A small product has one big advantage over a large company: it can change quickly. That advantage disappears if each change takes a month, or if nobody can tell whether the change helped.
This article explains how to run short build-measure-learn sprints with an AI team, what to measure, and, because it matters, which parts are automated today and which are still on our roadmap.
The short answer. Each week: look at three numbers, pick the single biggest drop-off, change one thing, ship it behind your approval, and measure again. An AI team can do the building, the measuring and the proposing; you do the deciding.
The loop in one paragraph
Build the smallest change that tests a guess. Measure what real people do with it. Learn whether the guess was right. Decide what to try next. That is the build-measure-learn loop from lean startup practice, and its point is speed of learning, not speed of building. A feature nobody uses is expensive however fast it was made.
Step 1: write the guess down before building
A sprint starts with a sentence in this form:
“We believe [change] will make [who] do [action] more often. We will know because [number] moves from [a] to [b] in a week.”
Example: “We believe a shorter sign-up form will make more visitors create an account. We will know because the share of visitors who finish sign-up goes from 12% to 18% in a week.” (The numbers are an illustration of the format, not data from any real product.)
Without the guess, every result can be explained away. With it, you either learn something or you do not.
Step 2: measure three numbers
For a small product, three numbers carry most of the signal:
- Visits. People who arrived. Tells you whether your message reaches anyone.
- Activation. People who did the one thing the product is for: booked a slot, made a first habit, sent a first request. Tells you whether the product delivers.
- Return. People who came back within a week. Tells you whether it is worth coming back to.
Read them in order. Low visits is a marketing problem. Good visits and low activation is a product or first-screen problem. Good activation and low return means the product is not yet a habit.
Measurement has to respect visitors. On products built on equipe, the analytics use no cookies and store no IP address, only a two-letter country and a random anonymous id kept in the visitor’s own browser, and events are deleted after 180 days. You get counts, not profiles. The team can also run simple A/B tests, splitting visitors between two or more variants and measuring a goal event, which is how you test the guess from step 1 without arguing about it.
Step 3: find the biggest drop-off
Lay the steps of your main flow in a row and look at where the most people disappear:
Visit → sees the offer → starts sign-up → finishes sign-up → does the main action → returns
The biggest drop is usually the best next sprint. Not the most interesting idea, not the feature a customer asked for last week, but the place where most people leave. This one habit prevents most wasted effort.
Step 4: change one thing
Change one thing per sprint when you can. If you change four and the number moves, you will not know which one mattered. Typical single changes:
- A clearer headline on the first screen
- One fewer field in the sign-up form
- A bigger, better-labeled main button
- A confirmation message that tells people what happens next
Small changes are cheap for an AI team: an engineer builds it, the designer checks it on phones, QA tries it the way customers would, and it waits for your approval before going live.
Step 5: let the team propose, and you decide
This is where an AI team differs from a tool that only builds what you type. With a team, a data analyst role reads usage, and the PM turns it into a proposal in plain words: what the data shows, what to change, and the expected cost. You answer yes, or not now.
An example of the card (illustrative, not a real customer’s data):
Proposal. Most visitors who start sign-up stop at the third field. Suggestion: remove it and ask later. Expected cost: a small change. Approve / Not now
You do not have to understand the code or the analytics tool. You have to be able to say whether the proposal fits what you want your product to be.
What is automated today, and what is not
We want to be exact about this.
- Works today: the roles exist and build and ship real products. Measurement tools (analytics, funnels, retention and A/B tests) are available to a team. In the open-source engine, the growth loop (evidence, cards, tickets and measured outcomes) shipped in release v0.103.148 on 19 November 2026; the release notes are on the changelog.
- Still coming: weekly usage reports with proposals delivered to you automatically. Until then, you ask the PM to review usage and propose what is next.
If you read a competitor’s claim of “autopilot”, ask the same question: what happens without you prompting it? We would rather you hear our limits from us.
A weekly sprint you can copy
| Day | What happens |
|---|---|
| Monday | Look at last week’s three numbers. Pick the biggest drop-off. Write the guess. |
| Tuesday | The team builds the change in a preview. You open it on your phone. |
| Wednesday | QA tries it like a customer. Fix what breaks. |
| Thursday | You approve it, and it goes live. |
| Friday to Sunday | Let real people use it. Do not touch anything. |
| Next Monday | Measure. Did the number move? Write down what you learned. |
Two rules keep it honest: do not judge a change before enough people have seen it, and keep a short log of guesses and results. After ten weeks, the log is your product’s most valuable document.
A worked example of one sprint
Imagine a booking page for a solo tutor (an invented example, to show the shape of the process). In week one, 40 people visit, 12 start booking, and 3 finish. The drop is between starting and finishing, so that is the sprint. The guess: asking for a phone number is scaring people off. The change: make the phone number optional. The team builds it in a preview, QA tries a booking on a phone, you approve, it goes live. Next Monday you compare the share of starters who finish. If it rose, you keep the change and look for the next drop. If it did not, you write down that the phone number was not the problem and try the next guess. Either way, you learned something in a week.
Common mistakes
- Measuring everything. Ten dashboards, no decision. Stick to three numbers.
- Changing too much at once. You cannot tell what worked.
- Judging too early. With a handful of visitors, noise looks like a trend. Collect enough before you decide.
- Chasing requests over data. One loud customer is not a pattern.
- Skipping the write-up. If you did not record what you learned, you will repeat it.
Where to go from here
If you are setting up your first sprints, start with how to launch a real product without code, and put the loop on a calendar with the 30-day plan for solo founders. To see who does what on the team, read what an AI startup team actually does.