There’s a version of the AI conversation in accounting that stays very abstract. Transformation. Disruption. The future of the profession. It’s easy to nod along to and hard to act on.

The more useful conversation is the specific one. Not what AI might do someday, but what firms are actually doing with it right now, where it’s delivering real value, and where it still falls short.

Transaction Review and Anomaly Detection

This is one of the clearest wins for AI in accounting workflows today. Instead of a staff member manually reviewing every transaction in a client’s books, AI can flag the ones that look out of place: unusual amounts, vendors that don’t match historical patterns, duplicate entries, mis-categorized expenses.

The human still makes the call. But they’re making it on a curated list of exceptions rather than combing through everything line by line. For firms handling high transaction volumes across multiple clients, that’s a meaningful reduction in review time without a reduction in quality.

Document Summarization and Extraction

Firms deal with a lot of documents. Contracts, bank statements, prior year returns, financial reports. AI is being used to extract key information from these documents and surface it in a usable format rather than requiring someone to read through everything manually.

This shows up in practical ways: pulling specific figures from a lengthy document, summarizing a contract’s key financial terms, or extracting data from bank statements that don’t export cleanly into accounting software. It’s not glamorous, but it saves real time on tasks that used to be purely manual.

Draft Client Communications

Writing takes time, and a lot of what firms write follows predictable patterns. Follow-up emails, document request lists, status updates, explanations of financial concepts for clients who need context.

AI is being used to produce first drafts of these communications that a team member then reviews, adjusts, and sends. The human judgment is still in the loop, but the blank page problem goes away. For firms with large client bases, that compounds into meaningful hours saved across the team.

Categorization and Coding Assistance

AI-assisted categorization has been part of accounting software for a few years now, but it’s getting meaningfully better. Firms are using it to speed up the bookkeeping side of their workflow, with AI suggesting how transactions should be coded based on historical patterns and the team confirming or correcting.

The correction piece matters. AI categorization isn’t perfect, and firms that treat it as a rubber stamp rather than a starting point run into data quality problems. Used correctly, as a suggestion layer with human review, it reduces the time spent on routine coding without introducing the errors that come from full automation.

Where It Still Falls Short

AI is not doing the judgment work. It’s not reviewing a client’s financial position and forming an opinion about what they should do differently. It’s not navigating a complicated tax situation, managing a client relationship, or making the kind of call that requires context, experience, and professional accountability.

The firms getting the most out of AI right now are clear-eyed about this. They’re using it to reduce time on the tasks that don’t require judgment so their people can spend more time on the ones that do. That’s a different framing than replacing accountants, and it’s a more accurate one.

What This Means for Your Firm

If your firm hasn’t started experimenting with AI in any of these areas, the gap between you and the firms that have is already opening. That doesn’t mean rushing into every new tool that comes out. It means being deliberate about where AI can reduce friction in your existing workflow and starting there.

The firms that will look back on this period as an advantage are the ones that treated AI as a practical operational question rather than a theoretical one. The practical question is simple: where is your team spending time on work that a well-configured AI tool could handle just as well?

Start there, and build from what you learn.