Read the question
It resolves your words against your own event catalog: which events you meant, which properties, which date range, which segment. Ambiguity is surfaced, not guessed at.
signup_completed · source · last 7 days// ai agent
Ask a question, get an answer you can defend
Type it the way you would say it out loud. The agent works out which events answer the question, picks the right chart, runs the query, and shows you exactly what it counted.
How it gets to an answer
Four steps, and you can see the output of every one of them. The agent is a query writer with a chart library, not an oracle.
Choose the shape
A single number, a trend, a breakdown or a funnel — decided by the question, not by whichever chart you happened to open. Comparisons get a baseline automatically.
bar chart · grouped by sourceRun the query
The query runs against your workspace's events, over the full window, with no sampling unless you ask for it. Most answers come back in a couple of seconds.
1.2M rows scanned · 1.4sShow the work
The answer arrives with the events, filters, date range and row count that produced it — and a button to open the whole thing as a report you can keep.
open query · save as report“Shows its work” means something specific
Every answer has a panel behind it listing the events it touched, the filters it applied, the window it used and the number of rows it scanned. If a number looks wrong, you can find out why in about ten seconds instead of arguing about it in a meeting.
Because the query is right there, an answer can also be handed to someone else. Save it as a report, share the link, and the person who opens it sees the same definition you did — not a screenshot of a number whose origin nobody remembers.
It is also how you catch a misread question. If the agent resolved signups to signup_completed when you meant trial_started, the panel says so and you can correct it in a follow-up.
- Events
signup_completed- Filters
source is set,is_internal = false- Window
- Last 7 days, workspace timezone
- Grouped by
source, top 5 shown- Rows scanned
- 1,243,880 · no sampling
- Tokens used
- 412
What you can ask
Anything your events can answer. These five run in the console above — the answers below are the ones it gives, unedited.
Which sources drove signups?Checkout after the pricing changeActivation for users in IndiaWhere do users drop off?Anything unusual today?
Which sources drove the most signup_completed last week?
Web SDK led with 4,821 signups — up 18% week over week. iOS and Segment added another 3,140 between them.How did checkout_started trend after the May pricing change?
checkout_started climbed steadily after May 12, finishing the month 32% above the prior period.What is activation for users from India this month?
Activation for IN users is 41.3% — 4.2pp above your global average. Most activate within two days.Where do users drop off in onboarding?
The steepest drop is between Created project and Invited — 45% never invite anyone. Only 27.9% of signups finish activation.Anything unusual in the last 24 hours?
error_occurred spiked 6.4× at 02:10 UTC — 92% of it from Android SDK 4.2.1 on checkout. It settled by 05:00.Follow-ups keep the context, so “now split that by plan”, “only paying users” and “what about last month?” all work without restating the question.
Guardrails and data handling
An analytics agent is only useful if you can trust what it reached for. These are the limits it works inside.
- It sees one workspace at a time
- The agent queries the workspace you asked from, with your own permissions. It cannot reach another workspace of yours, another customer's data, or anything outside the events your SDKs sent.
- Answers are reproducible
- Every answer carries the query that produced it. Run it again over the same window and you get the same rows. The row count is always shown, and nothing is sampled unless you ask for a sample.
- It reads your properties, and nothing else
- The agent works from event names, property names and property values — the data you chose to send. There is no third-party enrichment, no access to your production database, and no request bodies. If you never send an email address, it never sees one.
- Everything is inspectable
- Open the query behind any answer, page through the underlying rows, export them as CSV, or save the answer as a scheduled report. There is no step you have to take on trust.
- It answers; it does not act
- The agent does not create, change or delete events, users, funnels or reports on its own. Anything it produces is saved only when you save it.
- Your data stays yours
- Your events are not used to train models that serve anyone else, and you can delete a workspace and everything in it at any time.
What a question costs
Questions are metered in AI tokens, not per seat — so the whole team can ask without anyone counting logins.
Per question200–800
tokens, depending on how much the agent has to resolve. A straight metric lookup sits near the bottom; a multi-step funnel with a segment split near the top.
Pro plan pool2M
tokens a month, included. Unused tokens do not roll over, and the pool resets on your billing date.
Which buys~4,400
questions a month at a typical 450-token average — about 145 a day across the whole team, every day of the month.
| Plan | Monthly token pool | Approximate questions |
|---|---|---|
| Free | 50K tokens / month | ~110 questions |
| Pro | 2M tokens / month | ~4,400 questions |
| Team | 10M tokens / month, pooled | ~22,000 questions |
| Enterprise | Custom pool | Sized to your team |
Estimates assume a 450-token average. Every answer shows the tokens it used, and top-up packs are available from workspace settings — see pricing for the full breakdown.
Ask your first question tonight
Connect an SDK, wait a few minutes for events to land, and ask the agent what changed. The free plan includes 50K tokens a month.
No credit card. Answers come with the query that produced them.