A CFO at a logistics firm asked me last month why the “AI” tool his team had bought was still routing half the invoices to a person. The vendor had promised 99% accuracy. The team was drowning. Both things were true, and the gap between them is the whole story of AI in finance right now.
“AI in finance” has become a slide, not a capability. Almost every accounting tool, ERP add-on, and document vendor has bolted the word onto a product page. Some of it is real. A lot of it is OCR from a decade ago wearing a new label, sold with a number that measures the wrong thing. This post is my attempt to separate the two, using the finance processes I actually build agents for.
What finance teams are actually buying
When a finance team says it wants AI, it usually means one concrete thing: stop making people retype data off documents and into a screen. Invoices, bank statements, expense receipts, vendor forms. The work is high volume, low joy, and error-prone at 5pm on the last day of the month.
There are two very different products that both claim to solve this. One reads the pixels on a page and hands you text. The other reads the document, works out what the values mean, checks them against your other records, and only asks for help on the cases it cannot justify. The first is OCR. The second is agentic, and the difference shows up in your team’s workload, not on the pitch deck. If you want the longer argument for why agents differ from scripted automation, I wrote the pillar piece here: what agentic AI actually is.
The OCR-in-a-trenchcoat problem
OCR turns a smudge into a character. That is the entire job, and modern engines are very good at it. The trouble starts after the text exists. Which string is the invoice number? Do the line items add up to the total? Is the GST number on the master vendor list? OCR walks away at that point and leaves the reasoning to you, which usually means a wall of regex rules that hold for the first vendor and crack by the tenth.
So when a finance tool says “AI-powered” and the demo is just text coming off a scan, you have bought the easy 10% and kept the hard 90%. That is the trap the logistics CFO walked into.
Where agentic automation earns its place
Here are the four finance processes where I have seen agents pay for themselves quickly. None of them are exotic. They are the high-volume, well-bounded jobs where the right answer is knowable, which is exactly where automation should start.
The agent extracts the invoice, matches it against the PO and goods receipt, checks the maths, and posts the clean ones. It only flags what it cannot reconcile.
Transactions get parsed, normalised across formats, and matched to your ledger. Unmatched items come with a suggested reason rather than a blank cell.
Routine accruals and recurring journal entries are drafted continuously instead of in a panic on day one. The agent surfaces what looks off before you sign.
Duplicate payments, out-of-pattern amounts, and vendors that suddenly changed bank details get caught by behaviour, not by a fixed threshold someone forgot to update.
Accounts payable, end to end
AP is the clearest win because the right answer already exists somewhere in your system. The invoice should match a purchase order. The total should equal the sum of the lines. The vendor should be on file. An agent can check all of that, which OCR cannot.
What the agent does without a human
On a clean invoice from a known vendor, the agent reads the document, identifies each field, runs a three-way match against the PO and goods receipt, confirms the tax treatment, and queues it for posting. No person touches it. For a large share of a typical AP queue, that is the whole journey.
The one step that still needs a person
When the agent cannot reconcile something, a price that does not match the PO, a vendor whose bank details changed last week, it stops and routes the case to a human with its reasoning attached. That is the design goal, not a failure. An honest system raises its hand on the few cases that deserve a second look instead of quietly guessing.
Bank statement reconciliation
Statements are messier than invoices because every bank exports a different shape, and half of them arrive as scans. An agent normalises the formats, matches transactions to your ledger, and when something does not match, it proposes the likely counterpart instead of dumping the exceptions on your team. The reconciliation that used to eat a full afternoon becomes a short review of the genuinely ambiguous rows.
The numbers vendors quote vs. the ones you live with
This is where most finance automation projects quietly disappoint. The accuracy figure on the deck and the accuracy your team feels are measuring different things. Three numbers matter, and only one of them is what gets quoted.
On these numbers. These are ranges we have seen across PTAS deployments and internal benchmarks, not published industry figures. Your own mix of vendors, scan quality, and document types will move them, sometimes a lot. We share field-level results on your documents before anyone signs, precisely so the number you plan around is yours and not ours.
Character accuracy is not field accuracy
Character accuracy asks whether an “8” was read as an “8”. Field accuracy asks whether that 8 landed in the right box as part of the right total. Document accuracy asks whether the entire invoice came out correct with nothing missing. A vendor can be honest about 99% character accuracy and still leave your AP team correcting one invoice in three, because the reasoning that turns characters into trustworthy fields was never part of the product. When you evaluate a finance AI tool, ask which of the three the headline number refers to. The answer tells you most of what you need to know.
Closing the books and catching what rules miss
Month-end close
The close is painful because it is bursty. Everything happens in two or three days while the rest of the month sat idle. An agent flattens that curve by reconciling accounts as the month runs, drafting recurring entries, and flagging the items that look wrong early. You still own the close. You just walk into it with most of the work already done and a short list of things to check.
Anomalies a static rule never sees
Rule-based fraud checks only catch what someone thought to write a rule for. A flag at “over 50,000” misses the careful fraud that stays at 49,500. An agent works from behaviour: this vendor, this amount, this timing, this sudden change of bank account. It is not magic, and it generates its own false positives, but it catches a class of problems that a fixed threshold simply cannot. For finance teams that is the part worth paying for.
What to automate first, and what to leave alone
My advice is boring on purpose. Pick one process that is high volume, well bounded, and where you already know the correct answer. Invoice capture and matching is almost always the right first move. You can measure it, the payoff is obvious, and the agent has something concrete to check itself against.
Leave the judgement-heavy work alone for now. Anything that depends on negotiation, relationships, or a one-off call is a bad first project. So is low-volume work, where you will spend more time building the automation than you will ever save. Get one process paying off, learn how your team works with an agent in the loop, then expand. The firms that try to automate everything at once tend to automate nothing well.
How we approach finance automation at PTAS
We build on agentic document extraction rather than an OCR pipeline, because the reasoning step is the part finance teams actually need. Our product, DocPro, reads the document, validates it against your records, and escalates only what it cannot justify. For the finance workflow specifically, we wrote up how that plays out in a real AP queue here: agentic invoice processing for finance teams.
None of this replaces your finance team. It removes the retyping and the reconciliation grind so the people you hired for judgement get to spend their time on judgement. That is the version of AI in finance worth buying. The rest is a slide.
See it on your own documents. Send us 20 invoices, bank statements, or expense files. We will run them through DocPro and return a field-level accuracy breakdown with confidence scores, so you can plan around your numbers instead of ours.
Common questions
Is AI in finance the same as RPA?
No. RPA follows a fixed script and breaks the moment a layout or screen changes. An agentic system reads the document, reasons about what it found, checks it against your other records, and escalates only the cases it cannot justify. The difference matters most on the messy documents, which is where RPA quietly fails.
Why does a vendor’s 99% accuracy not match what my team sees?
Because the 99% is almost always character accuracy on clean documents. Your team lives with field accuracy on real mail, which is a different and lower number. Ask whether the figure is character, field, or document accuracy before you sign anything.
Where should a finance team start?
One high-volume, well-bounded process where you already know the right answer. Invoice capture and matching is the usual first move. Avoid judgement-heavy or low-volume work until the first project is clearly paying off.