How to talk to all your customers using AI: proactive email research with Claude Code

Preliminary note: the text received is truncated. It ends at the start of the section on how to add a live evaluation loop (eval loop). This summary covers only what appears in that portion of the transcript of a podcast episode. It is a conversation between Aakash Gupta, author of the Product Growth newsletter, and a guest named Greg. Greg was chief product officer at Bolt and talks about his experience with Woven, a customer conversation tool.

The starting thesis is that engineering is accelerating with AI, but product management can too if it helps teams talk to far more customers. Aakash notes that many PMs already use Claude to synthesize transcripts. The episode's question is whether they can use it to launch automated, proactive research with all their customers.

Greg says that when he was CPO at Bolt, he was satisfied talking to 5 to 10 customers a month. Now, he says, it is possible to hold real back-and-forth conversations with thousands or even hundreds of thousands of customers a month. Previously that would have required an army of researchers and support agents.

To illustrate, he shows a sample report from Ancestry.com. The recipients are people who abandoned the purchase of a DNA kit. The report is organized in several layers:

• Impact on the customer journey from the perspective of business metrics. Greg claims that this kind of personal conversation not only captures information but also makes it possible to convert a larger share of the audience. • Main concerns of those who drop out of the funnel: what they mention most often and what drives them to leave. • Cross-referencing with subsequent conversion data, to see the conversion rate for each category of concern against the baseline of everyone who converses.

That cross-referencing makes it possible to distinguish two types of problem. A question such as shipping and delivery time is solved with messaging: the AI answers the question and it is enough to communicate that information better. Others, such as confusion about the subscription, still fail to convert even when the AI answers. That points to a product problem that deserves a deeper approach.

Conversations start when someone abandons a flow, for example after leaving their email or phone number. The AI reaches out through that channel and, according to Greg, a high percentage reply and continue the conversation. Aakash sums it up as the evolution of in-app surveys plus AI. Greg agrees and adds the advantage of being able to ask follow-up questions, try to resolve the concern and see how the customer reacts. He gives the example of the fear that an unexpected relative might turn up and make things awkward.

On how to build it, Greg says that for many teams it won't be worth buying a dedicated product like Woven. With existing tools such as Claude Code, something similar can be put together. He built a demo in about 30 minutes before recording. Its components are:

• A customer data source. In the demo it is a CSV of people who abandoned a trial, with their drop-off point and usage data. A CDP or a data warehouse such as Snowflake or Segment would also work. • A CLAUDE.md file with instructions on how to approach the conversation and format the email, so it sounds personal, like a message written directly to the customer. • Markdown documents with company context. In the demo it is a fictional company, Cadence, offering AI-powered fitness training. They include escalation processes, policies on what the agent can and cannot do, and the available programs. If these don't exist, Greg suggests having Claude generate them from the website. • A skill that sends the outreach. It can be launched manually, scheduled or triggered by new rows in the CSV or database.

The agent goes through the list, identifies those who have not yet been contacted, personalizes each email according to their situation and the campaign goals, and sends it. In the example, it writes to someone who started using the app but did not convert after the trial. It asks about their experience, for example whether the program felt too intense. The agent replies within a couple of minutes and keeps asking questions.

Greg stresses response speed. Based on what he says he has seen with Woven, when delays are introduced the response rate plummets and conversion drops with it. His explanation is that the person is no longer in their inbox.

On why to use email rather than in-app chat, Greg offers several reasons. He recommends adding in-app chat as well, but points to these advantages of email:

• The audience that dropped out is no longer in the app. • Inside the app, you need to avoid distracting from the main goal and hurting conversion without yet knowing whether the person will leave. • Email is easy to deploy and scale, and it is a channel everyone uses daily. That makes it possible to be more proactive. • He says research campaigns get response rates of 20, 30 or 40%, far above what typically engages with a chat on a website.

On the difference between a 20% and a 40% response rate, Greg points to two factors. The first is the type of campaign: research-oriented ones further down the funnel, with users who have already shown intent, get more replies. He cites an experiment with a fitness company that got around 35% on trial abandonment. Higher up the funnel, for example when someone only left their email without completing onboarding, the response tends to be lower. The second factor is how personal and relevant the first message is. He compares it to B2B sales cold email.

His advice for the template is to use plain text and ask simple, direct questions tied to the point in the journey where the user is. The typical mistake is including too much content, trying to cram the entire survey into a single email. It is better to go slowly and let the customer signal what matters to them. Many of those questions end up coming out over the course of the conversation.

Greg proposes three moments in the journey, illustrated with a fitness app offering a free trial:

• Start: those who begin signing up and don't finish. This is the widest part of the funnel, with the lowest response rate but the highest volume. It also provides information for everything that follows. • During the trial, perhaps several times: helping set them up for success and checking how they are progressing. • End of trial without conversion: contacting those who don't move to paid. Here he recommends offering an incentive for sharing feedback, such as 20% off for a few months. He says this greatly increases responses, including among the undecided. That audience can resolve their doubts and receive a promotion that was probably going to be sent anyway.

For active users, Greg sees more sense in in-app chat, where the experience is captured in the moment. Channels such as email fit better at drop-off points:

• An active user who explored a new or upsell feature for the first time without converting. • Someone previously active who has reduced their usage in recent weeks. • Someone who canceled their subscription. According to Greg, some of the most interesting findings come from this audience, which decided the product wasn't for them.

As for who sends the emails, Greg ran his tests with a standard Gmail account. That limits volume, but lets you do it on your own, without depending on marketing. With large audiences, you would need to coordinate with the marketing team or set up Postmark or another tool. He notes that it is common for research emails to go out from a system separate from marketing's. He mentions Medallia and Qualtrics, whose emails are not sent by the marketing team. Aakash sums it up as setting up your own product research email system, and Greg confirms it as a common best practice.

Aakash recaps that email is a very powerful channel. It should be triggered at the right moment and with the right lifecycle trigger: trial, incomplete signup, inactivity and cancellation. On top of that come instant AI replies in your voice, to dig deeper and handle objections. Greg adds that for an individual PM it helps to better understand the audience for their product or feature and drive its adoption. It also builds an engaged audience to return to. For example, you can ask Claude to go through the inbox and find those who mentioned something related to a new feature. Then you write to them with a follow-up question or an invitation to a call.

On the prompts, Greg says it was straightforward. He gave Claude the context of setting up an outreach program from his inbox using the Gmail MCP. It would be similar with a Postmark MCP. He pointed it to the customer data source, which he says is the main thing to decide, and the instructions. One example was contacting people who were 10 days into their trial. The first version needed tweaks: reply threading wasn't linked properly and some of the language was inappropriate. Even so, within about 30 or 40 minutes he had something working to test with real customers.

The transcript cuts off just as Aakash introduces the section on adding a live evaluation loop. Its content and what follows are unknown. Note that the response and conversion figures come from the guest himself, with no supporting data in the excerpt received. Nor are consent and privacy discussed here when writing to customers who dropped out.