01 · The longlist
The need. The goal isn't a perfect list. It's a broad, qualifiable base, produced in hours rather than days, and cleaned of what you already know.
The tools. The specification, the supplier charter, the panel export. Read-only: at this step, the agent writes nothing that leaves your walls.
The prompt.
I have a new need. Attached documents: the specification, our supplier
charter, the export of our current supplier panel.
My selection criteria, in order of importance:
[e.g. price, lead time, certifications, production capacity, proximity].
Step 1: propose a longlist of candidate suppliers.
- Use the specification to understand the need (product, volume, zone,
constraints).
- Rule out any candidate incompatible with the supplier charter, and state
which one and why.
- Compare your list to our panel: separate "already known" from "new
entrants".
- For each candidate: name, why it could fit against the criteria, and what
remains to be verified. Do not fill in any unverified data: mark it "to be
confirmed". Present it all as a table.
I validate the longlist before we move to the RFI.
To reach beyond the model's memory, plug in live sources: search by NAF/APE code via INSEE's Sirene API (free), the legal and financial data from Pappers to check a candidate's health, the French government's company directory. Europages and Kompass for industrial Europe. And a field note: for manufacturing suppliers, Grok is surprisingly effective at finding candidates.
The guardrail. An AI can cite a supplier that doesn't exist, or credit it with a capability it lacks. Models fill gaps: it's their natural slope. The counter is threefold: the "to be confirmed" instruction in the prompt, the Sirene or Pappers cross-check on every new entrant, and your own review. Treat the longlist as a lead to validate, never as an established fact. You prune, you complete, you own the list that remains.
The measure. Time from expressed need to validated longlist, and the number of serious candidates a manual pass wouldn't have found.
02 · The RFI
The need. A questionnaire whose answers compare, with no re-keying and no interpretation. The golden rule: every RFI question must feed a cell of your scoring grid. A question that doesn't feed the grid is one question too many.
The tools. The scoring grid (with its weighting) and the supplier charter, attached to the prompt. The agent writes a document; you alone send it.
The prompt.
Here is my validated longlist, my supplier scoring grid (with weighting) and
our supplier charter.
Write a structured RFI document, ready to send:
- Every question must map to a criterion of the scoring grid, in its
weighting order. Flag any grid criterion no question covers.
- Favor closed, quantified or multiple-choice questions; avoid vague open
questions.
- Append the supplier charter and require a formal compliance commitment in
the answer.
- Structure: statement of the need, questions by theme (capacity, price, lead
time, certifications, references), imposed answer format, deadline.
- Professional, neutral tone. Reveal neither target budget nor negotiation
strategy.
The imposed answer format is not a detail. Answers aligned to the grid on the way out, mechanical processing on the way back. The time you spend reviewing the RFI, you recover threefold at the synthesis. And if you want to go further: ask the agent for a simple web page that hosts the questionnaire and writes each answer to a database. Better experience for the supplier, zero re-keying for you.
The guardrail. Review the RFI as if it were going out under your signature, because it is. Check that no sensitive information leaks, that the need is described without exposing your negotiation position, and that the charter's knockout criteria are indeed attached. The agent drafts fast; you stay responsible for what goes out.
The measure. RFI drafting time, and the share of answers usable without a clarification follow-up. That second figure tells you whether your questions were the right ones.
03 · The synthesis and the shortlist
The need. The answers come back, uneven in format and detail. It's the longest step by hand, and the one where the agent helps most: mapping onto the grid, scoring, gap detection.
The tools. The RFI answers, the scoring grid with weighting, the supplier charter, and the public sources from step 1 for cross-verification.
The prompt.
Here are the suppliers' RFI answers [attach], my scoring grid (with
weighting) and our supplier charter.
1. Map each answer onto the grid: one row per supplier, one column per
criterion, in weighting order. Use the supplier's answer; if it's missing,
ambiguous or contradicts another, flag it explicitly, don't fill the gap.
2. Check the consistency of declared data against public sources (Sirene,
Pappers): headcount, financial health, legal existence. Flag any
discrepancy.
3. Check the compliance commitment to the supplier charter: present, absent,
or with reservations.
4. Propose a shortlist of 3 suppliers ranked by the grid's weighting, with,
for each, the ranking justification, the detected gaps and the points to
dig into in the next phase.
The weighting stays mine. The final decision is mine.
The guardrail. This is the step that steers the decision, so the most sensitive. Three precautions. Spot-check that the grid cells match the answers received: the agent can misread a phrasing. Read the justifications with a critical eye: a candidate can rank first on score and be a knockout on a criterion only your operator judgment sees (dependency, history, field constraint). And keep the weighting: it's your trade that sets the weight of criteria, not the model.
The measure. Time between receiving the answers and the shortlist being ready, number of suppliers actually compared, and number of corrections made to the agent's grid. Three figures that tell you whether the agent won you a cycle.
04 · The ROI, quantified
Replace these assumptions with your own measures: the method matters more than the values.
| Assumption | Starting value |
|---|
| Frequency | 1 new-need sourcing per month |
| Manual time (longlist 1 d + RFI 0.5 d + processing 1.5 d) | ~3 days per sourcing |
| Time with the agent (framing 1.5 h + RFI review 1 h + synthesis supervision 2 h) | ~0.6 day |
| Loaded daily cost of a buyer | €360/day |
| Token cost per sourcing | a few euros (public vendor pricing, 2026) |
The calculation. Manual: ~3 days, i.e. ~€1,080 of buyer time per sourcing, close to €13,000 over 12 sourcings. With the agent: ~0.6 day, i.e. ~€215 per sourcing, close to €2,600 a year. The gain: ~2.4 days per sourcing, about 80% of the time, i.e. close to 29 buyer-days returned a year, around €10,000.
The gain the spreadsheet doesn't show. First, speed: a shortlist in two days instead of two weeks means a business need served before it drifts into off-procedure maverick spend. Then, breadth: a rushed buyer compares 3 or 4 suppliers; an agent clears 10 at no extra cost. More candidates compared means better competition, so a better price. That gain lands in savings, not in time.
Report both in the savings tracker, with the baseline measured on a real sourcing and the calculation method. One figure that holds, with its source, beats ten hoped-for figures.
05 · The trap to avoid
Never let the agent decide the shortlist, even when its grid is flawless. The formatting, yes; the weighting and the final call, no. A supplier can tick every box on paper and be a knockout on a criterion only your operator judgment sees. In every case, the agent is an execution tool, not a decision tool.
The agent gets you to the decision faster. The decision stays yours.
Before your next sourcing, do one thing: pull out your scoring grid and your supplier charter, and give them to the agent. It's the context that turns a guess generator into a framed colleague.
A question about your specific case? Write to me: alex@theprocurementor.com. I answer every message.
The ROI figures are assumptions to replace with your own measures. Always spot-check the data the agent produces before you present it.