Every conversation about AI in finance hits the same question: “does this actually work, or is it just a demo?” The most useful answer isn’t a thesis. It’s cases, with names, titles and numbers.
Datarails, a US financial planning software company, gathered 14 of them in an ebook. They range from Microsoft and Walmart to fintechs and an independent CFO serving several startups. Below is our read: what each group of cases shows, and the pattern that runs through all of them.
Where this comes from
This is vendor material, and the figures are the ones the companies reported. The ebook says nothing about cost, implementation timelines or projects that failed. Even so, it is one of the most concrete collections of AI-in-finance cases out there.
FP&A: less gathering, more deciding
The first five cases are FP&A. The standout is Microsoft: the quarterly forecast took 1,000 people three weeks, with spreadsheets rolling up to the CFO. With machine learning, introduced back in 2015, it now takes 30 minutes, and the variance rate fell from 3% to 1.5%.
At Shopify, an agent wired to three sources (BigQuery data, product documentation and the analysts’ written workflow) does in minutes the weekly analysis that used to take days. The most interesting effect is on the meeting: discovery is already done, so the conversation opens with “what should we do?”
Lyft uses AI for a first pass at the month’s variance explanations and to digest reports from the 47 banks covering the stock. Wasabi turned six months of old executive summaries into the template for a generator: every Friday a button drafts the text with the right numbers, and a person just edits.
“It’s not yet something you can just forward to your board of directors, but it’s very helpful because it helps you get to this first pass.” Aurélien Nolf, Lyft
The close: repetitive volume is the bottleneck
The ebook notes that the average company takes 6 to 10 business days to close, and the best take 3 to 5. The bottleneck isn’t accounting rules; it is the sheer volume of manual work that needs no judgment.
- Fanatics Betting & Gaming: a custom workflow identifies vendors and prepares journal entries. AP work at month-end went from 20 hours to 2.
- Workiva: each journal entry went from 10-15 minutes to 1-2 minutes. With hundreds a month, that is days back.
- Walmart: automating invoice matching against point-of-sale data removed the need for several hundred people. Non-routine exceptions still go to someone.
Financial operations: finance stops being a help desk
The third group deals with the daily flood of questions and messy data. Citi improved treasury cash forecasting accuracy by 50%. Mercury Financial cut analysis from days to hours with the same team.
The most instructive case is PayJoy: 80% of transactions had no vendor name in the right field, across 10 countries and several languages. A multi-step solution cleans and translates, normalizes, pulls the vendor name out of free text and clusters whatever won’t match. The monthly report that took an analyst 40 hours now takes a few, and the freed-up time went into fixing the data at the source.
At Billerud, a Swedish paper maker, replacing an old OCR tool with AI for PDF invoices cut manual validation from 15% to 9% in three months. Today over 90% of invoices are validated automatically, and handling cost fell 66%.
Technical accounting: AI searches, the expert decides
The last two cases matter most to anyone who works with standards, opinions and memos. An independent CFO drafted a regulatory filing for the SEC in under two minutes, then used her own expertise to validate it.
At Tipalti, a language model trained on IFRS and GAAP lays out what each technical memo needs to cover. The accountant decides the treatment and writes it. Then the finished memo goes back to the AI, which summarizes it and flags what’s missing, as a quality check. That saves about two days per memo.
Judgment stays with the person who knows the rules. AI clears the searching out of the way and looks for gaps.
The pattern that repeats
Read together, the 14 cases tell one story. The ebook boils it down to three principles, and we agree:
- Data quality isn’t optional. Per the ebook, 70% of even top performers report data quality and governance problems when integrating AI. You can start with messy data, as PayJoy did, but fixing the source belongs in the plan.
- Augment before you cut. Nearly every case gave time back to the team for analysis and decisions. Only Walmart, because of its scale, cut hundreds of roles. Starting with augmentation lowers resistance and lets you learn.
- One process at a time. Nobody rolled out “AI across finance.” Fanatics fixed vendor identification; Tipalti, its memos. Pick the most annoying bottleneck, fix it, measure it, move on.
And a fourth, present in almost every case without being written down: AI prepares, a person decides. Lyft calls it a first pass; Wasabi, editing a couple of paragraphs; the CFO behind the SEC filing, validation. It is the same preparer-and-reviewer split every CFO already uses in internal controls.
What this means in practice
Half of these cases aren’t even agents: they are predictive models or AI tools used by hand, one request at a time. The next step is turning a one-off play into a workflow that runs on its own, with rules, approval thresholds and a log of every action.
- An AI Analyst finds the bottleneck that repeats: the weekly report, month-end AP, the pile of PDF invoices, the memo that takes days.
- A Forward Deployed Engineer builds the agent inside your operation, connected to the right sources, with a person approving anything outside the rules.
- Side-by-side training teaches the team to review instead of execute.
Forecasts, invoices and memos are just examples. If a task repeats, in finance or anywhere else, it can be automated.
The question isn’t which “AI strategy” to adopt. It’s which task your team repeats every month that nobody can stand anymore. Start there.
Enjoyed this article?
Get the next one before it hits the blog.
A short email per issue with what actually matters about AI agents at work. No spam.