How Fanatics Betting and Gaming's finance team built ChatGPT adoption from the CFO's chair

What Ellis actually did, not what she believes about AI

Most executive AI interviews trade in generalities. This one, published by OpenAI on February 13, 2025, contains an unusually specific sequence of steps. Ellis, CFO of Fanatics Betting and Gaming, describes deliberately narrowing the surface area of experimentation before scaling it inside her own department.

Her stated reasoning is worth pinning down: rather than pursuing every possible use case, the team chose to concentrate. Two areas got the attention — finance and automating customer operations — and she volunteered her own function as one of the deep bets.

When we got started, we soon realized there's a decent chance of spreading yourself too thin across all the different use cases with AI. That's why we decided to go deep and land some big wins in key areas.Montana Labs

That is a resource-allocation decision, framed by a CFO in the language CFOs use. It is also the most transferable part of the account: the constraint was attention, and the response was concentration.

The adoption mechanics: task force, training, GPT-athon

Ellis lays out a four-part structure. First, an AI automation task force asked every finance team member to document processes they thought AI could improve. Second, that list became a project roadmap. Third, every team member completed basic ChatGPT training. Fourth, a full-day 'GPT-athon' paired teams with data scientists to build custom GPTs for specific projects.

The detail that matters here is the ordering. Training preceded building, and building was done collaboratively with data scientists rather than left to individuals to figure out. The bottom-up intake (ask people which of their own tasks hurt) was paired with top-down structure (a roadmap, mandatory training, a scheduled event).

Ellis also describes a maintenance loop: regular AI updates and celebrating new use cases during monthly all-hands meetings. Adoption, in her telling, is not a launch but a recurring agenda item. That reflects a realistic view — enthusiasm decays without a cadence to reinforce it.

The one number in the piece: 18 hours a month

The single quantified outcome is VendorID GPT, a custom tool that automates vendor identification and contract summarization. Ellis says it saves roughly 18 hours of monthly work, and frames the value in terms of the month-end close — the period when finance teams are most time-constrained.

It is worth reading that figure precisely for what it is: one custom GPT, one recurring task, a repeatable time saving inside a defined workflow. It is not a claim about revenue, headcount, or forecast accuracy. The rest of Ellis's benefits — faster data analysis, faster reading and communication of performance — are described qualitatively, and she is candid that 'it's still early.'

That honesty is useful. The measured win is a narrow, well-scoped automation; the broader claims about strategic thinking and scenario analysis are placed explicitly on a 'forward-looking roadmap,' not counted as delivered.

The recurring theme is execution-to-strategy, and it has a cost

Ellis returns repeatedly to one idea: use AI to reduce manual execution so the team can focus on strategy — the 'So what, what are we going to do about this?' question. It is the organizing principle behind why she raised her hand for finance in the first place.

The question that I posed was really, how can we make our day-to-day jobs less manual which ultimately will allow our teams to focus on strategy, and less on execution.Montana Labs

The implication for teams reading this: the payoff Ellis describes came from building bespoke GPTs against specific, documented finance processes, not from generic chatbot access. VendorID GPT exists because someone identified vendor review and contract summarization as a concrete, recurring cost and then built a tool for exactly that. The 'focus on the big picture' outcome was purchased with unglamorous work — process inventory, training, and pairing domain staff with data scientists to build the tools. That upfront structure, more than the model itself, is what this account is really about.