How I AI: Webflow CPO Rachel Wolan's AI Chief of Staff & Org Adoption Playbook

In this episode of How I AI, Webflow CPO Rachel Wolan demonstrates a local AI chief-of-staff app for calendar and inbox triage, a research workflow for meeting preparation, and the Builder Day format her organization uses to encourage hands-on AI adoption.

Rachel started coding at 16, and AI pulled her back into building. She walked me through the custom "AI Chief of Staff" app she built for herself: a local tool that helps her triage her calendar, manage her inbox, prep for networking dinners, and get surprisingly blunt feedback about where she's wasting time. It's messy, personal, and incredibly practical.

Rachel also described two Builder Days at Webflow. The first focused on design, while the second included product, design, data science, research, analytics, and engineering. The event format paired tool access with assignments, support, prototypes, and prizes.

build a personal AI chief of staff for daily triage

How to Build a Personal AI Chief of Staff for Daily Triage: Build a personal executive assistant that turns calendar and inbox signals into a focused daily brief, while keeping permissions narrow and every consequential action visible to you.

Executives are constantly context-switching, usually while running late and underprepared. I often call myself a "just-in-time executive,", prepping for meetings minutes before they start. Rachel built her AI Chief of Staff to handle exactly that chaos: calendar review, email triage, prep work, and delegation.

connect approved APIs with limited permissions

The first step was giving the Chief of Staff access to the systems she already lives in. Rachel connected the app to her work tools using official APIs and authentication tokens.

• Google Calendar & Gmail: She uses the Google Calendar API and Gmail API to give her agent access.
• Environment variables: Rachel keeps API credentials in an `.env` file and does not display it because exposing the file would require rotating the tokens. An `.env` file is a configuration mechanism, not a security boundary, so it must remain out of source control and be protected like any other secret store.
• Permission limits: Google Calendar access is read-only. Gmail access can read, archive, label, and create drafts, but it cannot send messages. Rachel reviews actions that could affect other people.

review the calendar and identify delegation options

Every morning, Rachel runs a simple prompt to review her upcoming schedule and look for opportunities to delegate, decline, or simplify. She does it directly in her terminal using Claude Code.

"Tell me about my day tomorrow. What can I delegate?"

The agent reviews her calendar and returns a prioritized set of recommendations:

• Make Meetings Async: It identifies meetings that could be handled via a document or Slack update.
• Delegate Tasks: It suggests specific meetings that a director or another team member could cover, even drafting the message to send them. For example: "Can [Director's Name] cover this and send me a summary?"
• Decline Optional Invites: It flags meetings from unknown organizers or with unclear agendas that she can likely skip.

What I like about this workflow is that it forces a real review of how her time is spent. The value isn't just automation. It's getting an outside perspective on whether her calendar actually matches her priorities.

ask for the "brutal truth"

One of the funniest parts of Rachel's setup is the "Brutal Truth" module. She explicitly trained the agent to call her out when she's operating below the level of her role, and the feedback is sharp in a way most humans would probably avoid.

After reviewing her calendar, the agent came back with this:

The Brutal Truth: "You're operating as a senior PM, not a CPO. You're reviewing PRDs, approving scripts, and recording marketing videos... The only thing that will matter in six months is [a specific strategic conversation]."

The module does more than summarize the schedule. It compares Rachel's activity with the expectations of her role and offers a deliberately blunt interpretation that she can accept or reject.

triage email without sending

Rachel applies the same approach to email. Her agent helps manage a 500+ email backlog by:

• Archiving ruthlessly: It gets rid of newsletters and non-essential communications.
• Highlighting what's important: It identifies emails where someone is waiting on her for a document or a decision.
• Drafting responses: It prepares draft replies for recurring requests or partnership inquiries, saving her from typing the same thing over and over.

prepare for meetings and networking events on demand

How to Use AI for On-Demand Networking Event Preparation: Turn a guest-list screenshot into an evidence-linked event brief by resolving each attendee's identity, researching current work, and combining verified facts with your own goals and context.

Rachel also uses her AI Chief of Staff for just-in-time prep work. One of my favorite examples was how she prepared for a networking dinner using a mix of vision models, web research, and personal context pulled from markdown files.

extract a guest list from a screenshot

The workflow started with something delightfully low-tech: a screenshot of the dinner guest list.

Rachel dropped a screenshot into the app and used vision and OCR to extract attendee names. She had not fully reviewed the generated research during the recording, so the output remained preparation material rather than verified biography.

run a multi-step research agent

Once the names were extracted, a research agent took over. Instead of making a single LLM call, Rachel chained together multiple research steps to build a fuller picture of each attendee:

1. Initial Web Search: The agent performs a general web search for each person to get a broad overview.
2. LinkedIn Search: It then specifically searches LinkedIn to find their professional profile, current role, and career history.
3. Deeper Web Search: Finally, it conducts another, more targeted web search for recent news, articles, or talks they may have given.

The agent runs a general web search, a LinkedIn search, and a more targeted follow-up search. Rachel still needs to confirm that each result belongs to the correct person.

add reusable context from Markdown files

This is where the workflow becomes truly personalized. To ensure the output is relevant to her, Rachel feeds the agent several markdown files containing key information:

• `about_me.md`: Contains her professional bio, communication style (from a workshop she attended), and key career highlights.
• `webflow_products.md`: An auto-generated file, updated monthly from release notes, that details all of Webflow's latest products and features.

By feeding those markdown files into the workflow, the agent could generate talking points tailored both to Rachel and to Webflow. I especially liked this approach because it's simple. No complicated database setup, just well-maintained markdown files that become reusable context across different agents.

assemble the prep document

The final output was a structured prep document inside her web app that included:

• Priority Connections: A list of the most relevant people she should try to speak with.
• Personalized Conversation Starters: Tailored opening lines for key individuals based on their background and her own.
• Hot Discussion Topics: Relevant industry trends, like Answer Engine Optimization (AEO), to bring up.
• Venue Information: Details about the restaurant for easy reference.

For anyone who finds networking dinners mildly painful, this is a pretty compelling use case. The workflow turns her from a "just-in-time executive" into someone who walks into the room already knowing who matters and what to talk about.

run a hands-on Builder Day

How to Run a Company-Wide AI 'Builder Day' to Drive Tool Adoption: Run a focused builder day that gives every function approved tools, a relevant starting track, live support, and time to produce a real prototype, then measure whether usage continues afterward.

Rachel's organization used Builder Day to give people dedicated time, tools, assignments, and support for building prototypes. She presented it as an adoption experiment that the team was still refining, not a finished universal playbook.

provide tools, tracks, and support

The goal of Builder Day was simple: help people get over the initial friction and see what these tools could actually do in their day-to-day work. Rachel's team created a structured environment with clear assignments, dedicated support, and access to the right tools.

• Tool Access: The company provided access to a curated set of AI-powered tools, including Cursor for coding, Figma for design, Make for automation, and of course, Webflow.
• Dedicated Tracks: They created different assignments and tracks for various functions (Product, Design, Data Science, User Research) so everyone could work on something relevant to their role.
• Support Channels: A dedicated Slack channel was staffed with engineers to help troubleshoot technical issues, ensuring no one got stuck.

make the event participatory

The event worked in part because it felt more like a creative sprint than mandatory training. The structure gave people enough support to get unstuck while still leaving room to experiment.

• Warm-up Assignment: Participants started with a simple exercise to get comfortable with the tools before diving into their main project.
• Judging Panel: To add a little friendly competition, a judging panel including Rachel and the CEO reviewed the projects.
• Prizes & Recognition: Winners were awarded prizes and recognized in different categories, celebrating the effort and creativity across the company.

inspect usage and participant feedback

Rachel's team used a Hex dashboard to compare tool usage before and after the events. She emphasized that the important question was whether usage remained sustained after the event spike.

After the design-focused Builder Day, roughly half of the design team was using Cursor, up from very little usage beforehand, and the dashboard showed continued use in the following weeks. Rachel did not claim that this percentage represented the entire organization.

A post-event survey described the day as fun, empowering, motivating, and eye-opening. The second event produced more than 80 prototypes across several functions, while long-term adoption was still being observed.

The operating pattern

Rachel first built software for her own work, learned where permissions and review mattered, and then helped create a supported environment where other functions could experiment.

The personal app and Builder Day solve different problems. One creates leverage for an executive with a specific workflow. The other reduces the initial cost of learning by giving a group time, tools, support, and a concrete assignment.

A useful first step is one local, bounded tool that you can evaluate yourself. For organization-wide adoption, measure what happens after the event rather than treating participation alone as success.