The State of AI in Accounting: What Changed This Year

The calendar turned on another year of AI-in-accounting hype. The honest read on what changed in the last twelve months is smaller than the marketing would suggest and larger than the skeptics want to admit.

The big changes are three.

Document extraction got materially better. A year ago, the best tax-document ingestion tools were reliable for cleanly scanned W-2s and 1099s and unreliable for anything handwritten, multi-page, or unusually formatted. Today, the leading tools handle messy K-1s with multi-state addenda, handwritten engagement notes, and multi-column statements of partner capital accounts with accuracy that approaches a junior preparer’s first pass. The work of extracting data from source documents, which was a bottleneck in every firm’s workflow, is now mostly automated. The firms that still have staff typing numbers off PDFs are running infrastructure a year behind current state.

Research assistants crossed the trust threshold. Claude, ChatGPT, and a handful of purpose-built tax research tools now produce code-section lookups, case summaries, and interpretive explanations at a quality where we use them as a starting point. The output still requires human review, and we never take an AI’s word as final. But the time from question to first-draft answer has collapsed. A question that took 45 minutes of reading a year ago now takes 8 minutes of reading a well-sourced AI response plus 15 minutes of verification. That is a meaningful productivity shift, especially for judgment-heavy work.

Workflow automation reached small firms. Twelve months ago, workflow automation tools were built for large national and regional firms. They required IT resources, configuration, and integration work that small and midsize firms could not justify. Today, lighter-weight tools (Zapier, Make, Airtable with AI integrations, purpose-built firm-ops products) can automate recurring engagement workflows at a cost and complexity level that fits a ten-person firm. The firms that implement these are running on fewer hours per engagement. The firms that do not are running on last year’s cost structure.

The non-changes are real.

Judgment work did not get automated. A return that requires thinking about the operating agreement, the partner mix, the state apportionment, and the owner’s non-tax objectives still requires a human to sit with it. AI output on these questions is directionally useful and specifically dangerous when trusted unreviewed. The professional judgment layer did not shrink.

Client relationships did not change. The client who calls at 9pm with a panic about a notice still wants a human on the other end. The client deciding whether to convert entities, structure a partner buyout, or sell the business is not going to accept AI output as the decision. The relationship portion of the work, which is most of the value clients pay for, did not automate.

Advisory strategy is still manual. The year-end meeting we described in an earlier piece, the strategic conversations about structure and direction, the judgment calls on complex returns — none of these got touched. AI did not attend any of our client meetings. It contributed to preparation and follow-up. It did not replace the meeting itself.

The honest summary: AI made existing good firms faster at what they already do. It did not make bad firms good. And the firms not using AI at all are increasingly behind on their cost structure, which eventually becomes behind on their competitiveness.

---

This briefing is for informational purposes only and does not constitute tax advice. The topics discussed depend on specific facts and current law, both of which change. A proper analysis of your situation requires professional review. Contact us to discuss whether this applies to your business.