The fastest objection to using AI in procurement is not "it's expensive" or "it's wrong" — it's "how do I know the AI isn't making this up?". That objection is right to ask, and a tool that can't answer it loses every enterprise procurement evaluation. The answer is source citations: every AI-extracted value is paired with the exact sentence from the source PDF the AI used to derive it.
The hallucination problem in procurement context
Large language models occasionally make up facts that sound plausible. In casual use that's annoying. In procurement it's contract-killing. If the AI says "Vendor A commits to 99.9% uptime" and you build a recommendation around it, but the actual document says "we will use commercially reasonable efforts" — your CFO loses trust in the next AI report, your lawyer reads every PDF themselves anyway, and the time savings evaporate.
Two ways to solve this. The wrong way: more careful prompts and stronger models. They reduce hallucination but never eliminate it. The right way: grounding. Force the AI to surface the literal text it used, alongside its claim. The user can spot-check, the AI knows it's being audited, and the workflow stays fast.
What grounded AI looks like in practice
In POCsheet's comparison reports, every cell in the comparison table now carries a small info icon. Hover or tap it and a popover shows:
- The verbatim quote from the source PDF (≤ 300 characters).
- The document name the quote came from.
- Original casing, punctuation and language — preserved.
If the AI says "Vendor A: SLA 99.9% with 10% credit on breach", the popover shows the exact sentence — usually from Section 4 of the Service Level exhibit. If the AI flagged a red flag ("🚨 ALERTA: liability cap below industry norm"), the popover shows the exact clause the AI judged as below norm. The user verifies in three seconds and moves on.
Why this is harder than it sounds
Asking an LLM to "include the source quote" naively often produces paraphrases or near-quotes. The implementation that works requires three constraints in the prompt:
- Verbatim, not paraphrased. The snippet must be copy-paste from the source. If the AI rephrases, the verification breaks.
- Original language preserved. The rest of the report can be translated (English vendor PDFs into Spanish reports, say), but the snippet stays in the document's original language so the user recognises it.
- Hard length limit. Long snippets defeat the "glance to verify" UX. Cap at ~300 chars; if more context is needed, the user opens the source PDF.
Each one of these is a separate sentence in the system prompt and each one matters. POCsheet's snippet contract is enforced for every cell in every report.
Three workflows source citations unlock
1. Legal review takes 4× less time. Counsel reads the AI-aligned table, spot-checks the snippets on the 4-5 clauses that matter most, and signs off. They don't re-read 80 pages of MSA.
2. CFO presentation has audit trail. When the recommendation slide shows "GCP is 22 % cheaper", the underlying snippet ("...3-year commit price USD 1.10M / year, see Pricing Annex §2.1...") is one click away. Nobody disputes the number.
3. Vendor pushback becomes evidence-based. If a vendor objects to a flagged red flag ("we don't have auto-renewal without notice"), you reply with the snippet from their own document. Conversation ends.
The bigger pattern
Source citations aren't a procurement-specific feature; they're how grounded AI should work everywhere a human is making a decision based on the AI's output. Legal review, medical diagnosis, financial analysis — same principle. The AI surfaces the candidate answer plus the evidence; the human keeps the judgement. POCsheet ships this in every comparison run on the platform from v3.6 onward; older comparisons gracefully omit the affordance.