Every issue after this one is a self-contained update kit: a regulation, the verbatim text, a copy-paste block, and the prompt to bring your P&P current. But before any of that, there is a meta-skill worth ten minutes: knowing which compliance jobs your AI can actually do, and which tool fits which job. Get this right and the rest is mechanical. Get it wrong and you will either waste a license you do not need or trust an answer the tool was never able to give.
Here is the whole map.
The one fact that runs everything: read is not rewrite
Microsoft caps Copilot’s rewrite at about 3,000 words — roughly 6 pages — and notes quality drops past that. But Copilot can read and summarize a document set up to about 300 pages. Those are very different numbers, and the gap is the key to everything.
- A job that reads (find every place AI is used in my P&Ps; which sections does this rule touch; where are my gaps): Copilot handles it, even across a big policy library.
- A job that rewrites (redline this whole P&P against the new rule): Copilot tops out around 6 pages before it starts dropping sections or inventing text.
So the question is never just “how big is my P&P.” It is “how much am I asking the tool to rewrite in one pass.”
How much each plan can redline in one pass
| Plan | Whole-P&P redline, one pass |
|---|---|
| Copilot (free or M365) | up to ~6 pages |
| Claude Pro / Team | up to ~300 pages (200K tokens) |
| Claude Enterprise | up to ~750 pages (500K tokens) |
These are one-pass limits. A bigger P&P can still be done on a smaller plan section by section — read with Copilot to find the sections, then redline each. The table is about effort and cleanliness, not a hard wall.
What you can do in Copilot today
Of a full library of ~50 mortgage compliance programs, these short ones you can redline whole, right now, in the Copilot you already have (each is ~6 pages or less):
AVM governance · TCPA communications · Reg-N advertising · Red Flags / identity theft · periodic statements · FHLMC loss-mit operations · e-sign disclosures · ARM rate-change notices · GLBA privacy · adverse action (Reg B) · UDAAP.
Most of the rest — about 39 programs, from 6 to 57 pages — want Claude for a clean one-pass redline (or section-by-section in Copilot). And the very biggest job, reviewing your entire policy manual at once — all ~50 programs, ~700+ pages — is the Claude Enterprise case. No single policy needs Enterprise; the whole-shop review does. That is the exam-readiness job.
The decision rule
Under ~6 pages of rewrite, use what you have. Over it, that is where you graduate to Claude. Simple as that.
And three ways to feed your P&P so you stay under the line when you can:
- Net-new rule (a brand-new requirement): nothing to compare against, just draft the new section. Any tool.
- Focused change (revises one section): pull out that section and redline it, not the whole manual. Any tool.
- Broad rule (touches many policies, like an AI-governance mandate): run an inventory first — let the AI read your library to find every affected policy — then fix each. The reading finds the list; the fixing is small per policy.
The safety layer (this is what makes it defensible)
None of this is safe because the tool is smart. It is safe because of discipline you bring:
- Quote-only prompts: “use only the regulation text provided; do not invent or paraphrase.” That is the difference between a grounded redline and a confident hallucination.
- Section by section on anything large, so nothing gets skipped.
- Human review, always. The output is a draft for a compliance officer, never the final word.
That discipline is baked into every kit we send.
What’s coming
Each issue from here is one regulation, done: the verbatim text, the copy-paste block, the “how to feed your P&P” instruction, and a note on which tool the job needs. Next up: the FNMA AI Lender Letter (LL-2026-04), which takes effect around August 6 — a broad one, so a good first test of the inventory-then-fix method.
Subscribe and you will get each kit as it ships.