AI in finance is limited less by capability than by consequence. Accuracy has to be exact, the underlying data is often inconsistent, responsibility for a number stays with a person, and the oversight rules are still being written. Those constraints shape where it can safely be used.
Accuracy: close enough is not good enough
AI is a trendy topic, but finance is behind in adopting it because of the high risk attached and the critical importance of accuracy, unlike other areas where close enough is good enough.[1] That is the whole problem in one sentence. A marketing tool that is right 95 per cent of the time is useful. A reconciliation that is right 95 per cent of the time is a defect you now have to find.
In practice AI can handle 80 to 90 per cent of basic finance tasks like checking, reconciliation, analysis, forecasting or modelling, but the last 10 per cent requires a qualified human who understands what the challenge is and what good looks like as an outcome.[2] The residual is not evenly distributed either. It is concentrated in exactly the judgement calls that carry the consequences.
The data underneath is usually not ready
The businesses at the front of AI adoption are very digital and tech heavy, like software, because they are simpler businesses where it is easy to automate and plug AI in. When you are inventory-heavy and selling physical goods it is much harder, because of the nuances, and if you have legacy data issues or imperfect infrastructure it is very difficult, particularly at scale, so most larger businesses with an inventory or stock component are not very advanced at all.[3] Vendor case studies rarely come from those companies, which is worth remembering when you read one.
Nor is early stage the easy alternative. Early finance roles will always require a human in the loop because companies at that stage have not yet set up their processes.[4] You cannot automate a process nobody has defined. The middle is where automation works, and the middle is narrower than the marketing suggests.
| The claim | What holds up in a finance function | |
|---|---|---|
| Accuracy | AI can do the reconciliation work. | It can do 80 to 90 per cent of it. The last 10 per cent needs a qualified person who knows what good looks like, and that is where the consequences live. |
| Readiness | Any business can plug it in. | Digital businesses can. Inventory-heavy businesses with legacy data and imperfect infrastructure largely cannot, particularly at scale. |
| Accountability | The system produces the number. | A qualified person signs off anything reaching a board, investor, auditor, client or regulator. Directors remain personally responsible either way. |
| Cost | Pick the best tool for the job. | Budget and internal data policy usually decide it, which often means the tool already in the stack rather than the one that tested best. |
| Jobs | Finance headcount goes away. | Roles change rather than disappear. Backward-looking processing is most exposed; work that depends on talking to the business is least. |
Accountability cannot be delegated to a model
This is the constraint that does not move. My one rule for using AI in finance is that it produces a first draft, not a final answer. A qualified person must check every number and sign off anything that reaches a board, an investor, an auditor, a client or a regulator, so accountability stays human.[5] A model cannot hold a duty, and in Australia the duty is real: directors carry personal responsibility for the financial reports their company produces, which ASIC states plainly. No tool changes who is answerable.
The harder version of this is coming. As AI starts recommending financial actions, finance leaders will need to reassess where human oversight sits in relation to liability and risk governance.[6] Reviewing an output is one thing. Being accountable for a recommendation you did not derive and cannot fully reconstruct is a different exposure, and most delegated authority frameworks do not address it yet.
That oversight question belongs alongside your other controls, in financial governance and who should own it.
Data privacy, compliance and cost
Simply playing around with AI in finance is not simple. There are risks with internal data, compliance issues that are sharper in sectors like healthcare, and financial constraints that may push a business towards a tool bundled with its existing stack rather than a premium option.[7] Those three arrive together, and the third one usually decides the outcome. The tool that gets used is the one already covered by an enterprise agreement, not the one that tested best.
The transparency problem sits underneath all of it. If you cannot explain how a number was derived, you cannot defend it to an auditor, and a finance output you cannot defend has no value regardless of whether it happened to be correct.
What this means for hiring
Not that finance jobs disappear. Everything is moving towards automation and AI, which means finance roles change rather than the industry going away.[8] The distinction that matters is between backward-looking processing work and forward-looking work that depends on talking to the business.
What employers are actually asking for reflects that. They highly value candidates who are adaptable and translatable with tech, particularly around automation, process improvement and a genuine interest in AI, without expecting expert-level capability.[9] That is a reasonable bar, and it is a different bar from hiring an AI specialist into a finance seat. The person you want can read a machine-generated reconciliation and tell you why the last 10 per cent is wrong. Our guide to getting started with AI in finance sets out the workflows someone at that bar would be expected to know.
The safest place to start applying any of this is the assembly work inside a month-end close.
Common questions
What are the main disadvantages of AI in finance?
Five things. Accuracy standards are unforgiving, because finance is one of the few functions where close enough is not good enough. The underlying data and systems are often not ready, particularly in inventory-heavy businesses with legacy infrastructure. Accountability cannot be delegated to a model, so a qualified person still has to check and sign off. Budget and internal data policy usually decide which tool a finance team actually gets, which is rarely the one that tested best. And the oversight rules around AI-generated recommendations are still being worked out, which leaves a liability gap most delegated authority frameworks do not yet cover.
Why is finance slower than other functions to adopt AI?
Because the risk attached is significant and accuracy is critical, unlike areas where an approximately right answer is useful. A reconciliation at 95 per cent right is not 95 per cent useful, it is a defect you now have to locate. That is a rational reason to move slowly, not a cultural failure, and it is why the sensible starting points are low-risk, high-frequency workflows that do not touch sign-off-grade numbers.
Can AI replace a finance team?
No, but it changes what the team does. AI can handle 80 to 90 per cent of basic checking, reconciliation, analysis, forecasting and modelling work. The last 10 per cent needs a qualified human who understands the actual problem and what a good outcome looks like. Everything is moving towards automation, which means finance roles change rather than the industry disappearing. Backward-looking processing shrinks the most, and the work that survives depends on talking to the business.
Who is accountable when an AI tool gets a number wrong?
A person, always. The rule I apply is that AI produces a first draft, not a final answer: a qualified person checks every number and signs off anything that reaches a board, investor, auditor, client or regulator. In Australia, directors carry personal responsibility for the financial reports their company produces, and no tool changes that. The open question is oversight of AI-generated recommendations, where liability and risk governance still need to be worked through.
References
- My read on adoption: AI is a trendy topic, but finance is behind in adopting it because of the high risk attached and the critical importance of accuracy, unlike other areas where close enough is good enough.
- Where I think the line sits: AI can handle 80 to 90 per cent of basic finance tasks like checking, reconciliation, analysis, forecasting or modelling, but the last 10 per cent requires a qualified human who understands what the challenge is and what good looks like for the outcome.
- What I observe about who is actually ahead: the businesses at the front of AI adoption are very digital and tech heavy, like software, because they are simpler businesses where it is easy to automate and plug AI in. Inventory-heavy businesses selling physical goods find it very difficult because of the nuances, and legacy data issues or imperfect infrastructure make it harder still at scale, so most bigger businesses with an inventory or stock component are not very advanced at all.
- Why early stage is not the easy case: early finance roles will always require a human in the loop, because companies at that stage have not yet set up their processes.
- My one rule for using AI in finance: it produces a first draft, not a final answer. A qualified person must check every number and sign off anything that reaches a board, investor, auditor, client or regulator, so accountability remains human.
- My view on where this is heading: as AI starts recommending financial actions, finance leaders will need to reassess where human oversight sits in relation to liability and risk governance.
- Why experimenting is harder than it sounds: simply playing around with AI for finance is not simple, given risks with internal data, compliance issues especially in healthcare, and financial constraints that might mean using a tool like Copilot over a premium option.
- What I tell finance people about the trajectory: everything is moving towards automation and AI, which means finance roles will change rather than the industry disappearing altogether.
- What employers actually ask for: candidates who are adaptable and translatable with tech, particularly in automation, process improvement and an interest in AI, without necessarily expecting expert-level AI capability.