Northexis was fortunate enough to be a sponsor of Vena Solutions' first London Strategic CFO Experience event, in Waterloo's County Hall.
I was then lucky enough to moderate a roundtable discussion on tech and AI in finance, with the senior finance leaders in attendance sharing their successes, and not-so-successes, together. Predictably, "tech and AI" was really just AI, and it was a fascinating discussion.
I know I learned a lot!
AI hallucinations are of course a known problem, and one attendee pointed out that the technology, after all, is based on large language models - not large numeric models! So currently, often, an AI interface can get a number right once, but drilling down subsequently can lead to varying degrees of issues. One way that I've seen tools address this is by not asking you to trust a single number; they provide the answer, but also the data that they used to get there, so you can check it if you wish. Vena Copilot does this well.
That leads to a second issue to be aware of, which is that AI tools are typically statistically based. If you have twenty versions of a spreadsheet, its first impulse will be to average the values. But for finance, normally twenty versions are twenty revisions - and there's just one, the last, which should be taken into consideration. Something to bear in mind when training AI models.
There was of course the expected amount of concern around security, payroll data of course the obvious sensitive data that we deal with on a daily basis; but most finance professionals I know have this at the top of their list already. The learning for me is that there are ways to satisfy any level of security concern; one attendee had a dedicated AI server, so that the organisation was happy that only data they specifically put on that server reached the AI model.
We heard several fascinating success stories, one claims handling business replacing 30 to 40 claims handlers with AI, and another dramatically reducing the time it took to review terms and conditions. But a common thread was iteration; the first go was never particularly great. Between training the models, and the improvements to the underlying technology, the attendees saw continual improvement, but had to keep at it.
That's not how we're used to evaluating technology - normally, we see if it does what we want, and if not we move on. AI needs to be approached differently.
The biggest surprise for me though was what wasn't a focus. I was expecting a lot of use cases for the day to day gruntwork of finance - journal entries, reconciliations etc. And there was one story about sales commissions, but that was more about comparing data sources than anything else.
So I think there remains a lot of scope for core finance automation with AI. We're seeing a lot of enthusiasm for Vena Copilot's FP&A agents for just that reason.
But I expect we'll see a lot more from the market in the fairly short term too!
