How to run lifecycle experiments across the customer journey with an AI Manager

A lifecycle experiment tests one hypothesis about one moment in the customer journey: accounts that do X get message Y, and we expect Z. With Zentesimal, each hypothesis becomes a test with one trigger, one message and one metric, and an AI Manager runs it for every account in that moment. A team can experiment across the whole journey and check each result against its baseline.

An AI Manager is an AI agent that follows every account the way an assigned customer success manager (CSM) would.

This post explains the method we use, with examples from the first experiments we built for an HR software client, a time tracking and absence management tool for small businesses. The client stays anonymous.

Why do most lifecycle experiments never ship?

Because each one costs a project. A list, a flow, a message, a way to measure, and someone who remembers to check. The idea takes a minute. Building it takes more time than most CS teams can spare.

So customer success (CS) teams end up with two options. They test in the few moments a person can cover, usually the biggest accounts. Or they send the same automated email to everyone and call it a nurture flow (a fixed sequence of emails sent on a timer).

An experiment that never launches produces no result.

Testing is worth the effort, because intuition is a poor judge of which ideas matter. In a Harvard Business Review article on online experiments, Ron Kohavi and Stefan Thomke describe an idea Bing's program managers had ranked as low priority:

It was the best revenue-generating idea in Bing's history, but until the test its value was underappreciated.

The same holds for lifecycle work. The message that keeps a downgrading account, or gets a stalled user to finish onboarding, can differ from the team's favourite, as the Bing case shows.

Controlled tests settle arguments that opinion can't. In a survey of controlled experiments on the web, an academic paper by Kohavi and colleagues, a senior marketing executive at Amazon opposed shopping cart recommendations. A controlled experiment showed the feature won by such a wide margin that not having it live was costing Amazon money (Kohavi et al., 2009).

Where do good lifecycle hypotheses come from?

From two looks at your own data, one backward and one forward. We use the same method we laid out in how to reduce and prevent churn with Claude.

• Backward analysis. Take the accounts that already churned, converted or stalled, and see what happened in the weeks before. That shows which moments carry the volume.
• Forward analysis. Take every account showing a signal (a behaviour you can see in the data, such as a downgrade or a failed attempt) today and measure how many still end that way when nobody acts. That number is the baseline.

How do you write the hypothesis?

With a template that has one field for each part, filled in before launch. The example is the expansion test at our HR client, where some Basic plan accounts kept trying to open Projects, a module only available in the Plus plan:

Lifecycle hypothesis template

Written as one sentence: "if we offer help with the task instead of selling Plus, more of them will accept a demo, because they are stuck on a task and not shopping for a plan."

Without a baseline, every account that would have stayed looks like a win.

How does the experiment loop work?

Five steps, and the last one feeds the first. We run all of them in a conversation with Claude, Anthropic's AI assistant, which builds and reads everything in Zentesimal through the Claude and ChatGPT connector.

The lifecycle experiment loop, step by step

How do you test before launch?

On invented accounts whose emails all point to the team's own inbox. Step 4 catches problems before any customer sees a message. For each experiment we asked Claude for 20 invented accounts that mix three kinds of case:

Demo test matrix

Reading those messages showed more than reading any configuration.

Step 5 closes the loop: each result becomes the baseline for the next idea.

What does an AI Manager look like across the whole customer journey?

One experiment per moment, all running at once. The AI Manager acts when an account reaches a moment where help matters.

These are the first experiments we built for the HR client, each with its own hypothesis and metric:

Lifecycle experiments across the customer journey

The first four form a chain. An account that signs up gets onboarding; if it stalls, a follow-up; if it stays silent, a second touch; and once it gets going, the AI Manager watches for signs it needs more.

Every message is written for one account. This is an illustrative example of the expansion message, not a real conversation:

Hi there, I saw your question about logging hours per client, and that you've tried to open Projects several times. Projects does exactly that: it records the hours the team spends on each client. Projects is part of the Plus plan. If it helps, I can show it to you in 10 minutes with your own data, or we can look at how to track those hours with your current plan.

The message names the signal, states the plan limit and offers an option that costs the customer nothing.

Why does the AI Manager need to reply as well as write first?

Because an experiment that starts a conversation has to finish it. Lifecycle experiments are proactive: the AI Manager opens the conversation at the right moment. When the customer answers, a second Worker (an AI agent dedicated to one job) replies from the client's knowledge sources, in the customer's language.

That reactive half matters for the experiment too. The replies are where the reasons show up: why an account downgraded, what blocked an onboarding, what the customer was really trying to do. Those reasons feed the next hypothesis.

While testing, the reply Worker only answers messages from our own email addresses, so we can write to it as customers without it touching a real conversation.

What does the team decide, and what does the AI Manager run?

The team owns the judgment. The AI Manager owns the execution.

The team picks the moments, writes the hypotheses, sets the tone and reads the results. Its judgment lives in the experiment's playbook, the written instructions the AI Manager follows. It is the same guide you would hand a new CSM in their first week.

Playbook template

Tone deserves that care. The Nielsen Norman Group, a user experience research firm, tested how tone changes what people think of a brand. Trustworthiness explained about 52% of the variation in whether people would recommend it, against about 8% for friendliness (Nielsen Norman Group, 2016).

Three lessons from our test runs shaped every playbook:

• Say who is who. In one test, the AI Manager welcomed the customer's company as if it were its own. Every playbook now states who the AI Manager works for and who the customer is.
• Show a bad example next to a good one. A message that blames or pushes, placed beside one that helps, sets the tone better than any list of rules.
• Let a person decide on weak signals. In the expansion test, the AI Manager writes only when the signal is strong or medium. On a weak one, it alerts the team and waits. Signal strength comes from the backward and forward analysis: how much of the outcome a signal explains and how often it ends that way. It changes by company and sector, and how to rank signals by share and risk shows the calculation.

How do you know an experiment worked?

By comparing each check period against two lines set before launch: the baseline and the target.

Each experiment is read over a check period, weekly or monthly depending on the client's volume. The period has to hold enough accounts to tell a real difference from noise.

The minimum depends on the baseline rate and on the smallest lift worth detecting: the smaller the lift, the more accounts each group needs. The usual convention is 80% power at a 5% significance level, and sample size determination explains the calculation.

Lifecycle metric definitions

Where do the results come from?

From the AI Manager's own record. Zentesimal logs every action: which account, which message, and what the customer did next. From Claude, the team reads those results and groups them by experiment.

What are the three possible readings of a result?

Every result lands in one of three places relative to the baseline and the target.

• Above the target. The hypothesis holds, and the experiment becomes a permanent part of the journey.
• Above the baseline, below the target. Something works. The team changes one thing (timing, segment or message) and runs it again for one more check period.
• At the baseline. The experiment changes nothing. It stops, the reason gets written down, and the next moment takes its place.

A group of accounts that receives no message (a control group) makes the comparison stronger than a before and after, as the churn method explains.

So far these experiments have run on demo accounts, so this post covers the method and the setup.

Key takeaways

How Zentesimal runs lifecycle experiments

Zentesimal is an AI Manager for customer success and customer experience teams. Workers, AI agents specialised in each stage of the customer lifecycle, run these experiments inside the tools the team already uses, from the chat tool to Slack to the product's own data.

With the connector, the whole loop happens in a conversation with Claude: the analysis, the hypothesis, the setup, the test and the results. Your team decides what to test, and the AI Manager runs each experiment on every account that shows the signal. The conversion and retention use cases show where these experiments fit, and how to convert more free users to paid goes deeper on onboarding.

Book a demo