AI for procurement: spend, supplier documents and quote comparison.

The purchasing meeting has three items on the agenda: why packaging spend went up, which new suppliers are still missing documents, and which of three quotes for a replacement part to accept. Each question sends someone into spreadsheets, shared folders and email attachments before anyone can apply judgment.

Procurement is heavy on documents and on judgment, and AI belongs on the document side. It can classify spend, read supplier documents and quotes, and compare offers line by line with sources. The choice of supplier and every commitment stay with buyers.

Where does procurement time go?

• Reading and re-keying: invoice lines, supplier documents, quotes and order confirmations, each in its own format.
• Chasing: missing or expired documents, unconfirmed orders, answers to questions sent to suppliers.
• Normalizing: turning offers into the same unit, currency and terms before they can be compared.
• Judging: negotiating, choosing suppliers and managing relationships, the part that deserves more of the week.

The first three are where AI helps. The aim is to give the fourth more time.

How can AI help with spend classification?

Spend lines arrive with free-text descriptions and inconsistent supplier names, so spend analysis starts with someone sorting them by hand. AI can do the sorting into your own category tree, the one your category managers use, not a generic list.

The method is simple. Rules go first where they exist, such as a supplier that only ever sells IT services. For everything else, the model proposes a category with a confidence score and the reason, usually the words in the description and the supplier’s past lines. Low-confidence items go to a category analyst, and each correction becomes an example for the next run. A reviewed sample shows accuracy by category.

Two details matter in practice. Classify at line level, not invoice level, because one invoice can carry several categories. And keep the model’s reason visible, so an analyst can fix the rule behind a wrong answer, not just the line.

Duplicate supplier records distort spend figures before classification even starts; our note on cleaning master data covers that step.

How can it help with supplier onboarding documents?

A new supplier sends a stack of documents: tax registration, a trade registry extract, certificates, insurance, a bank letter, signed policies. AI reads each one and records the document type, the issuer, the supplier name and tax number, and the issue and expiry dates. It then checks the set against the list your policy requires for that type of supplier, flags what is missing or doesn’t match the supplier record, and warns ahead of every expiry date. Documents arrive in Turkish, English and other languages, often as scans with stamps; anything the system can’t read with confidence goes to a person instead of being marked complete.

Bank details deserve their own rule. The system can flag that an account number on a letter differs from the one in the ERP, but it never updates bank details. Every change is verified by a person through a known contact, using a phone number already on file rather than one in the email that requested the change, because a fake change request is a well-known fraud pattern.

Which documents a supplier must provide is your policy, set with your legal and tax advisors; the system only checks against it.

How can it compare quotes fairly?

Fair means the same basis for every offer. For each quote, AI extracts the item, quantity, unit, unit price, currency, discounts, delivery terms, lead time, payment terms and validity date. It converts units and currencies to a common basis, using the exchange rate and date your finance team sets, and flags anything that can’t be compared, such as a different specification or a missing lead time. Two traps deserve a flag every time: an offer for a slightly different part, such as an alternative material or a newer revision, and a price that depends on volume tiers.

An illustrative example: three quotes for the same part, normalized into one table.

Every value links to the place in the document it came from, so the buyer can check any figure in one click. The table also shows what the price alone hides: B’s freight and C’s split delivery. A full version adds lead times, payment terms and validity dates the same way.

A comparison you can’t click back to its source is just another spreadsheet.

If the buyers want weighted scoring, they set the criteria and weights before the quotes are opened. The system applies them; it doesn’t invent them.

Where must buyers decide?

• Supplier choice. The comparison informs it; a buyer makes it.
• Negotiation and relationships. Quality history, strategic fit and trust are judgment calls.
• Every commitment. Purchase orders, contracts, framework agreements and acceptance of non-standard terms. For contract terms, see our note on clause extraction.
• Supplier approval and bank details. A person approves a new supplier and every change to how it is paid.

AI prepares, compares and flags. It never sends an order or accepts terms.

Which first workflow should you choose?

Pick by volume, clear rules and a named owner. Spend classification is low risk and quick to measure. Onboarding document checks suit a growing supplier base with clear requirements. Quote comparison is valuable per case but usually lower in volume, so it pays back where many similar quotes arrive.

After the order comes the confirmation. Reading supplier confirmations and checking dates and quantities against the open order is the bridge from procurement to supply chain work, described on the supply chain intelligence page. Whichever workflow you choose, start with read-only access to orders and supplier records, as the ERP and enterprise workflows page describes, and record the baseline before anything is built: hours per quote comparison, documents chased per new supplier, and the spend still sitting in “unclassified”.