How to Define Activation for an AI Product

You define an activation metric by picking the first action a new user takes that predicts they will still be using the product a month later, then writing it down as one sentence with a count and a time limit. For an AI product, that action has to be the user doing something with an output, such as sending it, editing it or pasting it somewhere. Seeing an output is not enough.

This matters because activation is the number that everything else in a self-serve product hangs from. Free-to-paid, payback and the retention curve an investor asks for all sit downstream of it. Most teams set the bar one step too early. They count a setup screen or a first result, get a flattering number, and then spend a quarter trying to fix a conversion problem the metric was built to hide.

The tile that says Activated: sixty-one percent

Picture an AI meeting-notes tool. You sign up, and the first screen has one button: Connect Google Calendar. The next screen says Notetaker will join your next meeting. After that call ends, an email arrives with the subject Your summary for Weekly sync is ready and a button reading View summary.

The summary page is good. It has a three-line recap, a list headed Action items (4), and two buttons under the list: Send to #product and Copy to Linear. There is also an Edit link beside each action item, because the model sometimes gives a task to the wrong person.

On the team's internal dashboard, the first tile reads Activated: sixty-one percent. In this example, activated means the calendar is connected. The number looks healthy. What the tile does not show is that only fourteen percent of new accounts ever open a summary a second time, and fewer than one in ten have ever pressed Send to #product.

Now the founder's side of it. He puts sixty-one percent in the monthly investor update, three months running. The growth hire ships a shorter calendar permission screen and the tile moves to sixty-six. Paid conversion stays in low single digits. So he opens a pricing experiment, because pricing is the only lever left that the dashboard points at. Nobody on the team has read the definition behind the tile since the week it was built.

Three events you could call activation

Almost every AI product has the same three candidates, and they sit one after another in the flow.

The first is setup complete. The calendar is connected, the workspace is named, the data source is linked. It is easy to count and easy to move, which is why it ends up on the dashboard. It tells you the user was willing to try. It does not tell you the product did anything for them.

The second is output received. The first summary was generated, or the first report was opened. This feels like value, and most advice would stop here and call it the aha moment. For a traditional app that is often fine, because the user had to build the thing they are looking at. In our notes tool, the summary gets made whether anyone asks for it or not.

The third is output acted on. The user sent the summary to a channel, pushed an action item into their tracker, or fixed the owner on a task. Each of those means they read it, trusted it enough to put their name next to it, and fitted it into how their team already works. That is the event worth calling activation.

Why seeing an AI output proves almost nothing

Here is the claim most activation guides would push back on. In an AI product, output received is a vanity event, and it should not be your activation metric even when it correlates with retention a little.

Two reasons. First, the product often creates the output on its own. A notetaker joins the call, a research agent runs overnight, a coding tool writes a suggestion before the user asked. Counting that as activation is counting your own server doing work. Second, AI output can be wrong in ways the user spots in seconds. Someone who opens a summary, sees their own task given to a colleague, and closes the tab has received an output. They have also just learned not to trust it.

Acting on the output is the only event that shows the user believed it. Slack is the well-known version of this idea. Stewart Butterfield told First Round Review that a team which had exchanged 2,000 messages had really tried Slack, and that 93% of those teams were still using it. Slack did not count signups or channels created. It counted people using the product with each other, over and over. Your version is the output leaving your product and doing work somewhere else.

What an early definition costs you

A definition set one step too early does not just flatter the dashboard. It points the whole team at the wrong screens. If activation is calendar connected, every onboarding change gets judged on the permission screen, and nobody measures whether people trust the first summary. The work that would move retention never gets scheduled, because the metric says the problem is already solved.

The money side is sharper for an AI company. Every passive account still costs inference. The notetaker joined the calls, transcribed them and wrote summaries nobody used, and all of it landed in cost of goods. AI product builders average around 52 percent gross margin against 70 to 80 percent for traditional software, so there is far less room to carry users who never act.

There is an upside hiding in the same numbers. Measured on SaaS and product-led companies, a ten point improvement in activation typically drives a 15 to 25 percent increase in free-to-paid conversion. You can only get that lift if the activation number measures something real. Ten points on calendar connected buys you nothing.

The week-4 test for any candidate event

You do not need a data team for this. You need an export of events for signups from at least two months ago, so every user has had a full month to stick or leave. Then do this, once per candidate event.

• Define retained first. For our notes tool, retained means the user acted on at least one summary in their fourth week. Do not use logged in, because the email brings people back even when the product is not working for them.

• List three to five candidate events from the first seven days: calendar connected, first summary opened, summary sent to a channel, action item copied to a tracker, action item edited.

• For each event, split the cohort into users who did it and users who did not. Write down the share of each group that was retained in week four.

• Pick the event with the widest gap between the two groups. If almost everyone does it, like calendar connected, the gap will be small and it will not tell you much.

• Check coverage the other way. Of the users who were retained, what share did the event early on? If it is low, the event is too narrow and you are missing another path to value.

• Add a count if one action is not enough. Sent two summaries in the first seven days may split retained from gone far better than sent one. Test two or three counts and keep the smallest one that holds up.

Expect the winning event to produce a smaller number than your current tile. In this example, activation might fall from sixty-one percent to something in the teens. That is the right result. Product-led benchmarks put activation around 20 to 40 percent for most products, and 40 to 60 for the best, so a new number below that range is a problem you can now see and work on.

Write it down in one sentence, with what does not count

A definition that lives in one analyst's SQL query will drift. Someone will widen it to hit a goal, and the tile will creep back up. Write it as one sentence the whole team can read, using this shape: a new account is activated when a person on it does a specific thing with an output, a set number of times, within a set number of days of signup.

For the notes tool that becomes: a new account is activated when someone sends a summary to a channel or copies an action item into a tracker, at least twice, within seven days of signup. Underneath it, list what does not count: connecting a calendar, opening a summary, and the summary email being delivered.

Then rename the old tile rather than deleting it. Setup complete is still a useful number for the permission screen. It just stops pretending to be activation. Put both in the investor update, and say which one you changed and why. A founder who can explain why his activation number went down usually gets asked better questions about it.

Once the event is chosen, the design work is getting more people to it faster, which is what our guide to designing onboarding toward an activation event covers. What happens after it, in weeks two to four, is the subject of why users leave after they activate.

Why you are the last person who can see this

The definition drifts early for a reason. You know what a good summary looks like, so when you watch one appear, you already know what you would do with it. It feels finished. A new user sees the same page and still has to decide whether to believe it, where to send it and whether it will embarrass them in front of their team. The step you count as done is the step they have not taken yet.

Studio Maydit designs the screens where that decision happens: the first output, the buttons under it, and the path from a result someone reads to a result they use. We work with AI and SaaS founders in the US, UK and Europe, on fixed-scope projects of three to four weeks that end with a diagnosis of where the product is leaking, and our AI product design work starts from the event you are trying to move. If your activation tile looks healthy and your paid conversion does not, book a 30-minute call with Studio Maydit.