Introducing the Papyrus Copilot

Today we're shipping the Papyrus Copilot to all tenants on Business and Enterprise plans.

What it is

A conversational AI assistant — accessible from a floating button on every page — that can:

• Answer questions about your documents in plain English
• Summarise long documents on demand
• Compare two versions or two documents
• Surface status across workflows ("what's pending my approval?")
• Aggregate across the corpus ("our total exposure with [Vendor]?")
• Cite every claim back to source documents

Why we built it

Search returns documents. Reading returns understanding. There was always a gap — a friction-rich step where you'd search, find 12 documents, and have to read them all to answer the question you actually had.

The Copilot collapses that step. You ask the question. It does the reading. It tells you the answer, with links to verify.

How it works (briefly)

The Copilot uses Retrieval-Augmented Generation (RAG):

1. Your question is embedded (turned into a vector representing meaning)
2. Papyrus searches your tenant's documents semantically + by keyword for relevant context
3. The top-K most relevant chunks are passed (along with your question) to a large language model
4. The model synthesises an answer, citing sources
5. We post-process to verify citations resolve to real documents you have access to
6. Answer rendered with clickable citations

RBAC is sacred

The Copilot only retrieves documents you are permitted to see. Nothing else. The model never sees content you don't have access to.

This is non-negotiable. We tested extensively against scenarios where a clever prompt might surface restricted content. The model refuses; the retrieval stage refuses; both layers must pass for content to enter the context window.

Language support

English and Swahili at launch. Mixing the two within a conversation works fine. French and Arabic on the 2026 H2 roadmap.

What we're watching

We're shipping with metrics on:

• Question-answer relevance (human-graded sample weekly)
• Hallucination rate (does the answer claim something not in the cited sources?)
• Refusal rate (does the Copilot decline questions it should answer?)
• Time-to-answer
• Per-tenant adoption (% of weekly active users who use the Copilot at least once)

We'll publish quarterly accuracy benchmarks. Watch the Release Notes.

Pricing

Copilot interactions count toward your plan's “AI operations” quota:

• Free: not included
• Starter: 100 interactions/month
• Business: 1,000 interactions/month
• Enterprise: unlimited

Each interaction is a full Q+A exchange. Follow-up questions in the same conversation count as additional interactions.

What's next

The Copilot's next milestone is action execution — not just “tell me how to share this folder” but “share this folder with the marketing team for 7 days, watermarked, view-only”. We're staging that capability behind an opt-in flag and rolling it out tenant-by-tenant over the next quarter.

For now: open the Copilot panel, ask a question, see what it does. Feedback to feedback@papyrus.io.