In 1983 an accountant with Lotus 1-2-3 could change one number and watch a month’s workings recalculate before the coffee went cold. The accountant at the next desk, with a ledger and a calculator, had the afternoon ahead of them. Within ten years the second accountant didn’t exist. Nobody called it automation. It was just the spreadsheet, and finance took it for granted so completely that we forgot it was the first time the ceiling on the job moved from the person to the process. (VisiCalc got there first, in 1979. Lotus is the one that reached every finance desk, so it gets the credit here.)
I qualified after 1985, so I’ve never worked in a finance function without one. I’ve spent 15 years in finance operations, and most of that time automating them: scheduled loads instead of exports, a model instead of linked workbooks, a script instead of the same Tuesday afternoon repeated forever. All of that was before AI was a thing. For the last 18 months I’ve been building AI systems inside finance teams, on real packs and real ledgers, and my view from inside is that the ceiling is moving again and most of the market hasn’t noticed. Gartner’s November 2025 survey of 183 finance leaders put AI use in finance at 59%, against 58% the year before. Adoption has stalled, and the top live use case, at 49%, is knowledge management, which mostly means search.
So this is the post I wish someone had handed me 18 months ago. By the end of it you’ll know nine things AI already does inside finance teams, with a picture of each so you can recognise it when you see it. You’ll know which one to start with (the least glamorous, and the cheapest to measure). And you’ll have a test, borrowed from the way you already judge spreadsheets, for telling a build that will help from one that will quietly make things up. None of it needs a new ERP.
One disclaimer before the list. I’ve built each of these for clients and they’re in production. But no two looked the same, because every finance team has its own weird and wonderful setup: the ERP that was customised in 2014, the planning tool nobody upgraded, the one supplier who still faxes. The pictures below show the shape of each thing, with made-up names and numbers. Your version will look different, and it should.
The nine, so you can jump to the one that hurts most:
What the spreadsheet actually changed
Before the spreadsheet, an accountant’s day held a fixed number of calculations. Work more hours and you got a few more; the ceiling was the person. After it, the number of calculations a finance team could do was set by what the business needed, not by how many hands were in the room. A model with ten thousand formulas recalculates as easily as one with ten. That is the whole trick, and it’s why the spreadsheet won: it didn’t make anyone a faster accountant, it made the count of calculations somebody else’s problem.
AI, built properly, does the same thing one level up, to tasks. Today a finance person can get through a fixed number of tasks in a day: so many questions answered, so many invoices checked. Build the systems well and that number stops being a property of the person and becomes a property of the process. The organisation asks for as many as it needs, the same way it asks for as many calculations as it needs. (I went back and forth on whether “tasks” is the right word. It is, as long as you count checking as a task.)
That’s the whole argument. The rest of this post is what it looks like in practice, one task at a time.
1. Answers with the source attached
The most basic thing that works is a system that has read every file the team owns and answers questions from them, with the exact cell or page it took each figure from. Retrieval-augmented generation, if you want the term. Finance teams spend a startling share of the week finding numbers that already exist in their own output, then checking that the version they found is the right one. The checking is the expensive part, because you can’t skip it.
I’ve written about how this works and how to test it elsewhere on this site, so I’ll keep it to one point here. The saving is the search and the validation together. An answer with the workbook and cell attached is an answer you can check in ten seconds instead of rebuilding in an hour, and a system that says “I can’t find that in the documents” when it can’t is one you can hand to someone who wasn’t in the room.
It’s worth knowing how far this scales. Cerebras, the chip company, wrote up the knowledge base they built for their own staff this year: more than 15,000 questions a day, three months after launch. The part I keep coming back to is who asks. Their phrase is “humans, automations and agents”. The retrieval is exposed as tools that other systems can call, so a script or an agent can ask the knowledge base a question mid-task the way a person would. That links this item to items 8 and 9 below. The reconciliation that runs overnight can ask the same knowledge base which supplier a mangled bank reference belongs to, or what the policy says about a duplicate, and carry on.
2. Reading the contract before deciding
Lawyers got here first. Researching precedent used to mean days in the case law, and systems that retrieve the relevant cases and cite them are now ordinary in the profession. It moved quickly once it saw them work. Thomson Reuters’ 2026 report has generative AI use across professional services at 40% of organisations, up from 22% a year earlier, with industry-specific tools up 14 points in a year. The detail I like is that only 17% of legal professionals are comfortable letting AI give the advice. They trust the retrieval and keep the judgment, which is the right split, and it maps straight onto finance.
Finance’s case law is contracts and policies. Can we exit this supplier at year end without a fee? What does the lease say about the break? Is this spend capitalisable under our own policy? What exactly did we promise the lender in the covenant? Today the process is “let me find the contract”, twenty minutes of reading, then maybe a message to legal. With the contracts, the policies and the board minutes readable and cited, the decision is made on what the documents say, with the clause quoted, and the conversation with legal starts from the clause instead of from a search.
3. The questions nobody asks
Every finance team has a class of questions that never get asked, because answering them would take two days and the person asking knows it. What happens to Q4 cash if the US hire slips two months and the dollar moves five percent? What’s our exposure if the largest customer pays 30 days late through the winter? Which three projects would I pause first if the raise is delayed?
These are the questions that decide things, and they get answered by feel, because the real cost of a proper answer is a weekend. That cost is invisible. It never shows up on a timesheet, because the work never happens.
When a system can read the forecast, restate the assumptions and run the arithmetic while you watch, the question gets asked. The scenario is not cleverer than the one an analyst would build. It just exists, and the one the analyst would have built doesn’t.
4. The gap between what the software promised and what it does
Every finance system was sold on a demo where the data flowed. Then it went live, and somewhere in the process there’s a step where someone exports to CSV, fixes it in Excel, and uploads it again. Into the same system, or into the next one along. The vendor’s answer is a change request with a quote attached. So the export stays, for years, and it becomes someone’s Tuesday.
This is the least glamorous item on the list and the one I’d start with in most teams (and the one vendors least like hearing about). The fixes someone applies in Excel every month are rules: rename these columns, split that field, map these cost centres, drop the rows with no period. Rules can be written down, run on a schedule, checked against a control total and logged. AI helps at the edges, working out the mapping from a messy extract or reading a layout that changed without warning, but the win is mostly that the loop runs without a person inside it. The expensive software finally does what the demo showed, with a bridge you own.
5. Knowledge that survives the people who hold it
Why does the lease sit in the UK entity when the office is in Dublin? Why is that supplier paid in euros? What did the board actually agree in March about the hiring freeze? Finance teams carry hundreds of these decisions, and most of them live in one person’s head, plus an email thread nobody will find.
A system that has read the board minutes, the contracts and the model change log can answer “why” as well as “what”. And if the team logs decisions into it as they happen, two lines each, the log becomes the memory.
The version of this I’d push hardest is small. Not a company-wide knowledge base, a project one. Every project a finance team runs, an ERP migration, a planning build, a re-org of the cost centres, produces a few hundred decisions and the reasons behind them. The reasons are the valuable part and they’re the part nobody writes down, because at the time everyone in the room knows them. Then the people leave. Six months after a contractor’s last day the reasons have gone with them, and the team is maintaining a mapping table it can’t explain. A folder per project, the documents in it, the decisions logged as they’re made, and a system you can ask “why” is a cheap habit that pays out for years.
6. New starters who don’t need to find the veteran
Onboarding into a finance team is mostly a hundred small questions asked of the one person who’s been there long enough to know. That person’s time is the most expensive in the department, and the questions arrive at the pace the new starter hits them, which is exactly when the veteran is busy.
Point the new starter at the system that has read the policies and the last twelve months of packs. They get answers with the source attached, at their own pace, and they turn up at the veteran’s desk with only the questions that need a human. The second answer in the picture is the one I care about: a good system will tell you the intranet page is out of date, because it can see the newer file.
7. The CFO’s questions, answered in the meeting
There’s a phrase every finance person has said in a board meeting or an exec call: “I’ll check with the team and come back to you.” It’s the right answer when the alternative is a guess. It also delays a decision by a day when the decision was ready to be made.
With the pack, the ledger detail and the forecast readable by a system that cites what it reads, the CFO or the FD can answer in the room. What was the gross margin on the enterprise tier last quarter? When does the Dublin lease break? How much of the marketing overspend is timing? The person asking gets a number with a source. The finance team gets one less item on Monday’s list. This is the same capability as item 1, pointed at the person who sits in the most meetings.
8. Work that runs while you’re not watching
Most finance automation I’ve built doesn’t have a user. It runs overnight or on a schedule, and the only time a human hears from it is when something is wrong. The silence is the point of it.
The invoice queue is the usual example and it’s a fine one, but the version I find more convincing is reconciliation. Bank feed against ledger, every morning. Supplier statements against the purchase ledger, when they arrive. Intercompany balances, both sides, before anyone starts the close. Each run matches what it can, posts what it’s allowed to, and writes one message when it finds something it can’t resolve.
The AI part here is small and specific: reading a supplier statement that arrives as a scanned PDF, working out which invoice a bank line refers to when the reference is mangled, drafting the query email to the supplier. The discipline part is bigger: every run logs its inputs, outputs and what it decided, so an auditor can replay any morning. I was building it that way before the models were good enough to help with any of it.
9. Work you kick off, that used to be a week
Then there’s the work you have to start yourself, but no longer have to do by hand. My standing example is supplier invoices. Forty suppliers, and between them they send PDFs, scans, spreadsheets, invoices typed into the body of an email, and the occasional photograph. Every one of them has to end up in the same shape in the ledger, with the right PO, the right tax treatment and the right period.
I cut one team’s invoice processing from 60 hours a month to about two, before any of this. What’s changed since is the reading. A model can now take a scanned invoice in a layout it has never seen and pull out most of the fields correctly, most of the time, with the checks behind it catching the rest. That used to be the step that needed a person for every new supplier. Validation, the standard format and the upload could always be automated. The intake couldn’t, until recently, so the whole chain now runs and the person sees the two exceptions instead of the forty invoices.
Three more I’d add
Looking at what other teams are running, and what I’ve built since, three more belong on the list.
Variance commentary, first draft. Month-end commentary is written by someone staring at a variance report and typing what they already know. A system that has the variance detail, last month’s commentary and the forecast assumptions can draft the paragraph, with the drivers it pulled from the ledger listed underneath. The analyst edits instead of composing, and the judgment is still theirs.
Policy checks before approval, not after. Expense claims, supplier contracts, new vendor requests: each one is checked against a policy that lives in a PDF somebody wrote in 2021. A system that has read the policy can screen each item on the way in and flag the ones that need a look. Gartner’s survey has error and anomaly detection as the third most common finance AI use at 34%, and this is the practical version of it.
The audit request. “Please provide support for the following 40 samples.” Each sample means finding the invoice, the approval, the payment and the contract, then bundling them. With the documents already read and linked, that becomes a query rather than a fortnight.
Better and worse ways to build it
The Excel analogy works a second time here. Everyone in finance knows there are good spreadsheets and bad ones: hard-coded numbers buried in formulas, a model nobody but its author can change, inputs and outputs on the same tab. The tool is the same. The discipline isn’t, and the bad model costs the team for years.
AI systems are exactly the same. The difference between a system that helps and one that quietly makes things up is mostly build discipline: whether every figure carries its source, whether the system refuses when the documents don’t support an answer, whether there’s a test set of questions with known answers that gets run every time something changes, whether each run leaves a log an auditor could follow. None of that is visible in a demo. All of it is visible six months in.
I spend a large share of my time on exactly this, learning how to build these systems the right way, because the market is going to arrive here eventually and I want the good version ready when it does. The technology is there already. The discipline is what’s scarce, and it’s the part I’d look at first.
What I’d tell your CFO
All nine are running for clients today, in ordinary finance teams, on ordinary documents, and none of them needed a new ERP. Each one was built around the setup that team already had, and most look boring from the outside. Start with the export-fix-upload loop your team has stopped noticing, because it’s the cheapest to close and the easiest to measure. Then put a system that cites its sources over the pack. Then let the reconciliations run overnight and only speak when something’s off. Lotus 1-2-3 took the ceiling off calculations four decades ago. This takes it off tasks, provided it’s built the way you’d want a model built.
Sources
- Gartner, Finance AI adoption remains steady in 2025, press release, 18 November 2025. Survey of 183 CFOs and senior finance leaders: 59% using AI; top use cases knowledge management 49%, accounts payable 37%, error and anomaly detection 34%. Summary at CFO Dive.
- Cerebras, How we built our knowledge base, 2026. More than 15,000 questions a day from “humans, automations and agents”; retrieval exposed as tools that other agents can call.
- Thomson Reuters Institute, 2026 AI in Professional Services Report. Generative AI use at 40% of organisations, up from 22%; industry-specific tools up 14 points; 17% of legal professionals comfortable with AI giving legal advice. Legal-team summary on the Thomson Reuters legal blog.
- The 12-days-to-8-hours close and the 60-hours-to-2 invoice figures are from my own work and are written up elsewhere on this site. The names and numbers inside the figures are illustrative.