AI for Accounting Firms: What's Real and What's a Sales Pitch

Every accounting software vendor just added "AI features" to their product. Intuit, Thomson Reuters, Wolters Kluwer — they're all in the game. Your inbox is full of pitches. Conferences are full of demos.

And you're running a 15-person firm wondering whether any of this applies to you.

Short answer: some of it does. But not in the way it's being sold.

What AI can actually do for your firm

Four things, specifically:

• Answer client questions from your knowledge base. Not generic tax advice — answers pulled from your firm's documents, policies, and procedures.
• Extract data from documents. Read W-2s, 1099s, K-1s, bank statements. Pre-populate your tax software. Not perfectly — a human reviews — but it cuts data entry significantly.
• Draft communications. Client emails, engagement letters, meeting summaries. Drafts, not finals. Your staff reviews and sends.
• Search your institutional knowledge. Every memo, training doc, email thread, and procedure your firm has ever produced — made instantly searchable.

That's the real capability set. Useful? Yes. Magic? No.

What AI cannot do

• Replace tax preparation judgment. Period.
• File returns autonomously.
• Handle complex advisory without human oversight.
• Work accurately 100% of the time.

Anyone selling you "autonomous AI accounting" is selling fantasy. AI is a productivity tool, not a replacement for your professionals.

Where to start

Pick one of these three:

• Client communication. Highest volume of routine work. Easiest to prove value. Start here if you're drowning in "when is my return done?" emails.
• Internal knowledge search. Best if your firm has lots of institutional knowledge locked in partners' heads and old memos. Makes everyone faster.
• Document data extraction. Best for firms with high-volume returns where data entry is a bottleneck.

Don't try all three at once. Prove one, measure the result, then decide on the next.

The only rule: test before you commit

Any vendor or consultant who won't let you test their solution on your actual data before signing a contract is a vendor to avoid. Full stop.