How I AI: Building a Real-World Business with Andrew Mason & Nabil Hyatt

A board-game club in Berkeley became an excuse to test how far AI can stretch beyond software work. Andrew Mason, founder of Groupon and CEO of Descript at the time of recording, and Nabil Hyatt, a partner at Spark Capital, used AI to help open Tabletop Library, a physical membership club built around modern board games, events, and player matchmaking.

The project involved permits, lease negotiations, financial models, a physical storefront, hundreds of games, and real customers. Andrew and Nabil treated AI less like an oracle and more like an always-available operating partner that could generate drafts, organize context, and handle tedious coordination while they made the actual business decisions.

In this episode of How I AI, they walk through three systems that made the project feasible as a side business: a Claude Project that became the operational memory of the company, a custom classification system for organizing the game library, and a text-message concierge that helps members coordinate games.

Andrew's most useful point is that the shift is not a special prompting technique. It is remembering to involve AI when an unfamiliar operational problem appears. Every draft, spreadsheet, ordinance summary, and customer hypothesis became something concrete to review, while landlords, lawyers, permitting guidance, and human judgment stayed in the loop.

Using Claude Projects as the operating system for a new business

Andrew and Nabil used Claude as a shared workspace for research questions, projections, planning documents, operating assumptions, and ongoing decisions. Both had experience building technology companies, but neither had run a retail board-game club before, so the project quickly became a test of whether AI could help close operational knowledge gaps fast enough to make the idea practical.

Building a persistent business context

Instead of treating each chat as disposable, they kept everything inside one Claude Project and continuously fed new artifacts back into the system. Early conversations produced mission statements, business-plan drafts, and market analysis. Those documents later became reusable context for questions about pricing, permitting, events, and operations. The result was less repetitive prompting and a growing institutional memory for the business. They still had to manage what information belonged in the workspace, update stale assumptions, and avoid dumping sensitive customer data into general-purpose AI tools.

Turning vague ideas into an actual operating plan

Their starting point was intentionally simple: describe the kind of place they wanted to exist and the experience they hoped members would have. From there, Claude expanded the idea into operational questions they had not fully considered yet, including staffing assumptions, pricing structures, event programming, lease considerations, and utilization planning.

I wanna open up a place to play board games with my friends in Berkeley. These are my goals.

The Claude workspace gradually accumulated a surprisingly broad set of operating artifacts:

• Financial projections: Draft budgets, pricing scenarios, break-even assumptions, and revenue forecasts tied to membership usage and event capacity. The models helped them reason about viability, but the outputs still depended entirely on the assumptions they supplied.
• Space Layout: Ideas for how to physically arrange the shop to fit the most tables while still feeling welcoming.
• Real estate and permitting: They pasted Berkeley ordinance pages, zoning details, and lease information directly into Claude to help interpret requirements and draft materials such as a landlord letter of intent. The AI-generated summaries gave them confidence navigating unfamiliar processes, but they still verified requirements with professionals and official agencies.
• Landlord presentation: A draft deck explaining the concept, operations, and business rationale for the space. The interesting part here was not the slides themselves. It was the speed. Instead of spending days figuring out what a Berkeley landlord might expect, they could iterate rapidly on a credible first pass.

Using AI to map customer behavior

One of the more concrete planning exercises was a customer-persona framework that mapped players across two dimensions: whether they preferred mastering a single game or constantly trying new ones, and whether they wanted quieter experiences or highly social play. That framework became a way to reason about programming decisions, membership structure, event design, and even how the physical space should feel for different groups.

Claude suggested rough proportions and events for each customer type, but Andrew and Nabil treated them as hypotheses rather than research. The AI business-plan workflow turns those early assumptions into an operating draft that still requires real market and financial validation.

Building a Dewey Decimal system for board games

Modern board-game stores often overwhelm new players with walls of unfamiliar titles. Andrew and Nabil wanted browsing to feel more like exploring a library, where related experiences naturally sit near one another. That led them to create a custom categorization system for their collection.

Designing the Tabletop Library Classification System

Their Tabletop Library Classification System assigns broad genre ranges alongside a decimal-style complexity score. Cooperative games occupy one section, adventure-focused cooperative games narrow the grouping further, and an added complexity marker indicates whether a game is approachable for beginners or much heavier strategically. The goal was not perfect taxonomy. It was helping customers discover adjacent games without needing deep knowledge of the hobby.

• 400s: Cooperative Games
• 420s: Adventure Co-op Games

The system also included a decimal for how complex or "heavy" a game was, from .1 (lightest) to .5 (heaviest). So, a really complex adventure co-op game might get the code 420.5.

Using Airtable and AI to categorize hundreds of games

They loaded the collection into Airtable and used AI-generated fields plus Claude prompts to propose classification codes based on each game's mechanics, themes, and complexity. Human review still mattered because the categories were subjective and occasionally inconsistent. But AI removed the impossible part of the task: manually researching and organizing hundreds of titles from scratch.

The resulting shelf layout lets customers browse by style and difficulty instead of alphabetically. The AI product-classification workflow shows how Airtable and Claude propose codes while staff review subjective categories and maintain the taxonomy.

Generating retail collections automatically

Airtable’s AI fields also help generate themed retail groupings from prompts such as "Games about cats" or "Silent Strategists". Instead of manually curating every display, the system proposes candidate games automatically, giving staff a faster starting point for merchandising decisions. Humans still review the final collections before anything appears in the store.

Building a low-code AI concierge for player matchmaking

The most ambitious workflow in the episode is a text-message concierge that helps members coordinate games with one another. A customer can text the club describing the kind of game they want to play and when they are available, and the system attempts to assemble a compatible group automatically. Because the workflow can contact people and update reservations, Andrew and Nabil designed explicit confirmation steps into the system rather than letting the model act autonomously.

Choosing a lightweight stack

Their setup intentionally used familiar tools instead of a fully custom application. Each layer handled a narrow operational role:

1. Airtable: Stores member records, games, reservations, availability, and preference data. Andrew pointed out that free-text fields became much more useful once AI could interpret them directly instead of requiring heavily normalized database structures.
2. n8n: Runs the orchestration logic, connects services together, and manages the workflow routing between messaging, AI agents, and Airtable actions.
3. Twilio: Handles inbound and outbound text messaging through the club's phone number, including member notifications and coordination messages.

How the concierge actually works

The demonstrated workflow was intentionally straightforward but surprisingly capable:

1. A member texts the Tabletop Library number describing the kind of game they want to play and roughly when they are available.

Hey, can you set up a game for me anytime this weekend? Maybe a deck building game.

1. Twilio receives the message and passes it into an n8n workflow for processing.
2. The workflow checks whether the sender is a member and retrieves only the relevant Airtable records needed to fulfill the request.
3. An AI agent interprets the request and uses a restricted set of Airtable actions to look up games, members, and reservation details.
4. The agent proposes a potential group and coordinates outreach, but invitations and reservations still require explicit confirmation rules rather than trusting the model to infer intent correctly.

Using plain-English instructions instead of heavy code

Much of the concierge behavior lives in plain-English instructions, but conventional safeguards still govern tools, permissions, validation, and outreach. The low-code concierge workflow connects Airtable, n8n, and Twilio while keeping invitations and reservations behind explicit confirmation.

What AI actually changed for this business

The important pattern across all three workflows is that AI reduced coordination overhead and unlocked work that otherwise would have been skipped entirely. The Claude workspace accelerated planning. The classification system made a giant game library navigable. The concierge enabled lightweight social coordination that would be too tedious to run manually. None of those systems removed the need for review. Financial assumptions still needed scrutiny, permitting still required verification, and customer-facing automations still needed guardrails.

What feels worth copying from this episode is the operational mindset more than any specific tool. Start with one annoying task that repeatedly stalls a project. Centralize the context, let AI produce drafts and structure, then define exactly where human approval belongs. This approach works especially well for messy operational work where the bottleneck is coordination, categorization, or first-pass synthesis rather than precision-critical execution. The systems here became useful because Andrew and Nabil continuously reviewed outputs, refined assumptions, and treated AI as a collaborator instead of an authority.