How to Build an AI Customer Research Agent in Claude Code

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit www.news.aakashg.com/subscribe

Brought to you by

Jira Product Discovery - Plan with purpose and ship with confidence.

Today’s Episode

Engineering is getting faster with AI. Is product management?

I’d argue one way it can is by helping you talk to far more of your customers. You’ve probably used Claude to synthesize interview transcripts. But have you used it to run research with every customer who drops out of your funnel?

Greg Greiner has. He was chief product officer at Bolt and head of product at Uber for Business. Back then, 7 to 8 customer conversations made a good month. Now he’s CEO of Woven, where AI follows up with customers over email.

Some Woven customers run tens of thousands of those conversations a month, and some run hundreds of thousands.

You should be using AI to create a scaled representation of yourself to talk to your customers for you.

Then he showed you how to build your own version without buying his product, using a Claude Code project he’d put together in about 30 minutes!

10 Key takeaways

1. Go from seven customer conversations a month to thousands - At Bolt, Greg "would have been happy to talk to seven or eight customers a month." AI now holds those conversations with thousands.

2. Split every concern into a messaging fix or a product fix - Shipping questions convert once the AI answers them. Fix those with messaging. Subscription confusion doesn't convert even after the AI responds, so that one is a product problem.

3. Email the customers who already left - Someone who abandoned isn't in your app anymore, and an in-app research chat can hurt conversion for people who haven't decided yet. Email is "a channel that all of your customers already use daily," which is why Greg's research campaigns get 20% to 40% of people to reply, and a fitness company saw about 35% on trial abandonment.

4. Ask one simple question per email - The common mistake is trying to "recreate your survey maybe in one email." Ask one short, direct question and let the customer lead.

5. Put your emails where people drop off - Start at the widest part of the funnel, people who entered an email and never finished signing up. Few of them reply. It's still your biggest group. When a trial ends without converting, offer something like 20% off just for feedback. Greg says that "massively increases reply rate," and the people who reply are the ones already on the fence.

6. Build the agent in Claude Code in under an hour - Greg's demo runs on a CSV of customers and a few markdown files about the company, and it sends through the Gmail MCP. The first pass broke reply threading, but it worked within 30 to 40 minutes. Keep replies fast. When Woven tested adding delays, reply rates dropped "precipitously."

7. Add an eval loop with one prompt - Greg asked Claude Code to check auto-sent replies for "groundedness in the documentation" and to save a fresh draft when one fails. Feed the failure into a full rerun, or you'll overfit.

8. Test customer-facing AI in a sandbox first - An internal tool can live with a lower bar, because a person reviews it first. Even a model upgrade on customer-facing AI means running trusted sample data outside production and checking guardrails still pass at the same rate. LLM judges want to please too, so make yours tie every action to the instruction that allowed it.

9. One PM can now run a team that ships like 25 engineers - Greg's most productive stretch at Uber was one PM to about 25 engineers. Now three to five engineers ship that much, and the PM keeps it all "greater than the sum of its parts."

10. Spend your time deciding what to build - At Woven, a team of four, building fast with good quality is the easy part. "What's hard is still building the right thing." So the team debates the problem at a whiteboard while recording, then feeds the transcript and the diagram to an AI skill that writes the PRD and the Linear tasks. It only works if everyone at the table has a baseline of product and business sense.

Go Deeper

In-app surveys: the playbook from 4M PostHog responses

How to do AI-powered discovery, with Caitlin Sullivan

Teresa Torres' step-by-step guide to AI product discovery

AI customer intelligence: the complete guide for product teams

Claude Code for non-technical PMs, with Andre Albuquerque

Why 90% of LLM judges fail, and how PMs can fix them

The AI product operating model

Where to find Greg Greiner

Where to find Aakash

Newsletter

Listen to The Growth Podcast