To AI or Not to AI in Finance
Teddy Collins, EVP of Finance at SeatGeek, has a memo in mind with that title. On the Agent CFO Podcast he explained which finance work he keeps away from AI, and where AI lets his team do work it never had time for.
The video starts at 18:09, where Teddy describes the memo. Scroll back for the full episode.
- Teddy Collins keeps repeatable, rules-based work, such as pulling a report and transforming the columns, on RPA and middleware.
- Giving that work to AI costs tokens on a solved problem and hands browser control to a tool whose result still has to be checked.
- At SeatGeek, AI goes to work the team never had time for: compliance, SOX controls and a scheduled agent that looks into every corporate card decline.
- Ashok Manthena adds that once a workflow is deterministic and runs at volume, the choice of model and the token bill start to matter.
A memo called To AI or Not to AI
Ashok Manthena used the word deterministic while talking about how teams pick AI models for fixed workflows. Teddy Collins, EVP of Finance at SeatGeek, picked it up straight away.
“I've had this idea for a memo for a number of months. The title's gonna be To AI or Not to AI. Because not everything needs an AI solution.”
If a task is deterministic, works as if this then that, and repeats, Teddy sends it to tools that have been around for years: RPA and middleware.
What it costs to give solved work to AI
Teddy used a plain example: going to grab a certain report, pulling the columns you need and transforming the data. “Yes, AI can do that, but we don't really need to be burning tokens to do that or give up browser control.”
He is also wary of the result. “It's at best going to get it right, but who knows if it's really going to get it right. So let's keep that in the deterministic bucket.”
A rule that has run correctly for years gives the same output every time. An AI agent doing the same steps adds a token bill and a result someone has to review.
Work the team never had time for
At SeatGeek, AI goes to new work. “We are doing things now that we just didn't have the time to do before.”
He describes the aim as doing more for the company with the team he has: running more analysis, making compliance and SOX controls easier, and taking that work off people’s plates.
The card decline agent
His example is corporate cards. SeatGeek uses one of the leading corporate card management systems, and when someone’s card is declined, a scheduled task looks into what happened.
It diagnoses the decline and uses it as a point of education: it points the person to a wiki article or a best practice, or looks at the reason and turns it into a training moment. Each decline needs its own diagnosis and its own reply.
“We certainly could do that, but I would need to give that level of white glove service, I would need a team of twenty to just be doing that. And now we can do that.”
Agents talking to other agents is the next step he mentions, and he left it there: “that's for another time.”
Choosing a model once the workflow is fixed
Ashok sees the same split from the model side. When a person asks a simple question, any reasonable answer will do, because they can ask another tool if the first one misses.
Fixed workflows are different. “When you build a deterministic workflow, at that time you need to find out what model actually makes sense and the cost optimization as well. If you're gonna spend hundreds of thousands of tokens, you need to optimize it.” In his view, few teams have a good way to make that choice yet.
At SeatGeek, Teddy’s team keeps internal best practices on which model and thinking mode to use, and an internally built harness can look at a prompt and decide how to approach it. The way they track what all of this costs is in How SeatGeek Tracks AI Token Spend.
The person still owns the number
Early in the episode Teddy compared generative AI to any other tool in a finance career. “This is an amazing augmentation tool, but ultimately you're accountable for the results and you need to know the numbers.”
You still have to learn how to write, know what the right answer is and have the judgment, he said.
Agents for the work rules cannot do
ChatFin’s AI agents work on top of the ERP a finance team already runs, across the close, AP and reporting. Bring one of your own processes to a demo and we will map it end to end, including the steps that are better left to a rule.
Book a demoWatch the Agent CFO Podcast
New episodes come out on YouTube and Spotify. Guests include Teddy Collins of SeatGeek, Craig Sullivan of NetSuite and Samantha Singer-Guindi.
Talk to the ChatFin team
ChatFin runs reconciliation, AP matching and variance review on top of the ERP your team already uses, with the data behind every draft attached for review.