Over the last weeks, we've been talking a lot about the intelligence we're building into our solutions. But before it became part of our platform, AI was already being used by our own finance team. We tried it on the work we knew best, with the data we had, under the deadlines we couldn't move. Some experiments stuck. Others taught us exactly where language models stop being helpful. One of the most useful lessons began with a number that was completely wrong.
102 days
The first thing I ever asked an LLM to do at Lucanet was reconcile payroll data against prior months.
Not a strategy. A Monday afternoon, a spreadsheet, and a vague sense that I should understand what these things could do before I started having opinions about them.
It went well, so I tried something harder. Booking funnel data, looking for the kind of trend that might quietly change our forecast. And it told me that one stage of our funnel took 102 days.
I liked 102 days. It was specific. It contradicted received wisdom. The chatbot flattered me and made me feel like I was being rigorous. Perhaps there was some nuance in the data, hitherto overlooked.
So I asked it to show me the calculation, and it couldn’t.
Not wouldn’t. Couldn’t. It told me there were two possible approaches it might have taken, but neither reproduced 102. Each time the chatbot rehashed the analysis, I got a different number, as confidently presented as the last.
I sometimes refer back to my philosophy studies and question whether an argument is valid or sound. This was a third category. An argument that was neither, presented in the register of both.
The division of labor
That afternoon produced the rule I have applied ever since, and it is the same conclusion our CTO Kevin reached from the engineering side in his piece on the Intelligence Core: do not ask a language model to do the arithmetic. Ask it to work out which arithmetic needs to be done, then let something deterministic do it.
The reason is not only accuracy. It's that I can't sign something I can't walk someone else through. A controller does not re-derive every formula in a consolidation from first principles, but they can tell you where each number came from and which judgements sit underneath it. That is the standard. An output I can't interrogate is not a shortcut, it's a liability with good manners.
Once we accepted that, the useful applications for AI in our processes became much easier to spot. They were all narrower and duller than the ones I had been imagining.
What actually stuck
Over the next few weeks we tried a load of new use cases in the team. Three of them we are still using today:
- A pre-close revenue estimate, drafted from Salesforce data, gives us a day zero view of recurring revenue. Not the final number we report after a more detailed process, but one that tells us whether we are about to be surprised, days before we would otherwise find out.
- IFRS 15 calculations, where the work is interpretation against a rulebook rather than invention.
- And a personal favorite: reporting checks. Two kinds: one for writing style and one for consistency, both within a management deck and against the decks that came before it. It reads the pack and asks whether page 14 is still telling the same story as page 3, and whether either of them agrees with what we said last month.
That last one is the least impressive but maybe the most valuable, because it catches the type of error a human reviewer stops being able to see. By the fourth readthrough, you are no longer reading, just recognizing.
It matters that it wasn’t me who built any of these. Our VP Finance, our Head of FP&A, and our controllers did. We have developed the muscle across our team so individuals can address their most urgent challenges quickly. But this is not the same as having production-quality tools that can be used in anyone’s hands. A workflow built in an LLM is not a fully fledged product.
What home-brewed tools exposed
After some initial success stories, progress on bigger challenges has been slow.
In almost all the tasks we’ve tried to point AI at, we effectively asked the tools to look for problems in our data, and they obliged. Extensively.
Inconsistent definitions. Duplicate records. Fields that meant one thing in 2023 and something slightly different now, in ways nobody had thought worth documenting.
This is the part of the story that doesn’t appear in the demo. AI is very good at flagging where the risk is. It is not, on its own, a fix if you want a predictable, even deterministic, output. Getting outputs we can rely on has meant real data engineering and real integration work, and with home-brewed tools I don’t believe there is a version of this story where we could have skipped that.
Every one of our largest gains has started with data quality and system integration, not with a third-party model – we quickly found that building tools which were properly integrated into our data with the right level of validation was a job for a professional engineer.
Being the difficult customer
In parallel with my team’s efforts to build workflows and tools in LLMs, our product team has been trying to do the job properly.
At Lucanet, a critical step before releasing any agent is to test it on our own finance and tax teams. When our product colleagues talk about dogfooding, they are talking about my month-end.
That has been more useful and more challenging than either side expected.
Challenging because accountants are not trained product managers and they are certainly not QA engineers. Product leaders define good engagement in their own terms: structured feedback, reproduction steps, clear articulation of the expected behavior. All entirely reasonable, and all landing on a subject matter expert on day three of close who has a different definition of urgent.
Both sides had to move. My team had to learn that “it felt wrong” is a start and not a finding. The product team had to learn that in this case MVP couldn’t mean “minimum viable” but rather “totally reliable” product.
Some of the family of agents that has now shipped came out of exactly those moments. The Close Agent removes a large amount of manual work from the cycle. The Analyst Agent takes on some variance analysis, and the Emission Agent has compressed a sprawling manual task down to the small fraction of her day that Tina, who owns our sustainability reporting, actually wanted to spend classifying data.
I am often asked why we report ESG to the standard we do rather than the minimum we could. Beyond the cash saving on our sustainability-linked loan, the honest answer is that we are proud of it. And now with the time savings of a dedicated workflow agent, pride is less expensive than it was before.
Talent, and what leverage actually looks like
It's too early to look at any one of our agents or home-brewed tools and quantify the savings delivered, either in my team or the potential savings for our customers. But suffice to say our business keeps growing at over 30% a year and my finance team is reassuringly stable in size. Nothing about AI has displaced the thing that still determines whether any of this works. Human review and management remains the single biggest driver of success we have.
The more interesting change is in what I now look for in people.
First, experience completing a task has become less valuable than an understanding of how the task is designed. If a process can be described precisely, it can increasingly be delegated or, perhaps more importantly, explained to our product team. If it only exists in someone’s habits, it cannot.
Second, I need people who can see the whole orchestra. The old model, where a manager arrives each morning with a list, does not survive contact with tooling that can generate the list. Knowing what needs to be true by Thursday is now the job.
And self-review has never mattered more. When the first draft of anything is free, the scarce skill is being the person who can tell whether it is any good.
I don’t think our finance team has been particularly brilliant. We are all roughly as ignorant as each other at the start of something new. What is working, though, is trying more things than is comfortable, then reviewing the results honestly enough to notice what works. Small wins, repeated, compound into outcomes that look from the outside like a plan.
The 102 days was never true, but it might be the most useful number anyone has given me.
Discover Lucanet's family of workflow agents
AI is already reshaping finance and tax, the question is how to adopt it without losing control of your numbers. Join our webinar series to explore Lucanet Lume and the family of agents built to close, plan, and report on your terms. Each session focuses on one agent, its workflow, and the results teams are already seeing.
