Practical AI in Finance: Data, Automation and Keeping Humans in the Loop

About this episode

Adam Swoverland runs finance at US Auto Force, where he oversees FP&A, accounting, collections, AP and customer onboarding.

He came up through treasury, credit and commercial banking instead of public accounting, then spent about four years as VP of Finance at Breakthrough, a logistics technology company in Green Bay. Business partnering is the part of the job he enjoys most.

Adam wants AI to make his team faster and more productive, and he says no jobs have been cut because a growing company keeps adding work. He argues finance cannot build the data foundation alone and needs IT as a partner.

His company runs a corporate AI initiative: IT supplies the technology, his team looks for the bigger use cases, and staff get training on using AI for daily work in Excel and Outlook.

He still warns that AI may not be the most efficient answer to a given problem, and that pricing and inventory management need specialized tools a general assistant can't match.

About the guest

Key moments

1. 03:37

   Collections work is mostly problem solving across systems

   Adam dislikes the word collections because most of the work is answering customer questions: where an invoice is, how a payment was applied, why something never arrived. The answers sit in disparate systems, so staff go hunting for them.

   He sees AI listening to the conversation and bringing the right data to the person handling it, with that person still in the loop.

2. 05:12

   IT brings the technology and finance brings the problems

   Adam says finance cannot set up its data for AI on its own, and that what companies built in the past is often a poor foundation for it.

   He describes a loop: IT knows what the tools can do, where their limits are and how fast they are changing, while finance knows the pain points inside finance and across the company. His team spends a lot of time with IT on governance and data infrastructure.

3. 10:32

   Building automations doubles as process documentation

   Writing SOPs rarely feels worth the time, so Adam's team protects tribal knowledge another way. They use Copilot and Python for what he calls desktop automation, removing tasks that eat time without adding much value.

   Building each automation forces someone to write down how the process works, which covers the risk when a senior person leaves.

4. 15:58

   An agent had to learn that site means warehouse

   When Adam asked an AI agent on the company's sales data which warehouse to look at, it said it had no warehouse, because the system codes locations as site names. Someone on his team told it the two were the same, but he says he can't build an alias table for every term.

   He also worries about hallucinations, since the agent states wrong figures with high confidence and only apologizes when challenged.

5. 24:48

   ROI is clear for processes and fuzzy for productivity

   For AI deployed on specific processes, Adam compares the old cost, the time saved and the cost to deploy, and a couple of those cases are clear cut. Thirty minutes saved here and there, or a meeting recorded for the team, is harder to count.

   He leaves room for focused experiments with no promised ROI. A blank check for everyone, he says, would be irresponsible.

Chapters

Selecting a time plays the video from that point.

1. 00:00Introduction
2. 00:04Adam's role at US Auto Force
3. 00:54The most valuable part of finance leadership
4. 02:16Customer experience in finance operations
5. 03:37Collections as problem solving
6. 04:42Data sources as the biggest AI blocker
7. 05:13Finance and IT working together on AI
8. 08:14Knowledge graphs vs AI assistants
9. 10:33Capturing what long-time staff know through automation
10. 11:22From Python to building workflows in chat
11. 15:59Giving AI context
12. 16:54Hallucinations and keeping a person in the loop
13. 18:20AI for productivity instead of job cuts
14. 19:39How AI adoption happens inside a company
15. 24:44Measuring ROI on AI projects
16. 26:08Small experiments before scaling
17. 27:21Data infrastructure first
18. 28:02Closing thoughts

Heard in this episode