AI In Finance
Five generations of automation. One thing none of them solved.
Each wave of finance automation reached further than the last. None of them remembered more. The reason is not technical, and it explains why AI in finance keeps stalling at pilots.
Finance has automated more than almost any other function, and it is the last function in the enterprise to deploy AI at scale. Both of those things are true, and the gap between them is worth understanding.
It helps to look at what has actually been built, one wave at a time, and ask a specific question of each: what did it solve, and what did it leave behind?
Five generations, one pattern
Automation in finance, by generation
| Generation | What it solved | What it left behind |
|---|---|---|
| Rules | Repeatable, well-formed transactions | Anything exceptional |
| Machine learning | Classification, matching, prediction | Why a pattern matters here |
| RPA | Swivel-chair labour across systems | It encoded no meaning |
| Generative | Extraction, drafting, summarising | Confident, ungrounded error |
| Agentic | Orchestration and follow-through | Traceability and accountability |
Read the middle column and the progress is real. Matching, posting, chasing and coding largely run themselves now. A finance team in 2026 does far less manual work than one in 2016.
Read the right-hand column and something else appears. Every generation extended reach: more formats, more systems, more steps in sequence. Not one of them extended memory.
Why the hardest part was never touched
The usual explanations for finance lagging are risk aversion, regulation and legacy systems. None of them survives much scrutiny. Finance has the cleanest data in the building, and regulated industries have automated plenty.
The actual reason
Automation only ever reached work that was written down. A checklist. A policy rule. A matching procedure. That is why AP, AR and reconciliation automated well: the work was already specified before anyone tried to automate it.
The connective work was never written down. That a particular vendor always short-pays freight. That a recurring variance was ruled immaterial two quarters ago. That this customer's payments land three days before the remittance advice. None of it was ever a documented procedure, so no automation project ever had a specification to attack.
Twenty years of automation, and that layer is untouched. Not because it is hard. Because it was never visible as work.
Reach without memory
This is why the newest generation is the most uncomfortable one. An agent that can reach six systems and act across them is reaching further than anything before it, while knowing no more. That is precisely the configuration that produces confident, plausible, wrong action.
There is a number that makes the point concrete. On ungoverned data, more than half of enterprise AI responses contain fabrications. Structure the context properly and hallucination rates fall by roughly 87%. Same models, same questions. The variable is not the model. It is whether the context around the data was ever captured.
You cannot reason your way to a defensible answer from entries alone. Entries are all most finance systems hold.
What a sixth generation would have to do
The next step is not a better model. It is a place to put judgement.
Three things would have to become properties of the work itself, measured and stored the way an amount or a posting date is:
- Anomaly, whether this is normal for this account, measured against its own history rather than a blanket rule
- Materiality, whether it matters enough for a person to look, applied on arrival rather than at close
- Memory, what the team concluded last time, stored against the item and its pattern
None of those are retrieved. They are accumulated. Which means they cannot be solved by connecting an AI to more systems, however elegantly. Access is a different problem, and it is the one currently being solved.
This is what connected finance means at the level of a single item, and it is why materiality and anomaly are worth treating as parameters rather than judgement calls.
The test
Take any tool your team has adopted in the last five years and ask what it remembers. Not what it can retrieve. What it remembers, without being asked, about the work it has already seen.
For most, the honest answer is nothing. That is not a criticism of the tools. It is the whole history of finance automation in one question.