The classic workflow when evaluating a vendor MSA: read the contract straight through (45–90 minutes), then return to it five times over the next week with specific questions ("what's the termination notice again?", "did they cap liability?", "is there a non-solicit?"). Each return-visit is 10–15 minutes of searching for one clause in 80 pages of legal text. AI chat over the contract collapses each of those return-visits to 20 seconds.
The shape of useful questions
When procurement and legal teams chat with an AI over a contract, the questions follow predictable patterns:
- Single-clause queries: "What's the liability cap?", "What governing law applies?", "Are there any non-compete clauses?"
- Comparative across vendors: "Which vendor has the strongest SLA?", "Whose price escalator is most aggressive?", "Compare the data-processing terms — who has the weakest commitments?"
- Risk surfacing: "Is there any lock-in risk?", "Are there any clauses that survive termination?", "Do they have unilateral amendment rights?"
- Summarisation: "Summarise the penalty clauses in plain English", "Give me the 5 most important obligations on us as the customer."
All four patterns benefit from the AI having read the full document already — which is what a vendor comparison tool does by default.
What separates good contract chat from bad
Three properties make AI contract chat trustworthy enough to actually use in a procurement workflow:
- Citations over assertions. Every answer should quote the verbatim text from the contract that supports it. Without citations, the AI's answer is just a guess.
- Comparative awareness. When the user asks "which vendor is better", the AI should structure the answer per-vendor with a one-line verdict at the end, not give a single fuzzy paragraph.
- Explicit unknowns. If the document doesn't address the question, the AI should say so — not invent an answer. "The contract does not specify a data residency commitment for the proposed workloads" is a useful answer; a made-up "Vendor commits to EU data residency" is dangerous.
The system prompt that drives the chat enforces all three. The AI is told that document text is untrusted data (defends against prompt-injection embedded in PDFs), that it can only answer from the documents, and that it must cite verbatim quotes when the wording matters.
How POCsheet's chat works
When a comparison finishes, a chat box appears below the report. The chat is grounded in a snapshot of the document text that persists alongside the comparison — the raw PDFs are still purged after 2 hours per the privacy policy, but the snapshot lives as long as the comparison itself.
The chat:
- Suggests 5 starter questions on first open ("Summarise the penalty clauses", "Which vendor has the strongest SLA?", etc.) so you don't face a blank box.
- Streams the response as the AI generates it, so you see progress.
- Quotes the source for comparative or specific-clause answers, with the vendor's document name.
- Is owner-only — never exposed via share links. The viewer of a shared comparison sees the static report, not the chat.
- Free plan: 10 questions per comparison. Pro: unlimited.
The pattern this fits into
AI chat over documents is the highest-leverage application of LLMs in procurement and legal — more than auto-drafting contracts, more than rewriting clauses. The reason is simple: the human stays in control of the decision. The AI does the "where is this clause and what does it say" work, which is 80 % of contract review time and 0 % of the value. The human keeps the judgement.
Combined with source citations and a negotiation playbook, AI chat closes the loop: extract → evaluate against your standards → answer follow-ups. That's the procurement workflow, made 5× faster, without losing any control.