See pricing
How to Build an AI Customer Research Agent in Claude Code
The Growth PodcastAI

How to Build an AI Customer Research Agent in Claude Code

Bolt's former CPO used to be happy with seven or eight customer conversations a month. Now an email agent he built in Claude Code in about 30 minutes talks to every customer who drops out of the funnel, and checks its own replies before they send.

November 13, 2026 · 59 min listen · 18 min read · Greg Greiner
0:00
–:––

Context

Greg Greiner was chief product officer at Bolt and head of product at Uber for Business. He is now CEO of Woven, a startup whose AI follows up with customers over email when they drop out of a funnel. Host Aakash Gupta asks him how PMs can use AI to talk to far more customers than a human ever could, and then has him build a working version live in Claude Code. The episode matters for any PM who knows they should talk to customers more but cannot find the hours. It covers the research workflow, the agent setup, how replies are checked before sending, and where automated research should not replace a real conversation.

The Big Idea

A PM can now run real back and forth conversations with thousands of customers a month by building an AI version of themselves that emails people at the moments they drop off.

Greg argues this used to need "an army of researchers" and support reps. The value is not just more data: each conversation can ask follow-up questions, try to resolve a concern, and show which problems messaging can fix and which need a product change.

Key Insights

From seven calls to thousands

At Bolt, Greg says he would have been happy with seven or eight customer conversations a month. Today, some Woven customers run tens of thousands, and some hundreds of thousands, of AI-led conversations every month. These start when someone drops off after giving an email address or phone number, and the AI reaches out on that channel. A large share replies and keeps the conversation going. Greg sees this as an evolution of in-app surveys, with one big difference: the AI can follow up on an answer instead of just recording it.

Messaging fix or product fix

Once you join conversation data with what people do afterwards, you can tell two kinds of concerns apart. Greg's example is a report for Ancestry.com on people who abandoned buying a DNA kit:

  • Shipping and delivery time: people who raised this converted well once the AI answered. That is a messaging problem, solved by talking about shipping and delivery more proactively.
  • Subscription confusion: people who raised this still did not convert at a high rate even after the AI addressed it. That points to a deeper product problem worth tackling more holistically.
  • How to see it: compare the conversion rate of each concern bucket against the baseline of everyone who had a conversation, not just how often the concern comes up.

Email beats in-app for leavers

Greg recommends in-app chat too, but email wins for drop-off research for three reasons:

  • People who abandoned are no longer in your app, so in-app tools cannot reach them.
  • An in-app research chat can distract someone who has not decided yet, and end up hurting conversion instead of helping.
  • Email is easy to roll out and scale, and customers already use it every day. Woven's research campaigns get 20% to 40% of people to reply, far more than would ever use a chat widget on a site.

Email also buys time. Nobody expects an instant email reply, so the agent can run quality checks for a few minutes before sending, which would feel broken in a live chat.

Reply rates depend on intent

What separates a 20% reply rate from a 40% one is mostly the type of campaign and where it sits in the funnel. Research-oriented campaigns later in the funnel, sent to people who already used the product, get the highest replies. A fitness company Woven worked with saw about 35% on trial abandonment. Earlier in the funnel, such as people who entered an email but never finished onboarding, replies are lower because intent is lower. Greg compares the rest to cold sales email: the more personal and relevant the outreach, the more likely someone takes time to answer.

Speed of reply matters

Woven tested adding delays before the agent responds, and reply rates dropped "precipitously." Greg's explanation is that people are likely no longer in their inbox or in the flow of doing email. Conversion drops alongside the reply rate. Fast, accurate replies at scale are where AI earns its place in this workflow.

The PM ratio is changing

At Uber, Greg felt the ratio of PMs to engineers was often too high, with one PM and three or four engineers, so PMs went looking for things to do. His most productive stretch was about one PM to 25 engineers, because there was no room for anything but building product. He now thinks an AI-enabled pod of three to five engineers plus a PM can produce what a 20 to 25 engineer team used to. That puts new demands on the PM:

  • Switch context very quickly.
  • Fill gaps beyond the usual job description, such as product marketing or design support.
  • Keep a flood of fast output cohesive, rather than shipping every customer request as quickly as possible.
  • Push back on meetings that are busy work, and use AI to cut the time spent preparing for the ones that remain.

Building the right thing is hard

At Woven, a team of four, Greg says building quickly with good quality has become the easy part. What is still hard is choosing what to build, so the team spends much more of its time in discussion. That only works if everyone at the table shares a baseline of product intuition and business understanding, even with different areas of expertise.

Mental Models & Frameworks

The research agent, step by step

Greg built his demo in about 30 minutes before recording, for a made-up AI fitness training company called Cadence. The setup:

  • Customer data source: a CSV of customers with their drop-off point and usage data, which lets the AI personalize outreach. In production this could be a customer data platform or a warehouse such as Snowflake or Segment.
  • Instructions file: a CLAUDE.md file with basic rules on how to approach the conversation and format the email, so it still reads like a personal email you would write yourself.
  • Company context: markdown files covering escalation processes, policies on what the agent can and cannot do, and the specific programs on offer. Many companies already have these. If not, Claude can generate them from your website.
  • Sending channel: the Gmail MCP, from a single demo address. For larger audiences, partner with marketing on an existing tool or set up Postmark or a similar service.
  • Outreach skill: a Claude Code skill that sends outreach. You can run it by hand, schedule it, or trigger it when new rows appear in the CSV or database.
  • What it does on each run: finds customers it has not contacted yet, writes a message from their metadata and the goal of the outreach, and sends it.
  • Reply handling: when a customer responds, the agent answers within a couple of minutes and keeps asking follow-ups. An existing classification step decides whether a reply is auto-sent or saved as a draft.
  • Notifications: Slack messages report each processed inbound reply.

The prompt that built it

Greg's starting prompt was mostly context. He told Claude Code he wanted to run an outreach program from his inbox using the Gmail MCP, pointed it at the customer data, described the conversations he wanted to start, and gave a trigger such as people 10 days into their trial. Then he let it run. The first pass needed fixes:

  • Reply threading was broken, so responses did not land back in the original email thread.
  • Some of the language was inappropriate.

Running it through tests fixed these, and he had something working in 30 to 40 minutes that was ready to test with real customers. His advice is to decide your trigger data source first, since everything else depends on it.

An eval loop before sending

Live on the show, Greg added a quality check with one instruction to Claude Code. The prompt, as he dictated it:

  • Add an eval step after classification for responses that are queued to be auto-sent.
  • Focus that eval step on groundedness in the documentation.
  • If the eval fails, rerun the generation process and save the new generation as a draft for review.

He then checked Claude Code's summary against what he meant. The evals run only in auto-send mode, since drafts get human review anyway. The groundedness eval checks generated replies for claims against the company documentation. On failure, it regenerates once with the specific issue corrected and saves the result as a Gmail draft. Without being asked, Claude Code also added a failure emoji to the Slack notification when an eval fails. Greg's next step would be to retry several times so more replies can still auto-send.

Feed failures into a full rerun

A failed eval should not just stop the process. Greg's point is to use the failure reason as input to another full run of the reply generation, not as an exact patch instruction. That way the agent still goes through its whole reasoning about how to respond, and you avoid overfitting to the one eval. Woven finds that for the vast majority of failures, the agent quickly reaches a reply that passes its evals and guardrails.

Four ongoing customer research activities

The host proposed this mix of research habits for a typical PM, and Greg agreed:

  • Direct conversations: Slack chats, texts or calls with customers, ideally watching them use the product, which the host suggested at least once or twice a week.
  • AI-led voice interviews: tools such as Listen Labs that run automated voice research conversations, best for later-stage companies with large user bases.
  • Email conversations at key journey points: the drop-off agent built in this episode.
  • Call and message analysis: a platform that analyzes sales calls and social messages.

Greg's addition is that the automated activities do not replace human conversations. They make them more powerful, because you arrive already knowing the basics and can spend the time digging into the interesting parts. Email conversations can also find the users with the richest insights to invite to a deeper voice interview.

Decision Principles

Principle: Ask one question per email

  • When: writing the first outreach email to a customer who dropped off.
  • Why: the common mistake is too much content and too many questions, trying to "recreate your survey" in one email. A short, plain-text, direct question relevant to where the user is in the journey opens the door. The customer can then lead you to what they care about, and the back and forth often covers all your questions at a higher rate.

Principle: Match the channel to engagement

  • When: choosing between in-app chat and email for research.
  • Why: for people actively and successfully using the core product, capture them in the moment with in-app experiences. For drop-off and abandonment moments, email reaches people who are no longer in the app.

Principle: Sandbox customer-facing AI first

  • When: an AI output goes straight to customers with no human review.
  • Why: internal tools, or tools whose output a person reviews first, can live with a much lower quality bar. Customer-facing AI needs guardrails, measurement of every update against them, and continued human QA. Even a model version bump needs a sandbox run on trusted sample data outside production, a check that guardrails still pass at a similar rate, and extra QA for unexpected regressions.

Principle: Automate only after early fit

  • When: deciding whether a PM should talk to customers by hand instead of through an AI agent.
  • Why: at low scale and in the early search for product market fit, direct conversations, including ones that are not text-based, build relationships and deeper understanding. The agent will never have the context the PM of an early product has. The AI approach fits later, when you already know you are creating value and want to maximize it.

Trade-offs & Nuance

Gmail is fine until scale

Greg ran his demo from a standard Gmail account, so one PM can do it alone without a marketing system. The catch is that Gmail limits how much outreach is reasonable. For a large audience, partner with the marketing team on an existing tool or set up a dedicated service such as Postmark. Greg notes it is already common for research tools like Medallia and Qualtrics to send from their own systems rather than through marketing.

Research versus conversion goals

Woven's paid product handles the same process at greater scale, with more of a conversion focus than a pure research campaign. The two goals can coexist. For trial users who did not convert, Greg suggests offering an incentive, such as 20% off the subscription for the next few months, just for sharing what could be better. It "massively" increases reply rate, and mainly among people already on the fence, since nobody replies for a discount on a product they had no interest in. You get insight and a likely conversion from the same email.

Fast teams can lose cohesion

AI lets a small pod ship like a large team, but the output can come so fast that the PM is just trying to keep up. Greg sees this as one of the biggest challenges for modern PMs. The risk is a product that ships everything customers ask for, quickly, without the features adding up to something greater than the sum of its parts.

Common Mistakes

Mistake: Trusting the LLM judge blindly

  • Why it happens: asking a judge model "has the AI done anything it is not supposed to?" feels like enough.
  • Why it hurts: both the AI and the LLM judge want to please. Greg recalls an early story, though he was not sure of the company, where a customer got a carmaker's chatbot to say they could have a free truck.
  • Better approach: have the judge break apart every specific action the AI took in a response and link each one back to the specific instruction that permitted it.

Mistake: Only studying the failures

Examining traces where the AI went wrong is important, but Greg says traces where it went right are an equally big opportunity. If almost nobody converts after raising a certain topic, and the AI's reply once helps someone convert, study the approach it took. That creative success can teach you as much as a major mistake.

Mistake: Treating process as the job

In large companies, PMs can feel trapped by meetings, approvals, stakeholder alignment and corporate planning boxes. Greg says the old trap was believing all of that was the actual job, instead of building product, talking to customers and shipping. With engineering output now arriving so fast, PMs have to push back on meetings that are busy work and use AI to speed up preparing for the rest.

Practical Application

Build a drop-off research agent

  • Do: pick one drop-off point, such as users 10 days into a trial who have not converted, and export those customers with their usage data to a CSV.
  • Then: write a CLAUDE.md with tone and formatting rules, add markdown files on your policies, escalation paths and offerings, and ask Claude Code to build an outreach program that sends through the Gmail MCP.
  • Check: test it on yourself first and look for broken reply threading and off-tone language, which were the problems in Greg's first pass.
  • Why it works: Greg had a version ready for real customers in 30 to 40 minutes.

Add a groundedness check before sending

Give Claude Code a prompt that adds an eval step after classification for any reply queued to auto-send, focused on groundedness in your documentation. Tell it that on failure it should rerun the full generation with the failure as input and save the new version as a draft for review. Then read its summary of the changes to confirm the eval runs where you intended.

Map outreach to the funnel

Set up research emails at the points where people drop off:

  • Signup abandonment: people who entered an email but never finished onboarding. This gets the lowest reply rate but the highest volume, and can inform the rest of the journey.
  • During the trial: one or more check-ins to see how people are progressing.
  • Trial ended, no conversion: pair the question with an incentive for feedback, such as 20% off for a few months.
  • Later in the lifecycle: an active user who tried an upsell feature but did not buy, a previously active user whose usage dropped in recent weeks, and a customer who cancelled.

Tag concerns with their conversion rate

For every concern customers raise in AI-led research conversations, record how often it appears and the conversion rate of people who raised it, compared with the baseline. Send concerns that convert well after the AI answers to whoever owns messaging. Treat concerns that still do not convert as product problems for the roadmap.

Reuse your engaged reply threads

Customers who replied to research emails become an engaged audience. When launching a feature, ask Claude to scan your inbox for everyone who mentioned something related to it. Then reach out with a follow-up question or an invite to a call about a feature built from their feedback.

Turn whiteboard debates into PRDs

Woven's team debates a problem together, with someone bringing an artifact that lays out the problem and opportunity. They record the discussion and whiteboard the solution, for example in FigJam. Then they feed the transcript and diagram into an AI skill that writes a detailed PRD, creates Linear tasks split into work streams, and flags areas that need more definition. In a large company, Greg suggests doing the same with the most relevant subset of people and having the skill output your required corporate PRD format.

Questions to Consider

  • Which of the concerns customers raise about our product would disappear with better messaging, and which would remain even after a clear answer, meaning they need a product fix?
  • Where in our funnel do the most people leave, and are we emailing any of them to ask why?
  • Do our outreach emails ask one simple question, or do they try to fit a whole survey into a single message?
  • If an AI agent replied to our customers with no human review, what source documentation would we check its claims against, and is that documentation up to date?
  • If our engineers could ship like a team five times larger, which of our current meetings and approval steps would get in the way?

Bottom Line

An AI agent built in Claude Code from a customer list, a few markdown files and the Gmail MCP can hold research conversations with every customer who drops off, which no PM could do by hand. It only earns trust with customers when every reply is checked against your real documentation before it sends, and it should sharpen your human conversations rather than replace them.

Case Studies Mentioned

Ancestry.com DNA kit abandonment

Woven's AI emailed people who abandoned buying a DNA kit with Ancestry.com. The resulting report showed the business impact and the most common concerns. One concern was an unexpected relative showing up in results and that being awkward, which the AI could follow up on. Layering in conversion data showed shipping and delivery questions converted after a reply, while subscription confusion did not. The lesson is that conversation data plus outcomes tells you whether to fix messaging or the product.

Fitness company trial abandonment

In a Woven experiment, a company in the fitness space emailed users whose trials ended without converting. About 35% replied. Greg uses it to show that later-funnel audiences with real product experience reply at the high end of the 20% to 40% range.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
Claude CodeAnthropic's agentic coding tool that works in your project filesGreg built the whole research agent and its eval step in it.
Gmail MCPLets an AI agent read and send email from a Gmail accountThe sending and reply channel in Greg's demo.
PostmarkEmail sending serviceGreg's suggestion for sending research email at larger scale, including through its MCP.
Snowflake and SegmentData warehouse and customer data platformAlternatives to a CSV as the customer data source that triggers outreach.
SlackTeam messagingThe demo posts a notification for each processed reply, with a failure emoji when an eval fails, and Woven sends research reports there.
Listen LabsRuns automated, AI-led voice research interviewsGreg's example of scaling deeper customer interviews at later-stage companies.
LinearIssue trackerWoven's AI skill turns discussion transcripts into Linear tasks.
WovenGreg's product that runs AI email conversations with customers at scaleThe managed version of the demo, with simulated examples, feedback-built guardrails and playbooks, and new engagement points set up in 5 to 10 minutes.

People to Follow

Greg Greiner

CEO of Woven, previously chief product officer at Bolt and head of product at Uber for Business. Worth following for a hands-on view of AI-driven customer research and how small AI-native product teams work. The host says he is on LinkedIn.

Notable Quotes

"I would have been happy to talk to seven or eight customers a month." (Greg Greiner)

"It's a channel that all of your customers already use on a daily basis." (Greg Greiner)

"So the eval should not just stop the whole process." (Greg Greiner)

"What's hard is still building the right thing." (Greg Greiner)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free