A UK firm asked us to look at its slowest process: quoting. Customers send their requirements as spreadsheets, and every spreadsheet is different. Some are clean. Most are not.
We started the way we start everything: by pricing the labour instead of admiring the technology. Turning one of those spreadsheets into a quote meant a buyer working line by line, identifying each item, checking availability, chasing supplier pricing and assembling the response. The workflow occupied roughly 60 people, and a single quote took hours. Against the heuristic we publish, this was the loudest arithmetic we have ever run. The recommendation was not a tweak. It was a platform.
Not a Workflow Fix. An Operating System for Sourcing.
What shipped is a fully live web application that the business runs its sourcing operation through, now in its fourth major version. It has three jobs, and it does all of them at once.
Agentic sourcing
A user uploads the raw spreadsheet exactly as the customer sent it. From there, agents take over: parsing the file, identifying every line item, matching each one against a knowledge graph of items, suppliers and their relationships, and pulling live supplier pricing automatically. Clean lines flow straight through to a built quote. Ambiguous lines are not guessed at: the system flags them, states what it is unsure about, and queues them for a person. The agents do the volume; they never do the judgement.
Multi-level management
The platform is task-centric, and the whole organisation works inside it. When a line needs a decision, it routes to the level that can resolve it: agent, salesperson, country lead or chief executive. Escalations move up the same ladder; nothing waits in an inbox, and nothing gets resolved by whoever happened to see the email first. Every level sees the queue that belongs to it, which is how a workflow that once occupied roughly 60 people became something management can actually steer.
Client liaison
The customer conversation lives in the platform too. From the first spreadsheet arriving to the final quote going back, the thread with the client is tracked and driven from the same system that is doing the sourcing, so the response that used to take hours of assembly goes back while the enquiry is still warm.
Built Like a Product, Because It Is One
The application runs on Next.js 15 with Supabase. The item and supplier relationships live in a Neo4j graph that deepens with every quote processed. The agentic parsing and matching runs on Claude through the Vercel AI SDK, and AG Grid powers the dense tabular screens the team works in all day. We iterated through four major versions with the client, tightening the product with each release.
Stack: Next.js 15, Supabase, Neo4j, Claude via the Vercel AI SDK, AG Grid, live supplier pricing integrations
The Outcome
A quote that took hours now builds in seconds, inside an operation that once occupied roughly 60 people, and the platform has grown from a quoting fix into the system the sourcing business runs on. Humans stay in charge of every judgement call: the system flags problems rather than hiding them, so people review exceptions instead of retyping spreadsheets.
That is the pattern our audits exist to find: the system does the volume, your team makes the decisions. The engagement is under NDA, so the client is not named here; we are happy to talk through the specifics on a call.
