored dates the same way everywhere, or late-evening deadlines end up on the wrong day.
Do not compare numbers as text. Many no-code tools compare text fields alphabetically. Store numbers as numbers, or let the workflow that owns the number write a simple flag that conditions can test.
Decide how a safety switch fails. Every on/off switch should have a deliberate behaviour when its setting is missing or unreachable. For monitoring, carrying on is usually safer than stopping silently.
Which service this is
This is n8n automation, with AI used for two narrow jobs: reading documents and checking specifications. Everything that has to be right every time, such as prices, ranking and deadlines, is ordinary code. A related example is Public Tender Monitoring with n8n and AI.
