The hook: stop asking AI to “transform procurement”
The most expensive mistake I see is teams using GenAI for big, vague ambitions (“make our contracts better”) instead of small, repeatable chores. Procurement doesn’t need a robot negotiator. It needs a tireless analyst that can read 50 pages without getting bored, then hand you a clean shortlist of issues to decide on.
Used well, GenAI turns messy documents into structured work: clause checklists, redline suggestions, deviation tables, and proposal comparisons. Used badly, it produces confident nonsense, copies confidential text into the wrong place, or “helpfully” changes legal meaning. The difference is your prompts, your inputs, and your review discipline.
A practical mindset: treat GenAI like a junior analyst who is fast, literal, and sometimes wrong. You wouldn’t send a junior’s draft to a supplier without review. Same rule here.
The prompts: three structures you can copy into your workflow
These are written to work in most enterprise chat tools. Replace bracketed text, paste only what you’re allowed to paste, and keep outputs in a format you can act on (tables, bullets, issue logs). If your tool supports file upload with approved security controls, still assume anything you upload could be discoverable later.
Prompt 1 — Summarise a 50-page MSA into a negotiation-ready issue log
Use this when Legal asks, “What changed from our template?” or when you need to brief a stakeholder in 10 minutes. The trick is to demand citations to section numbers and to separate “facts from the text” from “recommendations.”
Exact prompt structure (copy/paste):
ROLE: You are a procurement contract analyst. You do not provide legal advice. You only summarise what is in the text and flag negotiation risks.
TASK: Create an issue log for the Master Services Agreement below.
OUTPUT FORMAT: 1) A table with columns: Topic | Clause/Section | What it says (plain English) | Why it matters (commercial impact) | Suggested fallback position | Questions to ask the supplier. 2) A second list titled “Missing or unclear items” (items not found or ambiguous).
RULES:
Cite the clause/section number for every row.
If you are unsure or the text is ambiguous, write “UNCERTAIN” and explain why.
Do not invent terms that are not in the agreement.
Keep suggested fallback positions commercially oriented (e.g., cap, carve-out, notice period), not legal drafting.
CONTEXT:
Our preferred positions: [insert 5–10 bullets, e.g., liability cap = fees paid in 12 months; mutual confidentiality; IP ownership stays with customer; 30-day termination for convenience; data processing addendum required].
Deal profile: [category], [approx annual spend], [criticality], [data sensitivity].
MSA TEXT: [paste the relevant sections or the full text if permitted]
Prompt 2 — Extract SLA credits and penalties (and find the “gotchas”)
SLA language hides pain in definitions: measurement windows, exclusions, notice requirements, and “sole remedy” wording that blocks other claims. This prompt forces the model to pull the mechanics, not just paraphrase.
Exact prompt structure (copy/paste):
ROLE: You are a procurement performance and SLA analyst.
TASK: From the SLA text below, extract all service levels, credits, penalties, and enforcement conditions.
OUTPUT FORMAT: A) A table with columns: Service level name | Metric definition | Target | Measurement window | Reporting source | Credit/penalty amount | Max credit cap | Customer steps required (notice, ticket, time limit) | Supplier exclusions | Dispute process | Clause/section citation. B) A “Gotchas” list: anything that makes credits hard to claim or reduces value. C) A “Negotiation edits (plain English)” list: 5–10 changes to improve enforceability and value.
RULES:
Quote exact thresholds and percentages.
If a value is missing, write “NOT SPECIFIED” and cite where you checked.
Do not assume industry-standard terms.
SLA TEXT: [paste SLA schedule / exhibit text]
Prompt 3 — Compare four supplier proposals without losing the nuance
Procurement teams often do “apples to oranges” comparisons because proposals use different assumptions (scope, volumes, service hours, implementation included or not). This prompt makes the model surface assumptions first, then compare, then ask for clarification questions you can send back.
Exact prompt structure (copy/paste):
ROLE: You are a sourcing analyst supporting an RFx evaluation. You are strict about assumptions.
TASK: Compare the four supplier proposals below and produce an evaluation view that is usable in a steering meeting.
OUTPUT FORMAT: 1) Assumptions & gaps table (per supplier): Scope included/excluded | Pricing model | Volume/usage assumptions | Implementation approach | Service hours/time zone | Dependencies | Contract term | Indexation | Notable risks. 2) Side-by-side comparison table with columns: Criterion | Supplier A | Supplier B | Supplier C | Supplier D. 3) Clarification questions (minimum 5 per supplier) written as email-ready bullets. 4) A short recommendation paragraph that is conditional (e.g., “If we prioritise X, Supplier B is strongest because…; if we prioritise Y, Supplier D…”).
EVALUATION CRITERIA & WEIGHTS: [insert your real criteria and weights, e.g., total cost 35%, implementation risk 20%, security 20%, service 15%, commercial terms 10%]
RULES:
Do not normalise pricing unless you show the assumptions you used.
If proposals are missing data, do not fill it in—flag it.
PROPOSAL A: [paste]
PROPOSAL B: [paste]
PROPOSAL C: [paste]
PROPOSAL D: [paste]
The guardrails: where GenAI helps, and where it can hurt you
GenAI is a text engine, not a truth engine. It will produce a clean-looking answer even when the input is incomplete, contradictory, or outside its competence. In sourcing, that’s not a minor flaw—it can change negotiation positions, misstate obligations, or create a false sense of diligence.
Data privacy: Don’t paste supplier proposals, pricing, personal data, or contract text into consumer tools. Use an enterprise-approved environment with clear retention rules, access controls, and (ideally) no training on your inputs. If you can’t explain where the data goes, don’t upload it.
Confidentiality and privilege: Treat AI outputs as potentially discoverable. If Legal wants privilege protection, align on the workflow (tooling, access, labeling) before you run sensitive redlines through a model.
Hallucinations: Require clause citations and “UNCERTAIN” flags in your prompts. If the tool can’t point to the exact section, assume it may be wrong until you verify.
Redlining risk: AI can suggest edits that look reasonable but shift meaning (e.g., changing “material breach” to “breach,” or narrowing an indemnity without noticing). Use AI to propose options, not to finalise language.
Human-in-the-loop is not optional: Contract interpretation and negotiation positions have legal and commercial consequences. A qualified person must review, decide, and own the output. If your process can’t name the accountable reviewer, you don’t have a process.
Version control: AI summaries become stale the moment a new draft arrives. Tie outputs to a document version, date, and source (e.g., “MSA v3 received 14 May”).
Bias toward neatness: AI will smooth over messy issues. Force it to list missing terms, ambiguities, and contradictions so you see the rough edges early.
A counterintuitive tip: restrict the model’s freedom. The more you ask for “insights,” the more it improvises. Ask for extraction, tables, citations, and question lists. That keeps it in the lane where it’s genuinely useful.
How this buys you time (and what to do with it)
If you standardise these micro-tasks—issue logs, SLA extraction, proposal comparisons—you can compress work that normally drags across evenings: reading, re-reading, and turning documents into meeting-ready material. The time savings come from fewer blank-page moments and fewer missed clauses, not from skipping human judgement.
The best use of the reclaimed hours isn’t “more admin.” Put it into the parts of procurement that don’t autocomplete well: aligning stakeholders on trade-offs, pressure-testing supplier assumptions, shaping negotiation strategy, and building a category narrative that Finance and the business can actually support. That’s where a buyer earns their keep, and where GenAI is happy to stay in the background doing the grind.