Accountants are being told that AI will transform their profession, and sold tools to prove it. Some of the claims are true. Many are not yet. This article separates the two, from the perspective of people who build AI systems for finance workflows and have also done the work by hand.
What language models are good at
Large language models are good at four things that matter in finance.
Reading unstructured documents. Invoices, receipts, statements, contracts, letters. A model can extract the supplier, the date, the line items, the VAT and the payment terms from a PDF or a photograph, with accuracy that is now good enough for a first pass on most documents.
Classifying from examples. Given a few hundred historical transactions and their coding, a model can code the next one, and give a confidence score. The same applies to classifying queries, documents and exceptions.
Drafting narrative from numbers. Variance commentary, management account notes, client summaries. Given the figures and the context, a model produces a competent first draft that a person edits rather than writes.
Answering questions over a known body of information. With access to the right data and documents, and with citations back to the source, a model can answer "what did we pay this supplier last year?" or "what does this lease say about break clauses?" quickly and usefully.
What they are bad at
Arithmetic. A language model should never be the thing that adds up the total. Any system that touches the accounts needs deterministic code doing the calculations and the model doing the reading and writing around it.
Being right every time. Models make confident mistakes. The design question is not whether the model will be wrong, but how the system catches it when it is.
Knowing what it does not know. Confidence scores help, but a model will sometimes be confidently wrong. Hence the review step.
What to automate first
In practice, the best first projects share three qualities: high volume, clear rules, and a cheap way to check the output.
Supplier invoice processing. Extract, code, match to a purchase order or a recurring pattern, and queue exceptions. The volume is high, the rules are known, and a wrong coding is easy to spot and cheap to fix. For most practices and finance teams this is the right place to start.
Bank and card transaction coding. The same pattern. Code from history, score confidence, route anything uncertain to a person. Accuracy improves as the corrections feed back.
Document capture for onboarding and compliance. Identity documents, engagement letters, prior-year accounts, into structured records. Dull, frequent and well-suited to the technology.
First-draft commentary. Month-end narrative, client summaries, board notes. Low risk, because a person always reads and edits before it goes anywhere.
What to keep human, for now
Anything that is signed off. The accountant signs the accounts, not the model. AI prepares; a qualified person reviews and approves. This is a matter of professional responsibility as much as accuracy.
Judgement calls. Provisions, going concern, revenue recognition for unusual contracts, tax positions. The model can gather and summarise; it should not decide.
Client advice. AI can prepare the analysis. The conversation with the client, and the recommendation, remain with the adviser.
How to introduce AI without putting the numbers at risk
Keep the model away from the arithmetic. Numbers live in a database. Calculations run in tested code. The model reads documents and drafts text.
Set confidence thresholds. Above the threshold, the system acts (or proposes). Below it, a person decides. Start conservative and widen as you gain evidence.
Log everything. Every input, every output, every decision and who approved it. An auditor, a regulator or a client should be able to trace any number back through the system.
Keep data in your own environment. Use enterprise model endpoints that do not train on your data, within your own cloud tenant, with access controls and data processing agreements in place.
Measure accuracy, not enthusiasm. Run the model alongside the current process for a month and compare. Decide based on the error rate and the time saved, not the demo.
The economics
For an accountancy practice, automating invoice processing and transaction coding changes the cost of bookkeeping and compliance services, which is where fee pressure is greatest. For an in-house finance team, it removes the repetitive work around the close. In both cases the staff do not disappear; they move to review, exceptions and client or stakeholder work, which is the work they trained for.
A realistic view
AI in finance is neither the revolution the vendors describe nor the fad the sceptics dismiss. It is a set of capabilities that, applied to the right processes with the right controls, remove a great deal of repetitive work and make the remaining work more valuable. The firms and finance teams that benefit will be the ones that start with a specific process, measure the result, and build from there.
If you want a specific answer to "what should our finance team do about AI?", that is the question we help clients answer, and then build.