AI is not magic (it's pattern matching)
Let's start with the most important thing to understand: AI is not intelligent in the way humans are intelligent. It doesn't "understand" anything. It doesn't have judgment or common sense.
What AI does extremely well is pattern matching at scale. It looks at millions of examples and learns to recognize patterns. When it sees something new, it tries to match it against patterns it has seen before.
Think of it like a very experienced bookkeeper who has seen 10 million receipts. When they see a new receipt from "Tim Hortons," they instantly know it's probably a meal expense—not because they understand coffee, but because they've seen "Tim Hortons" categorized as meals thousands of times before.
That's essentially what AI does, just much faster and at a larger scale.
OCR: How computers read text from images
OCR (Optical Character Recognition) is the foundational technology that lets computers "read" text from images. When you photograph a receipt, OCR is what extracts the text.
How OCR works (simplified)
Image preprocessing
The system adjusts contrast, removes shadows, and straightens the image to make text clearer.
Character detection
The system identifies where text appears and isolates individual characters.
Character recognition
Each character is compared against known letter/number shapes. "That squiggle looks like a 7."
Context correction
If a word looks like "Tlm Hortons," it corrects to "Tim Hortons" based on common words and vendor names.
Modern OCR is remarkably accurate—often 98-99% for clean, well-lit documents. The challenges come with:
- Faded thermal paper (the kind most receipts use)
- Crumpled or folded documents
- Poor lighting or shadows
- Handwritten notes
- Unusual fonts or logos
Machine learning: How AI learns from examples
Machine learning (ML) is the technology that lets AI improve over time by learning from examples. It's what powers auto-categorization in accounting tools.
Here's a simplified explanation:
Training data
The AI is shown millions of examples: "This receipt from Staples was categorized as Office Supplies. This one from Shell was categorized as Vehicle Expenses."
Pattern extraction
The AI finds patterns: "When vendor name contains 'Staples' or 'Office Depot,' it's usually Office Supplies. When it contains 'Shell' or 'Esso,' it's usually Vehicle."
The key insight is that the AI doesn't "know" that Staples sells office supplies. It has simply seen enough examples to recognize the pattern.
This is why machine learning gets better over time—the more examples it sees, the more patterns it can recognize. It's also why your corrections matter: when you fix a categorization error, you're teaching the system a new pattern.
LLMs: The technology behind ChatGPT
Large Language Models (LLMs) are the newest AI technology, powering tools like ChatGPT, Claude, and Google's Gemini. They work differently from traditional OCR and machine learning.
LLMs are trained on massive amounts of text—essentially the entire internet—and learn to predict what words should come next. This gives them an impressive ability to:
- Understand context and nuance in language
- Answer questions in natural conversation
- Summarize and extract information from documents
- Handle ambiguous or messy data
For accounting, LLMs are increasingly used for:
Document understanding
Making sense of unstructured documents like contracts, letters, or unusual invoices.
Intelligent categorization
Understanding that "WestJet - YYC to YYZ" is travel, even without seeing that exact format before.
Answering questions
"What were total office expenses in Q3?" answered in plain English.
Anomaly detection
Spotting things that "look unusual" based on context, not just rules.
LLMs can "hallucinate"
LLMs sometimes generate plausible-sounding but incorrect information. This is why human review remains essential, especially for tax-sensitive categorizations and compliance work.
How these technologies apply to accounting
In modern accounting software, these technologies work together:
Step 1: Document capture
You photograph a receipt or upload a document. The system receives an image file.
Step 2: OCR extraction
OCR reads all text: vendor name, date, amounts, line items, taxes. Converts image to structured data.
Step 3: ML categorization
Machine learning matches the vendor and transaction type against patterns. Suggests a category.
Step 4: LLM validation (newer tools)
LLM reviews for anomalies, understands context ("client lunch" vs "personal meal"), flags potential issues.
Step 5: Human review
You review flagged items, correct any errors. Your corrections feed back into the learning system.
What AI can't do (and what's changing)
The landscape is evolving rapidly. With the rise of agentic AI workflows—systems where AI can reason, plan, and execute multi-step tasks—some previous limitations are disappearing. Here's the current state:
What's becoming possible with agentic AI:
- Complex multi-step workflows: Agentic systems can now chain together tasks: extract data → categorize → validate against rules → flag exceptions → generate reports. What once required multiple tools and human handoffs can happen automatically.
- Contextual judgment: Modern AI can understand that a $500 dinner at a steakhouse might be reasonable for a law firm entertaining clients but unusual for a sole proprietor plumber.
- Basic tax optimization: AI can now identify straightforward opportunities—unclaimed deductions, missing credits, common planning strategies—though complex tax planning still requires human expertise.
- Exception handling: Instead of just flagging anomalies, agentic AI can investigate them: checking supporting documents, cross-referencing patterns, and suggesting resolutions.
What still requires humans:
- Strategic advice: "Should I incorporate?" "What's the optimal salary vs dividend mix?" These require understanding the client's full situation, risk tolerance, and long-term goals.
- Client relationships: AI can process documents, but the trusted advisor relationship—understanding a client's business, their concerns, their aspirations—remains deeply human.
- Novel situations: First-time scenarios, unusual transactions, and edge cases without precedent still challenge AI systems.
- Accountability: AI can recommend, but you're signing the return. Professional responsibility—and liability—stays with you.
The bottom line
AI capabilities are expanding rapidly with agentic workflows. The routine tasks that consumed your time last year might be fully automatable this year. The firms that thrive will be those that continuously evaluate what AI can handle—while focusing their human expertise on relationships, strategy, and judgment calls that truly require a CPA.