How the network works, end to end
Bizily is an AI procurement network for technical supplies: a request arrives in whatever form the buyer sent it, five agents turn it into a quote draft in which every field is grounded in the source, and a person approves before anything is sent, priced or written.
One request, one reviewable draft
Technical supply runs on requests that were never structured to begin with: a QC lab sends seven lines in the body of an e-mail, a facility attaches a scanned PDF, a site manager photographs a handwritten list. The work that follows is the same every time — retype it, decide what each line actually is, find the ones you do not stock, and rebuild a quote. The network runs that loop as a pipeline of named agents, each of which proposes and then stops.
- 1 · Intake
- E-mail, PDF, spreadsheet, chat or a photo becomes one structured record. The original is kept as evidence and never rewritten.
- 2 · Extraction
- One line per requested item: quantity, unit, part number and the span of source text each field came from. Ungrounded fields are dropped and counted.
- 3 · Catalog match
- Rules match first and carry their rule id and confidence. The model only re-ranks what rules could not place, and abstains below your threshold.
- 4 · Sourcing
- Lines nobody stocked go to your own suppliers as drafted RFQs. Replies re-enter through intake and become offers with landed cost.
- 5 · Quote draft
- Approved lines and terms become a quotation in your format, with your letterhead and your numbering.
- Review gate
- Nothing is sent, priced or written until a person approves, corrects or rejects each proposal.
Three sides, one network
The same pipeline serves three groups, and what each of them sees is deliberately separate. Distributors run the back office on their own requests. Buyers use the buyer app to ask the suppliers they already have. Manufacturers list a catalog once so it is available to distributors as a canonical, human-reviewed record.
- Distributors
- A free console for answering structured RFQs, plus a paid back office covering intake, sourcing, quotes, orders and invoices.
- Buyers
- A list becomes structured RFQs to the buyer's own named suppliers, with quotes compared side by side. It never buys.
- Manufacturers
- One canonical record per product with vendor-authoritative provenance, in front of distributors who decide what to import and stock.
RFQs are buyer-directed: a buyer’s request reaches only the suppliers that buyer names, and a line is offered more widely only when the buyer explicitly asks to source something their own suppliers cannot fill. Requests a distributor receives from its own customers are visible to that distributor alone.
Why a person is still in the loop
A quote is a commitment: a price, a lead time and a part number that a laboratory will plan around and an auditor may later read. That is the wrong place for a system that is usually right. So the review gate is not a setting — agents produce proposals, and approval is the only path by which a proposal becomes a record.
The audit trail is built for the same reason. Every model call is ledgered per tenant and per call with prompt hash, model, tokens, cost, confidence and injection flags. An AI back office should be something you can audit rather than something you have to trust.
What it does not do
The platform is never merchant of record. It does not sell, hold stock, take a cut of the transaction or stand between a buyer and their distributor: purchase order, invoice, delivery, warranty, returns, disputes, import and regulatory registrations all stay with the distributor. Prices, costs and customers are private to a tenant and are not reused across tenants.
Where it fits
Multi-brand, quote-first technical supply: laboratory and scientific instruments and consumables, medical and diagnostics, chemicals and reagents, industrial MRO, safety and cleanroom — the categories where documents matter as much as the item, such as a certificate of analysis by lot, a calibration certificate by serial, or a safety data sheet. North America and Southeast Asia first.
Questions
- What is an AI procurement agent?
- An agent is a narrow, named step in a pipeline, not a chatbot. Each one takes a defined input, produces a structured proposal, and stops. The intake agent turns a message into a record; the extraction agent produces one line per requested item with an evidence span; the catalog match agent places lines against a catalog; the sourcing agent asks suppliers about what is left; the quote draft agent assembles the document. A person approves the result before it becomes a business record.
- What does grounded mean?
- Every extracted field carries the span of source text it came from, so each quantity, unit and part number can be traced to the words in the original e-mail, PDF or photo. Fields that cannot be traced are dropped and counted rather than guessed, and the source is kept as evidence, never rewritten.
- Can the agents send a quote or place an order on their own?
- No. Agents propose; people approve. No model sends a message, sets a price or writes a business record. Every write passes a human review gate, and every model call is on the ledger with prompt hash, model, tokens, cost, confidence and injection flags.
- What happens to a request the catalog cannot match?
- Off-catalog is the normal case, not an error. Rules match what they can; the model re-ranks the remainder and abstains below the threshold you set. Whatever is still unmatched goes to the sourcing agent, which drafts RFQs to your own suppliers and turns their replies into offers with landed cost.
- What does it integrate with?
- Requests arrive by e-mail, PDF, spreadsheet, chat and photo, so the buyer does not have to change how they ask. On the accounting side it is QuickBooks-first. Manufacturers can list a catalog from a spreadsheet or a feed. There is no punchout requirement and no EDI project to run first.
- How is prompt injection handled?
- Untrusted text stays data and is never treated as instructions. Requests arrive from outside the company by definition, so a supplier reply or a PDF that contains something resembling a command is extracted as content, flagged on the ledger, and still has to pass the same human review gate as everything else.