Building AI tools as a PM, Yatin Patel

In 2026, a product manager’s job changed. As I wrote at the time: “You’re not just defining and coordinating anymore, you’re building, testing, and iterating much closer to the problem.”

So I built. This page covers the tools I made for my own work, and the ones I handed to my team.

The problem

My work was spread across many tools. A data question meant filing a ticket and waiting three days. Each of the team’s dashboards ran as its own separate app. And AI tools made it easy to produce a lot of confident text that nobody had checked.

Context

Hertz’s product organization was adopting AI coding tools, through a company AI gateway with strict rules on data and on anything written to outside systems. Whatever I built had to live inside those rules.

What I built

An AI-native PM workspace. It takes a problem from discovery through data-backed sizing, a PRD, an automated quality review and approval-gated publishing. It runs on more than 30 written rules and 40 reusable agent skills, and every external write waits for my approval.
A data-investigation workspace for PMs, which I co-built. It gives a coding agent tracked investigations, a shared data catalog in three tiers, and a memory that learns from corrections. I ran a hands-on workshop that showed product managers how to answer their own data questions with it.
An AI agent for product managers. It connects PMs to company data and shared knowledge, and it shared the top award at Hertz’s first AI hackathon. I didn’t stop at the demo. I hardened it for a pilot myself: a security review, per-user cost caps, approval on every publish, and backend test coverage that nearly doubled.
The apps behind our contact-rate scorecard. They include a five-level drilldown of why customers call, covering about 500,000 calls a month, with independent cross-checks before any number goes live. Engineering built the transcription pipeline underneath.
A shared hub for my team’s data apps. PMs build, review and ship their own data apps there, with an AI assistant guiding each step so no one needs the terminal.
Field training made from phone recordings. For Hertz Premier Curbside, I turned screen recordings of the field process into an interactive training course for airport teams.

My PM workspace: I state the problem, the AI does the legwork, and nothing leaves without my approval. Illustrative.

Approach

Write judgment down as rules. The AI follows them every time, so quality doesn’t depend on my mood that day.
A person approves every external write. Nothing reaches a shared page, a ticket or a message without my sign-off.
Turn mistakes into rules. When an analysis goes wrong, the lesson becomes a rule. A correction that comes up three times becomes a proposed rule that I review. Some of these rules are on the Lab page.
Make it usable by people who aren’t engineers. A guided menu, one-step setup and plain-language help.

One shared home for the team’s data apps, with a guided path from idea to review to live. Illustrative apps.

Key decisions

Merge, don’t fork. When my manager built a team workbench with good ideas, I folded them into one home instead of keeping two.
Production, not just a prize. After the hackathon, I spent the time to make the agent safe for real users.
Loosen rules that become bottlenecks. The hub’s approval rules were too strict at first and made me the only reviewer, so we widened approval rights.

Results

Shared the top award at Hertz’s first AI hackathon (spring 2026).
Within eight days of launch, the team hub had five apps live and five people committing code.
The PM workspace runs my daily work.

I don’t have adoption numbers for the older tools, so I don’t claim any.

What I learned

Be strict about the shared parts, but never make one person the only approver.
Breadth before adoption is a trap. A tool with forty features and one user is a hobby.
The best rules come from real mistakes. The rules I trust most each have a story behind them.