Fractional growth team or in-house hire: which fits?
| Option | What they do on activation | Fits when | What you still need |
|---|---|---|---|
| Fractional growth team (such as Kiko) | Write the activation definition per task type, instrument first runs, route failures to onboarding or evals, run tests | No agreed definition yet, or no one has time to run tests | Engineers to ship product changes; model owners for quality fixes |
| Fractional growth lead (one person) | Set the definition and priorities, manage whoever runs the work | You have builders and an analyst but no senior owner | People to instrument and test |
| In-house growth PM or product lead | Own activation long term | The definition is settled and the work is ongoing | Data support and time away from the roadmap |
| Data or ML engineer | Log task outcomes, build the eval suite from real failures | Quality, not onboarding, is the main failure | Someone to decide what counts as useful |
| Product-led growth or activation consultancy | Audit and recommend | You want an outside diagnosis, and your team will execute | Execution capacity |
Go Fractional lists a median of $175 an hour for fractional chief growth officers (source). Kiko engagements start with a 90-day growth sprint. Pricing starts at $10,000 a month and scales with the budget deployed and the number of channels run in parallel. You get a dedicated human growth lead plus agents. For a wider comparison, see fractional growth vs agency vs studio.
Why doesn't "signed up and sent a prompt" work as activation?
A user can send ten messages and get nothing usable, and a task can "complete" and still be wrong. Agents also vary more than fixed software, because they choose their own steps and tools. Anthropic describes this as the difference between workflows, which follow predefined code paths, and agents, which direct their own process (Building effective agents, 2024-12-19). The more the product decides for itself, the more your activation event has to measure the outcome, not the activity.
We use three moments from our activation method: setup (accounts connected, permissions granted), aha (the first reliable, useful result on a task the user cares about, which is the activation event), and habit (useful results repeated at the task's natural frequency).
What counts as a reliable, useful result, by task type?
Useful means the user could act on the result as it stands. Reliable means the same kind of task works again without a human taking over. Each row is Kiko's activation rubric.
| Task type | Useful result means | Reliable means | Observable signal | Not enough on its own |
|---|---|---|---|---|
| Email or inbox drafting | A draft the user sends | A second draft sent within 7 days | Draft sent with a small edit distance | Draft generated |
| Scheduling | A meeting booked at a time all parties accept | No reschedule caused by an agent error | Event created and not cancelled within 48 hours | Availability suggested |
| Research and reports | A report the user shares, exports or cites | A follow-up report is kept too | Export, share or copy | Report viewed |
| CRM and data entry | Records updated correctly | Spot-checked records stay unedited | Not corrected within 7 days | Records touched |
| Customer support replies | A reply sent without a rewrite | The ticket isn't reopened for the same issue | Sent unedited or lightly edited; no reopen | Suggestion shown |
| Phone and receptionist agents | The call handled to a booking or correct handoff | The owner doesn't override the outcome | Booking created; no callback complaint | Call answered |
| Browser and web tasks | The form submitted or purchase made as specified | No undo, refund or manual redo | Confirmation with no rollback | Task started |
| Documents, slides and sheets | A file the user keeps and sends | A second file is kept | Download or share, no full regeneration | File generated |
| Code changes | A change merged or deployed | Not reverted within 7 days | Merge or deploy, no revert | Pull request opened |
Set the window to the task's natural frequency: a daily inbox agent can use 3 days, a monthly reporting agent needs longer. Keep the definition only if it predicts retention. If activated users don't retain better, tighten "useful". Kiko's default windows: 3 days for daily tasks (inbox, support, scheduling, phone), 7 days for weekly tasks (research, CRM, browser tasks, documents, code) and 30 days for monthly tasks (reporting).
Where do first runs fail, and who fixes each failure?
Setup and first-request failures are onboarding problems. Wrong or unreliable results are quality problems that belong in your evals, so each failure becomes a test case rather than an anecdote.
| Failure point | Symptom | What to check | Who fixes it |
|---|---|---|---|
| Setup | Drop at the OAuth or permission screen | Scopes requested versus scopes the first task needs | Growth team: ask for less, later |
| Blank first request | Users open the product and never submit a task | Empty state; whether suggested tasks are shown | Growth team: suggested first tasks |
| Out-of-scope request | The first task is something the agent can't do | Top first-task intents versus supported tasks | Growth and marketing: fix the promise |
| Tool or integration error | Task fails on an API or browser step | Error logs by tool | Engineering |
| Wrong or unsafe result | Low grades, edits, rollbacks | Failed tasks by type | Model and eval owners |
| Slow result | Abandonment before the result arrives | Time to first result; progress shown | Product: partial results, progress updates |
| Good result, no action | Result usable but not sent or applied | The last-mile step (send, export, apply) | Growth team: the handoff action |
When you read cohorts, log the model and prompt version each cohort first used. A model change can move activation more than any onboarding change, and it will look like an onboarding win or loss if you don't log it. Also watch the traffic mix: paid traffic whose first tasks are out of scope lowers blended activation even though nothing broke.
Which interventions move activation?
| Intervention | Failure it targets | Primary metric | Guardrail |
|---|---|---|---|
| Suggested first tasks matched to the signup answer | Blank first request | Share attempting a task on day 0 | Activation of users who pick a suggestion |
| Ask for one integration first; request the rest when a task needs it | Setup drop at permissions | Setup completion | Task failures from missing access |
| A sample-data run before connecting real accounts | Fear of granting access | Time to first useful result | Share who then connect real accounts |
| A clarifying question on ambiguous requests | Wrong results on vague requests | Usable-result rate on first tasks | Time to first result |
| Show the plan; confirm before irreversible actions | Rollbacks and lost trust | Rollback rate | Task completion time |
| A human rescue for stalled high-intent signups | Stalls | Recovery rate | Cost per activated user |
| Match ad and creator promises to supported tasks | Out-of-scope first requests | Share of first tasks in scope | Click-through on the ads |
For acquisition that sets up the right first task, see demo ads for AI agents, creators for consumer AI app installs and the AI agent creator list.
When don't you need outside help?
- You already log task outcomes (sent, merged, kept, rolled back) per user, and your eval suite is fed by real failures.
- You're pre-launch with few users. Watch ten first runs live before you define anything.
Kiko has run programs for 70+ venture-backed startups since 2023, 9 of them AI agent products. In agent work, the first 90-day sprint delivers a per-task activation definition, first-run instrumentation, failure routing into evals and the first two onboarding tests. Our case study for a seed-stage AI agent startup reports week-one activation up from 22% to 41% in a 90-day sprint.
Send us your top three task types and your current activation definition → Book a call
Frequently asked questions
Who can help an AI agent startup improve activation and early retention?
A fractional growth team, which defines activation per task, instruments first runs and runs tests, or a fractional growth lead who sets direction for your own team. Hire in-house once the definition is settled. Whoever you choose needs access to task-outcome data and engineering time.
How should an AI product define activation?
As the first reliable, useful result on a task the user cares about, with criteria written per task type: a draft sent, a meeting booked, a change merged. Keep the definition only if users who reach it retain better than those who don't.
Should activation mean the first successful task or repeated weekly use?
Use both as separate milestones. The first reliable, useful result is activation; repeated useful results at the task's natural frequency are habit.
How is activation different from evals?
Evals test whether the agent produces good results on a defined task set. Activation measures whether new users reach a good result on their own first tasks. First-run failures should become eval cases; see our evals page.
What is a good activation rate for an AI agent?
Compare activation within the same task type and acquisition source, using the same definition. In Kiko's agent programs, 20–45% of new signups reach the first reliable, useful result within 7 days.
Sources
- Anthropic, Building effective agents
- Go Fractional, Fractional chief growth officer rates
- Kiko, Activation
- Kiko, Evals