AI Can Do Half Your Accounting Work (But Don't Fire Your Accountant Yet)

The Problem You Recognize

You have a mountain of invoices, receipts, and bank statements to process every month. Your accounting team is stretched thin, working late nights on repetitive data entry. Human error is costing you money. You've heard AI can help, but you don't know if you can trust it with your company's financials.

What Researchers Discovered

Researchers put today's best AI models through real accounting work. They tested tasks like reconciling bank statements, calculating expenses, and preparing financial reports.

The results were clear: AI can handle about 56% of the work. But there's a critical catch.

The AI is wildly inconsistent. It might complete a task perfectly one time, then make critical calculation errors the next. Across multiple attempts, it only succeeded completely on a task 2.6% of the time.

Think of it like hiring a new accountant who sometimes produces perfect work but often makes mistakes that could cost your company thousands. You wouldn't let them work unsupervised.

The research also revealed something surprising: throwing more money at the problem doesn't solve it. Some efficient, less expensive AI models performed nearly as well as the most expensive ones. The most expensive option isn't always the best value.

When AI fails at accounting, it's almost always because it can't reason through the problem correctly—like misinterpreting a tax rule or making poor judgment calls. It has all the right tools (access to spreadsheets, financial software) but struggles with the business thinking that human accountants apply.

Finally, researchers found that building fancy, custom systems for AI provided almost no performance boost. Giving AI a specialized, custom-built platform didn't make it significantly better than using standard, off-the-shelf systems. The AI's capability mattered more than the tools it used.

You can read the full research paper here: APEX-Accounting

How to Apply This Today

Here's your 3-step plan to implement AI in your accounting department within the next 90 days.

Step 1: Identify Your "AI-Assist" Tasks

Don't try to automate everything at once. Start with the repetitive, procedural work that consumes hours but requires minimal judgment.

Your targets:

• Data entry from PDF invoices into your accounting system
• Initial bank statement reconciliation passes
• Categorizing routine expenses
• Extracting numbers from financial documents

Tools to use:

• For PDF processing: Try UiPath, Abbyy, or Microsoft Power Automate
• For data extraction: Look at Rossum, Hyperscience, or even OpenAI's GPT-4 with document understanding
• For reconciliation: Start with basic scripting in Excel or Google Sheets using their built-in AI features

For example: Set up a system where AI reads all incoming vendor invoices, extracts the amount, date, and vendor name, and pre-fills an entry in your accounting software. Your accountant then reviews and approves each entry. This cuts data entry time by 60-70%.

Step 2: Build Your "Human-in-the-Loop" Workflow

AI does the grunt work. Humans do the thinking. Design workflows where AI handles the repetitive parts and flags anything unusual for human review.

Here's how:

1. Map your current process: Document exactly how a task is done today
2. Identify the AI's role: Where can AI take over repetitive steps?
3. Insert checkpoints: Where must a human review the AI's work?
4. Create escalation paths: What happens when the AI is unsure?

For example: Month-end closing typically involves matching hundreds of transactions. Have AI perform the initial matching using rules-based algorithms. When transactions don't match perfectly (different amounts, dates slightly off), the AI flags them for your accountant. Your accountant reviews only the 20-30% of transactions that need human judgment, not 100%.

Step 3: Start With Cost-Effective Models

You don't need the most expensive AI. Start with mid-tier models that offer good performance at reasonable cost.

Your evaluation criteria:

• Accuracy: Can it achieve 85%+ accuracy on your specific tasks?
• Cost: What's the cost per transaction or document processed?
• Integration: How easily does it connect to your existing systems?
• Speed: How quickly can it process your volume of work?

Practical approach:

1. Test 2-3 different AI services on a sample of 100 documents
2. Measure accuracy, cost, and processing time
3. Choose the one that offers the best balance—not necessarily the highest accuracy
4. Start with a pilot of 1,000 documents per month

Teams using this approach typically see 30-40% faster processing of routine accounting tasks within the first three months.

What to Watch Out For

1. AI cannot be trusted end-to-end. The research confirms that AI still makes too many mistakes for complete automation. You must maintain human oversight, especially for complex tasks like calculating accruals, making judgment calls, or handling exceptions.

2. Real-world data is messier than test data. The research used synthetic company data. Your actual invoices, receipts, and statements will have irregularities, poor formatting, and edge cases that AI may struggle with. Expect performance to be 10-15% lower initially as you train the AI on your specific data.

3. The biggest gains come from workflow redesign, not just AI implementation. Simply dropping AI into your existing processes won't work. You need to redesign how work flows between humans and machines. This requires changing team roles, responsibilities, and approval processes.

Your Next Move

This week, pick one repetitive accounting task.

Gather 50-100 examples (invoices, receipts, bank statements). Test how long it takes your team to process them manually. Then try one AI tool—even a simple one like Microsoft Power Automate or Zapier with AI features—to automate part of the process.

Measure the time saved and accuracy achieved. If you save even 30% of the time with acceptable accuracy, you've proven the concept. Scale from there.

Question for your team: What's the single most time-consuming, repetitive task in your accounting workflow that you'd trust AI to help with first?