ChatGPT for procurement: which tasks, which prompts, which rules?

What can ChatGPT do for procurement?

ChatGPT can do the first draft of most procurement writing: RFP sections, questions to suppliers, contract summaries, negotiation notes and emails. It can also do a first pass on spend data. It cannot be trusted with totals, extracted contract values or facts about suppliers without a check against the source.

The table pairs each everyday task with what ChatGPT does well and the check it needs. The evidence comes from published studies, most of them on AI tools in general rather than ChatGPT alone.

For what AI does in procurement in general, beyond ChatGPT, eProcurement Toolkit’s guide What is AI in procurement? covers the use cases. The wider picture of strategy and governance is on the AI procurement strategy page.

Which ChatGPT plan should a procurement team use?

A procurement team should use a company workspace, ChatGPT Business or Enterprise, rather than personal Free, Plus or Pro accounts. The difference is not the model. It is what OpenAI may do with the data, who controls how long it is kept, and whether there is a contract on data processing.

OpenAI changes plans and terms. Check its enterprise privacy page and its help article on consumer data before you write your rules, and again once a year. If your company licenses Microsoft Copilot instead, Microsoft says prompts and responses are not used to train its foundation models, and Copilot only shows data the user already has the right to see.

What data should never go into ChatGPT?

In a personal account, no confidential or personal data at all. In a company workspace, your company’s own data rules and your supplier contracts still apply: a workspace with no training by default does not lift a non-disclosure agreement. Samsung banned generative AI tools for its staff in May 2023 after internal code was pasted into ChatGPT.

• Supplier prices and offers under evaluation. A price grid or a bid belongs to the supplier, often under an NDA or tender rules. Keep it in the company workspace, or replace names and figures before you paste.
• Personal data. Names, emails and phone numbers of supplier contacts, and anything about employees. Remove them; the task rarely needs them.
• Signed contracts. Fine in a company workspace if your rules allow it; never in a personal account.
• Public tender documents and bids. In public procurement, the rules of the tender apply. The UK’s PPN 017 suggests controls so that tender information is not used to train AI models.
• Your negotiation position. Your real target price and your fallback if there is no deal. For role-play, use made-up figures.

A simple test before you paste: would you send this text by email to someone outside the company? If not, it goes only into the company workspace, or not at all.

Which prompts work for procurement tasks?

Prompts that work for procurement give the role, the documents, the format of the answer and the limits: what to leave out and what to mark as uncertain. The examples below follow that shape. Replace the text in brackets, and keep confidential data in the company workspace.

Draft an RFP section from past RFPs.

You are a procurement specialist. Attached are two past RFPs for [category]. Draft the section "[section name]" for a new RFP. List the requirements as numbered items, each one testable. Add, in a separate list, questions the two past RFPs did not ask. Do not invent company names, figures or standards; write [to confirm] where you need information.

Prepare questions to suppliers.

We are buying [what], for [users, volume, countries]. Write 15 questions for an RFI, grouped by: scope, service levels, pricing model, data and security, exit. For each question, say in one line what the answer helps us decide. Plain English: many suppliers read English as a second language.

Compare two contract clauses.

Compare clause A (our standard) and clause B (the supplier's version) below. Give a table: topic, clause A, clause B, difference, risk for the buyer. Quote the exact words from each clause for every difference. If something is not in the text, write "not in the text"; do not infer it.

Clean and classify spend lines.

Below are [number] spend lines: supplier name, description, amount. Our taxonomy is: [list of categories]. For each line, return a CSV with: line number, supplier name cleaned (same spelling for the same supplier), category, confidence (high, medium, low), and a short reason. Mark as low confidence any line you cannot classify from the description alone. Do not change the amounts.

Prepare a negotiation.

Play the account manager of a supplier of [category]. I am the buyer. The figures below are made up for practice. Push back on my requests the way a real supplier would: lead times, minimum volumes, price indexation. After ten exchanges, stop and list the arguments I did not answer well.

Ask for quotes and formulas, not conclusions: “quote the clause”, “show the formula”, “mark what you are not sure of”. That makes the check that follows faster.

What does ChatGPT do badly in procurement?

ChatGPT does badly on the things procurement signs off: numbers, exact values from contracts, references and facts about suppliers. It writes a wrong answer in the same confident tone as a right one. Microsoft says the same of its own Copilot: its answers are not certain to be factual.

• Totals and price comparisons. Apple researchers found that the performance of all the models they tested fell when only the numbers in a problem changed, and dropped by up to 65% when a clause that does not matter was added. For a bid comparison, have the numbers computed in a spreadsheet or in code, and recheck the totals.
• Values extracted from contracts. In the Vals Legal AI Report, the best legal AI tools extracted data about as well as lawyers, at 73% to 75% against 71%. That is not good enough to skip the check: one wrong cap or notice period is enough.
• Legal answers. Stanford researchers found in 2024 that even specialised legal research tools gave incorrect information on more than 17% of their test queries, and one tool on more than 34%.
• References and sources. A report worth about AU$440,000, delivered by Deloitte to an Australian ministry in 2025, contained a false court quote and references that did not exist, and Deloitte agreed to refund part of the fee. Deloitte did not say AI caused the errors. Check that every source a draft cites exists and says what the draft says.
• Suppliers and market prices. A list of suppliers from ChatGPT is a list of leads. Check that each company exists, makes what you need and has the size you need.

How do you check ChatGPT’s work on procurement documents?

Check ChatGPT’s work against the source, not against your memory of it. The checks take minutes when the prompt asked for quotes and formulas, and they are where the team learns what the tool gets wrong.

1. For a contract summary, open the signed text and find each clause the summary cites. Note any clause the summary missed.
2. For extracted values, check every date, amount, cap and notice period in the text, one by one.
3. For a bid comparison, rebuild the totals in a spreadsheet from the bids themselves.
4. For spend classification, check a sample, for example 50 lines, against your taxonomy, and count the errors.
5. For any cited source or supplier, open it and confirm it exists and says what the draft says.
6. Before anything goes to a supplier or to a decision committee, a named person reads it and signs it off.

Keep a short list of the errors found. After a month, it tells you which tasks are worth it and which need more checking than they save.

How does a procurement team learn to use ChatGPT well?

A procurement team learns to use ChatGPT well by practising on its own documents, with the rules written down and the checks done together. BCG writes, about AI in general, that about 70% of the value comes from people adopting new ways of working, and only 10% from the algorithms.

There is a legal side too. Under article 4 of the EU AI Act, as amended in July 2026, a company whose staff use AI systems at work must take measures to support their AI literacy. The European Commission suggests keeping a record of trainings. The AI procurement strategy page covers the rest of the governance.

I run AI training for procurement teams: in-house, up to 12 people, from €2,500 for a half day or from €4,500 for a full day. The exercises are RFP drafting, contract review with a check against the text, supplier research with sources, and, in the full day, spend classification and negotiation preparation, on the AI tool your company licenses. The team leaves with a one-page charter and an attendance record.