cairn
About

We think the documentation problem and the support problem are the same problem.

Nobody has treated them that way, so every team in the world buys two tools that don't talk, and writes the same answer twice.

01 — The thesis

One problem, sold as two markets.

A customer gets stuck. Somewhere in your company, someone already knows the answer — they've typed it into a ticket eleven times this quarter. That knowledge exists. It's just trapped in a thread nobody can search, in a screen recording nobody can Ctrl-F, or in a wiki page that was true nine months ago.

The market's answer is two purchases. A capture tool that turns a workflow into a document and stops. A support suite that routes conversations and expects you to have written the content already. Between them sits an integration you maintain, and a gap where the actual value was.

Every competitor solves one link in the chain and hands you the integration bill for the rest. We wanted to build the chain.

So Cairn records the work once and produces all four artifacts from that single capture: the step-by-step guide, the branded help center article, retrievable knowledge for an AI agent, and an on-screen walkthrough that can spotlight the real button on the real page. One data model, no glue.

And because it's one model, the value can travel backwards. When the agent can't answer, that miss becomes a documentation work order with a named owner. When support resolves something hard, the resolution becomes a draft article. Usage improves the asset instead of wearing it out — which, as far as we can tell, is not true of anything else in this category.

02 — Why now

This was a good idea five years ago and it would not have worked.

Three things changed, and all three had to.

01Retrieval got good enough to citeA grounded answer that names its sources — and declines when it can’t find them — became buildable rather than aspirational. That single behaviour is the difference between a support agent and a liability.
02Capture got good enough to generateRecording a workflow now yields structured data, not a video: element roles, accessible names, ranked selector candidates, per-step URLs. Content can be generated from work instead of written separately from it.
03The category split hardenedDocumentation tools and support suites both matured into good products that solve half the problem each. The seam between them stopped being a gap and became an accepted cost. It shouldn’t be.

An AI agent is only as good as the knowledge behind it. That's why we built the knowledge machine first, and bolted the agent onto the end of it — rather than shipping a chatbot and hoping someone would feed it.

03 — What we refuse to do

The commitments are more useful than the values.

Anyone can claim to be customer-obsessed. These are things we've given up the ability to do, which is a different kind of statement.

No. 01No pricing you need a call to understandAI is metered per operation with the cost written down: about 5.5¢ for an answered conversation, 1¢ for a copilot draft. There is an estimator on the pricing page that converts your volume into a plan without talking to anyone.
No. 02No invented bonus tiers on creditsTop-ups are credited one-for-one — $10 buys $10. No "spend more, get 15% extra" ladder engineered to make you overbuy, and a hard cooldown on auto-recharge so a runaway loop cannot quietly drain a card.
No. 03No capturing what you typedCapture records the label of a field — "Card number" — and never the value entered into it. This is an architectural rule with tests behind it, not a checkbox someone can turn off during onboarding.
No. 04No AI that guesses to look helpfulBelow its confidence threshold the agent hands off to a human with full context and logs a knowledge gap. We would rather it say "I don’t know" than invent a step. The refusal behaviour is regression-tested.
No. 05No claim we cannot defend todayNo customer counts, no logo walls, no deflection percentages, no benchmarks we haven’t run. Where a number would help and doesn’t exist, we describe the mechanism instead — it’s checkable, which makes it stronger.
No. 06No cheap shots at competitorsScribe documents a workflow well. Crisp and Intercom run a support desk well. We will say where we differ and concede where we don’t, and we will never publish a competitor’s pricing to win an argument.
04 — How we build

Assertions are cheap. These are enforced by the build.

Every line below is a test that fails the pipeline, not an intention.

Contract-firstAn OpenAPI 3.1 specification is the single source of truth for every HTTP payload. Client types are generated from it and CI fails on any drift between the spec and the code.
Tenancy by constructionEvery tenant-scoped model carries a workspace scope and every route requires a policy. CI fails on any route without policy coverage, and automated cross-tenant probes run against the suite.
AI safety, testedAn evaluation harness gates prompt and model changes on groundedness, refusal, and escalation regression. A change that makes the agent more confident and less correct does not ship.
Adversarial by defaultA red-team suite runs prompt-injection and cross-audience leak probes as part of CI, with a nightly scan. The internal runbook surfacing in a public answer is a build failure, not an incident report.
Deny-by-default egressThe crawler, HTTP tools and webhooks run behind a dedicated gateway that blocks private and reserved address ranges and cloud metadata endpoints — independently of application code and DNS timing.
Secrets, properlyPer-workspace integration credentials are envelope-encrypted. API keys and webhook secrets are hashed at rest. Enterprise workspaces can encrypt screenshot originals, never serve them, and delete them once a redacted derivative exists.
05 — Where we actually are

Pre-launch, and saying so.

You will not find a customer count on this site, or a logo wall, or a deflection percentage. Cairn has not launched. The product is substantially built and tested; the customer base does not exist yet, and pretending otherwise would undercut the one thing we're actually selling.

In progress — not yet achieved
SOC 2 Type I in progressHIPAA-readiness posture, EnterpriseEU data residency on requestReliability figures are internal targets

Dashed means roadmap. Our reliability figures are internal service objectives published for transparency, not a contractual SLA — and they'll say so on the page where they appear.

What is real todayEvery capability described anywhere on this site exists as working, tested code: seven channel adapters, four redaction tiers, the walkthrough runtime, the agent's grounding and refusal behaviour, the credit ledger, the public API and MCP server.
reservedThe people behind itThis slot is intentionally empty until there's something true to put in it. No stock headshots, no advisor logos.

Judge the argument by using it.

Fourteen days with every Scale feature on. No credit card, no sales call, no demo gate — the same refusals, applied to ourselves.

Start freeTalk to us