Both cases of the procurement agent and their sub-agents Two cases start separately: case A begins with an article list from the shop and ERP and a sub-agent that normalises one specification per article. Case B begins with a list of candidate products and a sub-agent that derives a requirements profile from each idea. Both then run through the same path: a source router picks one or more fitting channels per product, and below it sit four research sub-agents for the internal database, web search, Alibaba and Europages or wlw. Their hits are merged, then a dedicated sub-agent verifies every single candidate. After that the cases separate once more: case A compares against the incumbent supplier, case B answers feasibility and shortlist. Both continue with one draft enquiry per supplier and one dossier per product. After the dossier the flow branches per supplier: either a negotiation sub-agent takes over and negotiates price, quantity and delivery terms across several rounds — by e-mail or through the platform the supplier was found on — or the case goes straight to procurement. Either way the decision sits outside the agent, with procurement. MAIN AGENT · ORCHESTRATION RANGE ON THE SHELF NEW PRODUCTS Article list from shop and ERP Title · GTIN · incumbent · sales volume List of candidate products Product idea · target market · target margin 01: sub-agent Normalise the article specification and gaps, per article 01: sub-agent Build the requirements profile specification from the product idea 02: sub-agent Source router picks one or more fitting channels per product and records why 03: sub-agent Internal database past enquiries and quotes, suppliers already known 03: sub-agent Web search branded goods, makers with their own website 03: sub-agent Alibaba private label and OEM, MOQ and unit cost 03: sub-agent Europages/wlw sourcing in Europe, standards, small lots 04: sub-agent Merge the candidates duplicates across channels, split makers from resellers 05: sub-agent Verification per candidate existence · category · standards · MOQ · lead time — each with evidence 06: sub-agent Compare with the incumbent switch, second source or stay 06: sub-agent Feasibility and shortlist go or no-go, three candidates 07: sub-agent Draft enquiry per supplier in the supplier's own language, open questions included 08: sub-agent Dossier per product recommendation · evidence · gaps · next step BRANCH PER SUPPLIER 09: sub-agent Negotiation rounds of negotiation (e-mail or platform) on price, quantity and lead time otherwise: direct to procurement Procurement decides — approval, order, contract

Figure 1 — The flow: one main agent orchestrates a number of sub-agents. Click to enlarge.

TL;DR

A short, number-based summary of where this stands today:

  • fully automated process, triggered by a daily check for changes in the input data (article list for the range on the shelf, list of new products)
  • purchase price down by an average of 6 % for new products compared with the manual process
  • 73 % faster turnaround between first contact with a supplier and the final negotiated result, through much shorter response times to supplier messages

Two questions, one research path

Procurement asks two different questions. The first: for the articles we sell today — is there another supplier for them? Cheaper, closer, more reliable, or simply a second source so that one outage does not empty the shelf. The second: for the products we want to add — can they be sourced at all, and from whom?

The questions sound alike but start from different places. In case A the specification already exists in all but name: the article is real, it has dimensions, a packaging unit, required standards, a sales history and a supplier. In case B there is an idea and a target margin. So the agent separates the two cases only at the start. Once a usable specification exists, the path is the same: channel choice, research, verification, dossier — and after that, per supplier, negotiation or handover.

Case A — alternatives for the range already on the shelf

The input is an article list from the shop and the ERP: title, GTIN, product category, current supplier, packaging unit, sales volume. The main agent does not work through that list itself. It starts one normalise the article sub-agent per article, which turns organically grown shop data into a specification you can actually search with: material, dimensions, tolerances, required standards and markings, packaging, quantity band. Whatever is missing from the data is not guessed, it is recorded as a gap — and every gap later becomes a question in the draft enquiry to the supplier.

The case ends with its second case-specific step: the compare with the incumbent sub-agent. It scores each verified candidate not in the abstract but against the supplier delivering today, along the dimensions that decide a switch: unit cost in the relevant quantity band, minimum order, lead time and shipping route, evidence for the required standards, switching effort (new article numbers, new packaging, new photography) and dependency risk. The result is one of three recommendations: switch, add as a second source, or stay with the current supplier.

Case B — suppliers for products that do not exist yet

The input is a list of candidate products, often no more than one line each, plus target market and target margin. Here the main agent starts a build the requirements profile sub-agent per idea. It first researches how the product is described in the market at all: which variants, sizes and materials are common, which standards and markings apply in the target market, which product categories it is listed under. Only then does a specification emerge, and it is deliberately written with ranges, because it is an assumption rather than a record out of the ERP.

The case ends with feasibility and shortlist instead of a comparison, because there is no incumbent to compare against. The sub-agent answers two things: can the product be sourced in the intended quality at the target margin — and if so, which three suppliers would you start with?

The source router — which channels fit which product

The most expensive mistake in supplier research is searching everywhere the same way. A branded article you resell and a private-label product that has to be manufactured first have nothing in common except the word supplier. So a dedicated sub-agent, the source router, picks one or more fitting channels per product — and records its reasoning, so the dossier says why nothing was searched elsewhere.

  • Internal database — search the house itself first: earlier enquiries and quotes, suppliers already scored, old price lists, delivery and complaint history from the ERP. The cheapest channel is the one that has already been paid for. A supplier approached two years ago for a different product category, too expensive at the time, is a hit today — with a route of contact, a known response time and a price base no external channel supplies. The search runs over product category, material and standards rather than free text, and every hit carries its age: a price from 2024 is a lead, not a quotation.
  • Direct web search — when the supplier is a manufacturer with its own website: branded goods, niche and specialist makers, authorised distributors, spare parts. The information sits on manufacturer sites, in dealer directories and trade-fair exhibitor lists, not in a catalogue. The sub-agent fans out queries for it: product term times synonym times role (manufacturer, wholesale, B2B, OEM) times language.
  • Alibaba — when the product has to be manufactured and unit cost, minimum order quantity and OEM capability carry the decision: private label, interchangeable commodity goods, packaging, accessories.
  • Europages and wlw — when sourcing is meant to stay in Europe: short delivery routes, small lot sizes, standards and documentation duties, customs and supply-chain risk.

The router deliberately picks more than one channel when the question has more than one answer. The internal database runs almost every time, because it is cheap and it secures the hit list against what the house already knows. For a private-label product with marking requirements, Alibaba and Europages run in parallel on top, because the question is not only who manufactures more cheaply but also what a European source does better on documentation and lead time. Every chosen channel becomes its own research sub-agent, and those run at the same time.

One sub-agent per step, one per candidate

The main agent does not research. It holds the product list, the routing rules, the budget and the state of the run — and starts, for every step, a sub-agent with exactly the slice that step needs:

  1. Normalise (case A) or requirements profile (case B) — one specification plus its gaps, per product.
  2. Source router — one or more fitting channels per product, with its reasoning.
  3. Research per channel — candidates from the internal database, web search, Alibaba, Europages or wlw, each in the same schema: company, country, role, source, supporting quote.
  4. Merge — spot duplicates across channels (the same company sits in the internal database, on Europages and on its own website), normalise company names and legal forms, split resellers from manufacturers when a manufacturer was asked for.
  5. Verification — one sub-agent per candidate. Does the company exist, with an imprint and a register entry? Does it really manufacture this category, or only trade in it? Are the required standards and test marks findable? Minimum order, lead time, export capability, route of contact?
  6. Comparison (case A) or feasibility (case B) — the two steps above.
  7. Draft enquiry — one enquiry per supplier in that supplier’s language, carrying the specification and the open questions from step 1; who sends it is decided by step 9.
  8. Dossier — one page per product: recommendation, evidence, gaps, next step.
  9. Negotiation or handover — per supplier: either a negotiation sub-agent takes over the correspondence (e-mail, or the platform the candidate was found on), or the case goes to procurement with its dossier and enquiry as it stands.

This split is not tidiness, it settles four problems at once. First, the context budget: forty candidates with a few pages of source material each fit into no context window, while forty sub-agents with a few pages each do. Second, parallelism: products, channels and candidates are independent of one another, so they run at the same time. Third, the blast radius: an invented supplier dies inside the verification sub-agent that checks only that one candidate, instead of riding along in a context where twenty other candidates distract from it. Fourth, testability: every step has its own success criterion, so it can be scored on its own and replaced on its own. A new channel is then a new research sub-agent rather than a rebuild — the internal database hangs off it in exactly that way.

After the dossier — negotiate or hand over

The dossier is the obvious place to stop, and for some of the suppliers it is. For the others it would waste a round. So after step 8 there is a branch, and it is set per supplier: either a negotiation sub-agent takes over the dialogue, or the case goes to procurement with its dossier and draft enquiry as it stands.

The branch is not a judgement about the supplier but about the conversation. A negotiation agent fits when it is transactional and bounded: first contact with no existing relationship, clearly specified goods, a manageable volume, platform suppliers who negotiate tiers in writing and in standard phrases anyway — by e-mail or in the platform’s own message channel. It does not fit when more hangs on the negotiation than a price: strategic suppliers, running framework agreements, exclusivity, regulated categories, anything with liability or contract terms, and every supplier where the relationship is worth more than the next three per cent. Which of those applies is decided by the main agent, from volume, product category, relationship status out of the internal database and an approval list from procurement — and the decision goes into the dossier with its reasoning, like every other.

Where the negotiation agent takes over, it is given a mandate rather than a target: target price and ceiling in the relevant quantity band, acceptable minimum order, maximum lead time, payment terms, Incoterm corridor, sample requirement — and the conditions for breaking off. Then it runs the dialogue a buyer would otherwise run: enquiry out, read the answer, convert the quote into the same schema the dossier uses, formulate a counter-offer, next round. In the supplier’s language, through a mailbox or a platform account that procurement reads along with.

It ends in one of four states, and all four are a result:

  • Mandate met — the negotiated position sits inside the corridor. The dossier gets the exchange and a recommendation to order.
  • No movement — two rounds with no concession on either side. Break off with the last position, so the next candidate on the shortlist is not left waiting behind a dead end.
  • Outside the mandate — the supplier asks for something the agent has no authority over: a higher volume, different payment terms, a contract clause. Escalation to procurement, with the open question first in the dossier.
  • Round limit — a hard ceiling, so that a friendly but fruitless exchange does not run on forever.

What the agent does in none of those four states: commit. No order, no binding volume commitment, no contract — not even by implication in an e-mail or a platform message. The negotiation agent moves the agent’s boundary one step back; it does not erase it.

Evidence or nothing

Supplier research is the process where a language model can do the most damage, because an invented company with a plausible address looks exactly like a real one. So the rule inside the agent is: no claim without a source. Every field in the dossier carries a URL and a supporting quote; anything that cannot be evidenced is reported as unevidenced and left unscored. A candidate with no findable register entry, or with nothing to show it actually manufactures the category, does not enter the dossier with a bad score — it enters with a note on what could not be checked. That makes the shortlist shorter and the handover to procurement faster.

For the negotiation agent the same rule holds in a stricter form: the source here is the supplier’s own message — the e-mail, or the exchange on the platform. Every number it writes into the negotiated position carries the round, and the sentence it stood in.

Where the agent stops

The agent decides nothing. It delivers dossiers, draft enquiries and — where the branch was set to negotiation — a negotiated position with its exchange attached; approval, order and contract stay with procurement. That is not a disclaimer but the division of labour that makes the agent useful: it takes the hours nobody wants to spend — searching, reading, reconciling, writing it all down, and now the first rounds of correspondence too — and puts the decision on the table with its evidence attached.

Four limits remain, and they are printed in every dossier:

  • Platform prices are indications, not offers. It only becomes binding with the answer to the enquiry.
  • A negotiated price is not a deal. The agent collects terms; they are committed to with procurement’s order.
  • A claimed certificate is not a valid certificate. The agent finds the claim; checking the document stays human.
  • No sample, no quality decision. For private label, research is the start of sourcing, not its end.