Agentic Onboarding
What changes when your new users are agents
Head of Growth at Amplitude
Sept 15, 2026
Who am I
Tommy Keeley is Head of Growth at Amplitude, where his team drives AI adoption & builds agentic capabilities into products like Slack & Teams. Before Amplitude he spent four years driving growth at Dropbox, including Principal PM for Sharing & Virality & interim Head of Growth. Prior to Dropbox, Tommy worked at high growth startups including Attentive & AlphaSense.
Agenda
- 1Onboarding 101
- 2Acquisitionfind › pick › signup
- 3Activationsetup › aha › habit
- 4Demo
- 5Conclusionwhat means success
Introduction
The Growth Funnel
Acquisition
How new users discover and enter your product. It is top of funnel. Not all channels are equal, some bring intent, others bring volume.
Activation
Activation is the moment a user experiences real product value for the first time.
Engagement
Engagement is how deeply, broadly, & often users keep using your product after activation. The behaviors users show in their first days & weeks are the strongest predictors of churn.
Monetization
Monetization is turning what users find valuable about your product into revenue. It is the outcome of getting acquisition, activation, and engagement right, not a separate strategy bolted on at the end.
Agentic Acquisition
Find › Pick › Signup
Your docs are your homepage now.
The question used to be "can a person find our website?"
Now it's, "can an agent find updated and accurate documentation?"
Are you in the candidate set?
- Web search rank and citationsRank for the task on Bing and Brave, not just Google.
- Docs, READMEs, registries, reposPlain text, listed where it looks.
- Robots and rate limitsGPTBot and ClaudeBot not blocked.
- What it already knowsSometimes it never even searches. It picks what it remembers from training, or what is already installed in the repo.
Mintlify, 2026, share of documentation requests across their hosted docs sites, agent user-agents vs browsers.
Agentic Acquisition
Find › Pick › Signup
Find puts you in the set. Pick is whether you win the comparison.
Once the agent has options it reads the same things about each one and picks the one it can use.
Does your product get picked?
- It can tell what you do.Same name, same description, same category everywhere it looks. If sources disagree, it moves on.
- It can tell how to use you.Step-by-step instructions it can follow, not a page that describes you. Every step you leave out, it guesses.
- It can tell you're alive.Recent updates, a package published this month, questions that got answered. Stale reads as abandoned.
- It can try you.From your docs alone, one call, one real result. That call is the interview.
- Says what it does
- Says how to use it
- Shows it is alive
- Lets you try it
- Says what it does
- Says how to use it
- Shows it is alive
- Lets you try it
- Says what it does
- Says how to use it
- Shows it is alive
- Lets you try it
The agent version is pick rate: when an agent has options, how often it chooses you.
Agentic Acquisition
Find › Pick › Signup
Signup is a credential, not a form.
A human types an email. An agent gets a scoped key it can use with nobody watching.
Account, project and key by API.
A screen is where the agent stops.
A temporary key on request. A quick math puzzle and a time limit replace the CAPTCHA.
Claim it later: terms, billing, ownership.
One job, one project, killable without a password reset. That's what lets a customer hand a key to an agent at all.
Least privilege is what makes "yes" safe.
A dry-run call the agent runs right after.
The step teams skip, so the agent reports done when nothing works.
Notice the order. The agent works first. The human signs up after, by claiming what already runs.
Test keys, no account.
7 daysstripe sandbox create --non-interactive
Project credentials in one call.
72 hoursneon claim createTemp account, token, live URL.
60 minuteswrangler deploy --temporaryAgentic Activation
Setup › Aha › Habit
Setup means your data is in the product.
For Amplitude that's the first event landing. An agent can install the SDK and send the first events on its own. It just needs three things.
Not a docs page it has to interpret. One versioned set of steps: install this, call this, name events like this.
Handed to it, not chosen from a dropdown. If it has to pick, it picks wrong.
One call that answers "did events arrive?" with a number, not a green checkmark.
The bar is not fewer clicks. It is less human work. If the first event never lands, the agent moves to the next product and never tells you.
Agentic Activation
Setup › Aha › Habit
The aha is the first meaningful thing done with the data.
Setup got the events in. Aha is what the agent does next: answer a question, build the first chart, and leave something behind that pulls someone back.
The first chart, the first funnel, the first answer to "what happened this week." Data sitting in a project isn't value. A question answered is.
A human's aha builds over a week. An agent's is one attempt. If the first meaningful action fails, it picks the next product and never comes back.
A weekly digest to the human's inbox. An alert when a metric moves. A scheduled check the agent runs itself. The aha that lasts is the one with a reason to return.
The best agent setups end by scheduling their own next visit.
Agentic Activation
Setup › Aha › Habit
Habit is a trigger, not a feeling.
A human comes back because the product is part of their week. An agent comes back because something called it.
A recurring job: the weekly report, the nightly sync, the Monday audit.
A webhook or an event: a new deploy, a new customer, a threshold crossed.
A human asks for the next thing, and the agent already has the setup and the credentials.
Design the trigger, or there is no habit.
Demo
Conclusion
Measure
Same funnel, new numbers.
The old metric is already in your dashboard. The new one has no old-world twin.
| Stage | Old metric | Agent metric | Why it matters |
|---|---|---|---|
| Find | Organic traffic | Agent share of docs traffic | Tells you who is actually reading |
| Pick | Share of voice | Agent pick rate | Training data is the new search results page |
| Signup | Signup conversion | Hands-off completion rate | Zero human steps, or it isn't agentic |
| Activate | Time to aha | Time to first trusted outcome, verification rate | "It ran" is not "it worked" |
| All of it | A/B test on the page | Conversion by instruction version | Your docs are the experiment now |
No numbers on this table on purpose. The agent metrics are definitions, not results. Setup completion is not activation and a first signal is not value: compare retention of completers to non-completers before claiming anything.
Conclusion
Homework
Point an agent at your own docs this week.
One hour. You will learn more about your funnel than a quarter of dashboards.
Open a coding agent. Ask it to implement your product end to end in a scratch repo. Do not help.
Run it again with a different model, a different harness, a different way of asking. The gaps that show up every time are the real ones.
What did you try? Where did you look? What would have had to be true for you to pick us? It will tell you. Nobody else will.
Every stall, guess, and quiet wrong turn. That list is your onboarding roadmap, and it's more honest than any survey you'll run this year.
Thank you
Appendix
A quarter of setup in one session
Point an agent at a scratch repo. Watch it find the path, run the skill, verify, and hand a human a dashboard. Messy legacy codebases still need people. The interesting part is where the work moved.
Simulated session for rehearsal. On the night this panel is the cue to switch to the live screen.