TL;DR
HR, legal, or security dropped a 40-page policy update in the channel at 4pm. You need to tell the team — before end of day — what actually changed, who has to act, what dates matter, and which 25 pages are restated theater. Paste the full text (not the PDF link), give the model your role and the previous version, and ask for deltas and required actions instead of “a summary.” A 40-page policy is roughly 20,000-30,000 words, which fits comfortably in Claude (1M-token window on paid plans) or Gemini 3.1 Pro (1M), but can silently truncate in the ChatGPT app (~32K-token / ~24,000-word ceiling on Plus). The brief is triage, not legal advice — flag ambiguous clauses with section numbers and route them to whoever owns the policy.
The task
A 40-page policy update lands in the team channel at 4pm. The team needs to know — by end of day — what actually changed, who on the team has to do something, what dates matter, and which pages are restated theater. You don’t have time to read it cover to cover, but you can’t get it wrong either, because compliance has teeth.
Where AI helps — and where it does not
AI is genuinely strong at three things here: structural summarization of long repetitive legal prose, pulling concrete “you must do X” requirements out of hedged language, and diffing two versions when both are pasted in full. It is good at separating informational boilerplate from binding clauses.
What AI cannot do: render a compliance verdict, decide whether your specific edge case is covered, or pick which clauses to escalate. It will also miss carve-outs buried in footnotes or annexes unless you explicitly tell it to look. Surface uncertain clauses with section numbers and route those to the policy owner. Your job is triage, not interpretation.
One specific, repeatable failure mode: the model tends to copy the policy’s style into the summary — passive voice, “shall,” hedged qualifiers — which makes the brief feel as bureaucratic as the original. Force plain English with active verbs in the prompt, or you get a 40-page document compressed into one page of equally unreadable legalese.
Pick a tool that can actually hold the whole document
The single biggest mistake is feeding a 40-page policy to a model whose effective context can’t hold it, then trusting a summary built on a truncated read. As of June 2026, here is what fits:
| Tool (June 2026) | Effective context for one document | Per-file cap | Best for |
|---|---|---|---|
| Claude (Pro $20 / Max $100+) | 1M tokens on paid plans (~500K in web chat, full 1M via Claude Code/API) | 30 MB / file, up to 20 files per chat | Long policies, version diffs, citing section numbers |
| Gemini 3.1 Pro (Google AI Pro $19.99) | 1M tokens | 100 MB / file | Big PDFs, side-by-side annexes |
| ChatGPT Plus ($20, GPT-5.5) | ~32K tokens in-app (~24,000 words); full 1M only on $200 Pro | 512 MB / file (but text capped ~2M tokens) | Short policies; risky for 40+ pages without splitting |
| NotebookLM (free / AI Pro) | Source-grounded RAG, up to 50 sources per notebook | per-source | Per-paragraph citations, repeated querying, audit trail |
A 40-page policy is typically 20,000-30,000 words. That clears Claude’s and Gemini’s 1M-token windows with room to spare, but it can exceed the ChatGPT app’s in-app ceiling (~32K tokens, roughly 24,000 words on Plus as of June 2026 — the API supports far more, the consumer app does not). If you must use the ChatGPT app on a long policy, split it by section (covered in the FAQ). For a policy you’ll re-query for weeks, NotebookLM grounds every answer in a specific paragraph with a clickable citation, which is the closest thing to an audit trail.
Tool guides on this site: Claude Projects (best for a policy you’ll reference repeatedly) and NotebookLM getting started.
What to feed the AI
- The full policy text — paste the text, not just the PDF link. Most consumer chat apps don’t reliably fetch a URL, and an upload can silently truncate past the context limit.
- The previous version, if it exists. Diffing finds 90% of what matters.
- Your role or function: “backend engineering team of 6,” “field sales contractors.”
- Geography / jurisdiction. US-only policies skip clauses; EU/UK add some.
- The 2-3 specific concerns you walked in with (data residency, contractor classification). These become explicit “answer this” line items.
- Any deadlines you already know — effective dates the policy probably implements.
- One example of how your team writes briefs internally, for tone.
Copy-ready prompt
Replace each [bracketed] placeholder with your own text before sending.
Read this policy. Produce a one-page brief for someone in [role] working in [function/geo].
Policy (current): [full text]
Previous version (if any): [prev]
Concerns to explicitly address: [2-3]
Return, in plain active English (no legal hedge):
1) The 3-5 things that genuinely changed vs the previous version — cite section numbers.
2) New required actions for [role], with deadlines and owners.
3) Clauses that are ambiguous to [role] — quote them and mark "needs confirmation."
4) Sections that are restated boilerplate [role] can skip — list section numbers only.
5) Specific answers to my 2-3 concerns, with section references.
6) One sentence on overall impact ("low / moderate / structural").
Mark anything you are uncertain about with [UNCERTAIN] and explain why.Shorter variant — Slack-ready 5-bullet TL;DR
Read this policy. Output exactly 5 bullets a team lead can paste in Slack:
- Bullet 1: what changed (one sentence)
- Bullet 2: what we have to do (with date)
- Bullet 3: what's ambiguous (with section ref)
- Bullet 4: what's NOT changing despite the noise
- Bullet 5: who to ping for questions
[paste policy]Sample output
A useful change-summary line: “Only one thing changed for engineering: starting Jan 15 (§4.2.a) you must record purpose-of-use in the access log when querying production data containing PII. Pre-existing access patterns are grandfathered for 30 days. Everything in §1, §2, §3 and §5 is restated 2024 language.”
A useful uncertainty flag: “[UNCERTAIN] §4.2.b reads ‘sensitive operational data,’ but the policy never defines that term. Under last year’s definition it included staging dashboards; the new glossary in Annex C is narrower. Confirm with the policy owner before assuming staging is exempt.”
How to refine
- Force the diff stance: “List only deltas from the previous version. Anything restated, even with reworded language, goes under ‘theater section numbers, skipped.’ Do not paraphrase.”
- Cite or it didn’t happen: “Every claim about the policy must reference a section number. If you can’t cite the section, drop the claim.”
- De-jargon: “Rewrite the brief in plain active English. Replace ‘shall’ with ‘must,’ delete hedging adverbs, use the role’s actual job verbs.”
- Press on ambiguity: “List every clause where a reasonable person in
[role]could read it two different ways. Quote both readings.” - Stress-test the brief: “Now imagine I’m compliance auditing this team in 6 months. Which 2 items in your brief would I question first?”
Common mistakes
- Trusting an AI summary as legal advice. The brief is triage; the policy itself is authoritative.
- Skipping the ambiguity section. That’s exactly where compliance gotchas live.
- Not citing section numbers. When someone asks “where does it say that,” you need to point.
- Forgetting the previous version. Without the diff, you re-summarize unchanged material as “new.”
- Letting AI default to legal-style prose. The audience is a busy individual contributor, not a partner at a law firm.
- Sharing the brief without marking it informal. Colleagues will quote your bullets in formal audits.
- Asking for “the summary” instead of “the deltas and required actions.” Different document, different value.
- Pasting only the PDF link and expecting the model to fetch it. Paste the actual text, or use a tool that extracts it.
- Uploading a 40-page PDF to the ChatGPT app and trusting the result without checking it read the whole thing — the in-app context can truncate past ~24,000 words on Plus.
FAQ
- Which AI handles a 40-page policy best? As of June 2026, Claude (1M-token context on paid plans) and Gemini 3.1 Pro (1M) hold a 40-page document — about 20,000-30,000 words — in one pass. The ChatGPT app caps in-app context near 32K tokens (~24,000 words) on Plus, so a long policy can truncate; use the section-split workflow below or move to Claude/Gemini.
- Can I share this summary with my whole team? Yes, but mark it “informal summary; the policy itself is authoritative” and include the policy link plus your name as the summarizer. People will quote it.
- The document is too long for one prompt — what do I do? Split it by section, summarize each with section numbers preserved, then ask the model to merge per-section summaries into a single delta list. Keep the merge prompt simple: “consolidate these section summaries; do not re-summarize.” This matters most in the ChatGPT app; on Claude or Gemini you can usually paste the whole thing.
- How do I handle a policy I can’t paste because it’s confidential? Use your company’s approved AI tooling — most enterprise contracts permit internal text. If unsure, summarize those specific sections manually and let AI handle the public boilerplate. Never paste confidential text into a free consumer account.
- The policy keeps updating mid-review — how do I keep the brief fresh? Add a “version + retrieved date” line at the top of your brief. Re-diff weekly until the policy stabilizes. A Claude Project or NotebookLM notebook lets you swap the source file and re-run the same prompt.
- What if a section conflicts with another policy? Flag it in the brief with both citations. Don’t try to resolve it — that’s the policy owner’s job.
Related
- AI PDF summary
- AI executive summary
- Policy Summary Prompts for One-Page Internal Policies
- Summarize a Textbook Chapter With AI
- Read Out a Funnel Analysis With AI
- NotebookLM getting started
Tags: #AI writing #Office #Workflow #Policy #PDF