You’re not wrong that the day‑to‑day feels mostly unchanged — AI is still climbing the “boring middle” of the adoption curve. But under the hood of the software you already use, the shift is gathering speed:
What’s happening right now (mostly out of sight)
| Where | What’s new | Why it matters to accountants |
|---|---|---|
| Big‑Four firms | KPMG just rolled out Workbench, a multi‑agent platform that sits behind its audit, tax‑advisory and velocity tools. | Your external auditors will soon expect cleaner digital data feeds and may ask for AI‑ready exports rather than spreadsheets. |
| EY and Deloitte have launched “agentic” platforms built with Nvidia to automate tasks across tax, risk and finance. | The tools start with document ingestion (returns, contracts, work‑papers) and kick out reconciliations, variance explanations and draft working‑papers. | |
| ERP / ledger vendors | Microsoft Copilot for Excel, Dynamics 365 Finance, Xero “Just Ask”, QuickBooks Gen‑AI assistant all moved from preview to paid add‑ons in 2025. | Instead of formulas/macros you’ll type “look up customers >90 days past due and draft collection emails”. |
| Point solutions | Invoice coding, expense review, and AP/AR matching are the first spots where GPT‑style models are beating rule‑based bots. | These tasks are repetitive, high‑volume and sit on structured data — a sweet‑spot for current models. |
Why you haven’t felt it yet
- Risk & regulation first, roll‑out later. CFOs are piloting behind closed doors because material‑error risk in finance is non‑negotiable. Deloitte predicts only ~25 % of companies will run agent pilots this year, rising to 50 % by 2027.
- Data plumbing is the bottleneck. A model is only as good as the chart‑of‑accounts mapping and document taxonomy you feed it. Most firms are still cleaning that up.
- UX is being buried inside products. You won’t “open ChatGPT for accounting”; you’ll click a new button in the software you already use.
What the next three years are likely to bring
| Horizon | Concrete changes you’ll notice |
|---|---|
| Next 12 months | Draft footnote summaries, variance memos and audit PBC lists pre‑written in your work‑papers; AI chat boxed into “suggest” mode so you keep final say. |
| 18‑36 months | Ledger‑to‑trial‑balance close largely automated for routine entities; continuous audit hooks running in the background; self‑service chat for policy queries (“Does ASC 842 apply here?”). |
| 36 + months | Agentic workflows that trigger filings, payments or journal entries autonomously once confidence thresholds and segregation‑of‑duty checks pass. |
Surveys already show FP&A AI usage jumping from 6 % in 2024 to 47 % in 2025 – but most of it is still in pilot status.
How to kick the tyres today
- Copilot in Excel – ask it to explain a complex formula or build a pivot; it’s faster than Google + trial‑and‑error.
- Upload a 100‑page lease to ChatGPT – prompt: “Extract payment schedule, indexation clauses, IFRS 16 inputs.” The first run will show you where your data is messy.
- Automate reconciliations – tools like BlackLine, Tipalti or even a small GPT fine‑tuned on your GL exports can flag mismatch rows and propose entries.
- Shadow‑mode a close task – let an LLM create a depreciation journal, compare it to yours, and measure the error rate. Low‑risk, high‑learning.