Procurement and purchasing
Comparing quotes, checking supplier documents, bundling demand: where AI saves time in procurement and where the decision must stay with a person.
What this is about
Procurement receives documents in every conceivable format: PDF quotes, price lists in spreadsheets, technical datasheets, framework agreements. The effort lies in making something comparable out of them.
What works well
| Task | Value |
|---|---|
| Bringing quotes into one table | very high |
| Finding deviations from the tender | high |
| Evaluating supplier self-declarations | high |
| Checking clauses against your own standard | high |
| Bundling and classifying demand requests | medium |
| Forecasting prices | low, worthless without your own series |
Introduction
- 01
One category, one format
Start with a category where quotes arrive regularly. Twenty repetitions a quarter give the data to decide on.
- 02
Define the target fields
Which twelve to twenty fields does the comparison table actually need? Writing that list first prevents an extraction that delivers everything and carries nothing.
- 03
Build in traceability
Every extracted value with page and position. Without it, checking costs more than the extraction saved.
- 04
Set a review rate and lower it
Check everything at first, move to sampling after four weeks if the error rate allows.
- Always check amounts and quantities against the source location, not against plausibility.
- Tender documents contain third-party trade secrets; settle the contractual position before processing.
- In public procurement, the decision and its reasoning are produced without AI involvement.
- A comparison table is an interim result, not an award recommendation.
Data flow and legal basis
Quote (PDF)
→ OCR with coordinates, see /en/engineering/ocr
→ field extraction with source references
→ checksums and consistency rules (net + tax = gross)
→ comparison table with provenance per cell
→ evaluation by procurement, documented
Legal basis. For quote data of legal entities the GDPR does not apply, but it does apply to the contact persons named in them. Art. 6(1)(b) or (f) GDPR carries the processing; what matters is purpose limitation and erasure after the procedure closes.
Trade secrets. Tender documents are routinely protected under the Trade Secrets Directive and its national implementations. Passing them to a processor needs a basis in the tender or the contract, and excluding use for training is not a negotiating position but a precondition.
Public procurement
- Evaluation follows the criteria published in the notice, is carried out by people, and is documented traceably.
- AI-assisted preparation is permissible but must be named in the award record.
- Bidders have a right to reasons. "The system scored it that way" does not carry.
- Equal treatment: a process that handles structured quotes better than unstructured ones systematically disadvantages. That must be checked and documented.
Metrics
| Quantity | Establish a baseline |
|---|---|
| Time from quote receipt to comparison table | across twenty cases |
| Share of mandatory fields extracted correctly | sample of 50 quotes |
| Share of cells corrected afterwards | continuously |
| Cycle time of the whole procedure | before and after |
Choosing a task in procurement
FREE ACCOUNT
Selection sheet and templates: procurement
A sheet that picks the first task out of the ones procurement actually has, with two tested templates and the review steps for them.
Worksheet3 items