Your business already knows the answer.
It is just scattered across a policy register, four Google Docs, a Slack thread and one person who has been here nine years. Paragraph makes it one governed source — and delivers it, permission-aware and cited, to your staff, your customers, your partners and their AI agents.
THE HARD PART
One question. One source. Four different right answers.
A customer, an agent and an employee asking the same thing are not entitled to the same answer. Most tools solve this by keeping four copies. We solve it with one paragraph and permission-aware retrieval.
Can I get a refund on a transfer that failed four months ago?
Manual refund requests can be made up to 90 days after the transaction date, so a transfer from four months ago falls outside the standard window. Send us the transaction reference and we will have it reviewed rather than closed automatically — if the original delay was caused by us, that changes the outcome.
Refunds are always paid back to the account the money came from.
GROUNDED IN
- PGH-0088Refund policy
One paragraph matching this topic was excluded: PGH-0089, the internal exception procedure, is not granted to the customer audience. The customer is offered the remedy, not the rule.
RETRIEVAL SCOPE
- PGH-0088 Refund policyGranted to the customer audience.IN
- PGH-0089 Refund exceptionsInternal procedure. Not granted.OUT
- Live account recordSigned-in session, read at query time.IN
Same store, same question. The customer answer is softer because the paragraph that would harden it is not theirs to see.
WHY THIS EXISTS
Nobody is short of documents. Everybody is short of answers.
- 01 · DUPLICATION
Four copies of one policy
The website says one thing, the macro says another, the handbook says a third. Nobody is lying; nobody is in charge. Every AI you point at this inherits the contradiction and answers with confidence.
- 02 · TACIT KNOWLEDGE
The answer lives in one person
The real rule is in a Slack reply from 2024 and in the head of an agent with nine years of service. It is not written down anywhere, so it cannot be governed, audited or delegated.
- 03 · NO AUDIENCE MODEL
Publish it or hide it
Knowledge bases give you public and private. Real businesses need the same fact stated four ways, with the internal rationale readable by staff and invisible to everyone else.
- 04 · SILENT FAILURE
Nobody knows what you fail to explain
Questions your content cannot answer become tickets, churn and a support cost nobody attributes to documentation. The evidence exists in the logs and no tool looks at it.
THE PLATFORM
Know. Understand. Act.
- 01
Know
Ingest what you already have, split it into paragraphs, and give each one an owner, a version, a lifecycle and audience grants.
- Docs, SharePoint, Zendesk, Slack, web crawls
- Every paragraph has one canonical copy
- Conflict detection on every sync
- 02
Understand
Answer questions from that knowledge only, scoped to who is asking, cited to the version used, with confidence stated and refusal preferred over invention.
- Permission-aware retrieval, not a filtered prompt
- Combines paragraphs with live systems of record
- Refuses when nothing authorises an answer
- 03
Act
Let an agent do the thing the answer implies — check the policy, verify authority, confirm with the human, call the system, log it against the paragraph that permitted it.
- Explicit confirmation before any write
- Authority derived from paragraphs, not prompts
- Full action audit trail
KNOWLEDGE GAPS
The only tool that tells you what you fail to explain.
Every question your knowledge cannot answer is evidence. Paragraph compares question volume against retrieval confidence and resolution rate, names the topic you are silent on, and drafts the missing paragraph from the answers your own team has already given in Slack.
Nothing publishes itself. A named owner approves every paragraph and grants its audiences.
International transfers and FX
- 1,840
- questions
- +47%
- vs last month
- 0
- paragraphs covering it
- How do I send money to the UK?412
- Can I receive USD into my account?380
- What is your dollar exchange rate?291
- Do you support international transfers at all?244
78% of these conversations became support tickets. Six Slack threads already contain a consistent answer — assign an owner and it becomes one paragraph.
DELIVERY
Edit once. It changes everywhere in seconds.
There is no help-centre article, macro or FAQ page to keep in sync. Those are just surfaces rendering the paragraph.
AI help centre
A help site that answers questions instead of listing articles, generated from published paragraphs.
PUBLIC · CUSTOMER
Embedded widget
One script tag. Authenticated, so answers can safely include the customer’s own account context.
CUSTOMER
Support copilot
Grounded suggested replies with sources, including internal procedures customers cannot see.
AGENT
Employee assistant
Policy, process and the reasoning behind decisions, for the people who have to apply them.
EMPLOYEE
Partner portal
Commercial and integration knowledge, scoped per partner, without a separate documentation site.
PARTNER
Agent-readable API
Structured, cited answers for external AI agents that will soon ask on your customers’ behalf.
MACHINE
GOVERNANCE
Built for businesses that get audited.
Every answer names the paragraph and version behind it. Every action names the paragraph that authorised it. When two sources disagree, the AI refuses to arbitrate and raises it to a human instead of averaging them into a number nobody approved.
Citations, always
Every claim names the paragraph and version it came from.
Refusal by default
No authorised source, no answer. Gaps are raised, not filled.
Human approval
Nothing publishes itself. Owners approve and grant audiences.
Immutable audit log
Answers, edits, grants, refusals and actions, tenant-scoped.
Tenant isolation
Your knowledge is never pooled, and never trains a shared model.
Region pinning
Data residency in the region you choose, including Nigeria.
Reads what you already have
Sources sync continuously. Nothing is published from a sync without an owner’s approval.
- GOOGLE DOCS
- SHAREPOINT
- NOTION
- CONFLUENCE
- SLACK
- ZENDESK
- INTERCOM
- FRESHDESK
- WEB CRAWL
- S3
- REST API
- WEBHOOKS
FAQ
Questions, answered
It’s an AI-native business knowledge infrastructure. It turns fragmented organizational information — policies, docs, tickets, decisions — into a trusted, permission-aware intelligence layer that serves employees, customers, partners, and AI agents from a single source of truth.
Companies of roughly 20–500 employees with multiple SaaS tools, significant documentation, and real support volume — especially fintech, SaaS, technology, professional services, e-commerce, healthcare tech, and logistics. Every unanswered question costs money; Paragraph connects knowledge directly to support cost, productivity, and compliance.
A wiki stores documents; Paragraph is the layer underneath. Where a wiki expects people to browse, Paragraph structures knowledge into reusable units, serves it to the right audience with the right permissions, answers questions with citations, detects gaps and conflicts, and can execute authorized actions.
NotebookLM says “understand these sources.” Paragraph says “understand this business.” It adds governance, publishing, audience separation, customer question analysis, and an action layer on top of grounded AI answers — and it revolves around organizations, not notebooks.
Every answer is grounded in source material with citations and a confidence label — Confirmed, Likely, Conflicting, or Unknown. When evidence is insufficient, the system prefers to say “I don’t have enough information” rather than invent an answer. Conflicting sources are flagged for human resolution.
Retrieval is permission-aware: access is enforced by tenant, user, role, department, group, customer, account, resource, and audience. Internal knowledge can never leak into a customer response, and every tenant is strictly isolated.
Paragraph is in concept validation. We’re onboarding a small number of pilot companies across the roadmap phases — join the early access list and we’ll be in touch.
Bring one messy policy. We will show you the rest.
A 30-minute demo runs on your own document. You will see it split into paragraphs, granted to audiences, answered from, and — usually — contradicted by something else you already publish.