For conversation-rich AI products

Your users are already telling you what to build.

Understand why users stay, why they leave, and what they wish your AI could do next.

Otis listens to conversations and combines them with real user behavior to surface the improvements that matter most.

Everything looked fine. Otis saw what really happened.

Example: AI support agent
What your stack saw
Resolved · 96% confidence
No ticket. No thumbs-down.
What the user did
Their last message: “How do I export all my data?” Then they never came back.

The agent closed the ticket. The user closed the account.

Example: AI coach
What your stack saw
Eval 9/10 · on-method · session time up
Engagement never dipped.
What the user did
Asked the same question four times across three sessions, then stopped opening the app.

It stayed on script. The user gave up getting an answer.

Example: AI copilot in a SaaS product
What your stack saw
Prompts per session up · copilot adoption climbing
Your healthiest-looking metric.
What the user did
Forty accounts asked it the same thing this month: “Can you just do this for all of them?” Then did it by hand when it couldn’t.

The copilot answered every prompt. The prompts were your roadmap.

What you can’t see is shaping your product.

Six signals that never look like signals.

Friction ►
Users rewrite and abandon what your AI produces, and nothing in your stack flags it.
Impact ※
You ship a prompt change, the eval moves, and you still can’t tell if behavior did.
Value ◈
Your best users go quietly stuck, and you find out at renewal, not before.
Discovery ◉
People use your product in ways you never designed for. That’s your next feature, and you can’t see it.
Churn ⠷
The greenest dashboard hides the users already on their way out.
Spend ◇
Rising token spend looks like engagement. Much of it is users retrying what didn’t work.

Your stack can tell you what ran. It can’t tell what delivered value.

That gap is where value is quietly won and lost.

Sees
Misses
Product analytics
SeesClicks, events, funnels
MissesWhat the user was actually trying to do
Observability
SeesSystem behavior
MissesWhether the user got any value out of it
Evals
SeesWhether the model passed a test
MissesWhether the user accepted, edited, retried, gave up, or wanted more
Tickets · Slack · Discord
SeesThe loud feedback
MissesSilent churn and the workarounds no one reports

Otis turns signal into ground truth.

Add Otis from your coding agent in minutes. First insights in hours.

# Claude Code or Cursor: npx @runotis/setup, then /otis-analyze
Setup

Your coding agent installs a light SDK and instruments your product surfaces; you review and merge. It runs async, so zero latency for your users, and there’s nothing to label or define. A short strategy onboarding, and you’re live. About 30 minutes.

The @runotis packages are private today: you’ll get access when you come on board. See the docs.

Privacy & trust

Otis knows what users do and say, not who they are. PII is redacted in the SDK and collector, before anything hits disk. SOC 2 in process, with HIPAA options for regulated teams.

See what your users are really trying to do.

Drop your email and pick a time with one of our cofounders: a look at how Otis works, and a conversation about what your users are actually doing.

Made in San Francisco for AI-native teams, by repeat founders who’ve built services from scratch and scaled them to 1B users.

Josh GreeneJosh GreeneTim MoretonTim MoretonAmy SlawsonAmy Slawson