How a Ford exec built a ChatGPT-based personal AI assistant stack and why it matters for builders

Whitney Stefko Dover, a Ford director and senior counsel, built a Claude-based family assistant named Claudette that turned emails and calendars into a 4 a.m. "Daily Dover" briefing. Three months later, Claudette is on a performance improvement plan. Its replacement: Gigi, a ChatGPT-based personal AI assistant built in ChatGPT Work. The switch wasn't about model superiority. It was about owning the context layer.

Claudette is on a PIP: a real-world agent swap

Stefko Dover initially "vibe coded" Claudette using Anthropic's Claude to manage school schedules, childcare, travel, birthdays, and household chores. It even drafted a morning text for her husband Business Insider. But after her husband and an a16z Substack post pushed her to try ChatGPT, she built Gigi. The transition was surprisingly painless because she had already created a guide explaining how Claudette worked and connected to her family's apps. She uploaded that material into ChatGPT, preserving much of the old assistant's context.

Gigi offered a better interface, stronger first drafts of emails and PowerPoint presentations, and more proactive suggestions. It also made fewer contextual errors. For example, it reads the family nanny's grocery list every Friday and fills online shopping carts. One recommendation included 12 boxes of bow-tie pasta, but Stefko Dover still reviews and pays for orders herself. She mentioned she might impose a $500 credit card limit to prevent Gigi from "purchasing diamonds on Rodeo Drive" Business Insider.

Why context ownership beats model loyalty

The key insight for builders: Stefko Dover owns her context, not her model. She created a portable guide for Claudette and reused it with ChatGPT. "I think it is our job as the users to sort of own the context layer," she told Business Insider. This means users can switch AI providers without rebuilding everything from scratch. The model becomes a commodity; the context and workflow design are the moat.

She didn't stop at one replacement. Since the interview, she added a third assistant: Ingrid, built with the invite-only Instinct AI, which can execute tasks through text messages and phone calls. Ingrid is layered on top of Gigi. For example, instead of texting her husband to-dos, Ingrid talks to his agent, which then performs tasks on his behalf Business Insider. This is a multi-agent workflow where each agent handles a different channel or capability.

What builders can learn from the Dover family AI stack

First, budget controls and human supervision remain essential. Stefko Dover still reviews grocery orders and would cap spending. Second, proactive drafting and continuity across projects are differentiators. Gigi remembers more about her identity and makes suggestions without being asked. Third, the ability to iterate quickly and swap agents is a feature, not a bug. Stefko Dover expects no final destination: "The power of being a user is this constant iteration."

For product teams building personal AI assistants, this case suggests that users will hop between platforms if the experience improves. The winning strategy is to make context portable and let users layer agents rather than lock them into a single model.

The limits of this case study

This account comes from a single Business Insider report and Stefko Dover's own statements. There is no independent verification of the claims about Gigi's performance or the capabilities of Instinct AI. The comparison between Claudette and Gigi is anecdotal, not a controlled benchmark. Still, it offers a rare, detailed look at how a sophisticated user iterates on personal AI tools in the wild.