Part 1 of 3 in our series on building the first accounting AI agents

Every week, I see a new “AI agent” launched. Demos promise 99% accuracy, instant reconciliation, or fully automated journal entries.

But here’s the truth: in accounting, none of these tools actually work the way finance leaders need them to. They look great in a demo, but when you put them into the messy reality of accounting workflows, they break.

Why? Because most of them lack context.

The Mirage of Accuracy

Accuracy is the favourite number in AI marketing. “Our agent matches 99% of transactions automatically.” Sounds impressive, right?

But if you’ve ever closed the books, you know that accounting doesn’t live in a clean dataset.

• A vendor changes their invoice format.
  
  

• A bank feed drops a transaction.
  
  

• A journal entry hits the wrong cost centre.

In those cases, “99% accuracy” means nothing. If the system doesn’t understand the why behind your data, that 1% of exceptions can eat up just as much time as doing the work manually.

And finance leaders know this. That’s why so many teams smile politely at AI sales pitches, then go back to Excel when close week comes.

Why Context Matters

Accounting isn’t just numbers. It’s relationships, timing, and judgment.

• An invoice might not match because of a foreign exchange difference.
  
  

• A journal entry might look unusual, but it’s tied to a one-time acquisition.
  
  

• A bank transaction might appear missing, but it cleared two days later.

Without understanding the broader context - history, related data, business rules - an agent is just guessing.

And when accountants can’t trust the system’s guesses, they’ll never adopt it.

The Limitations of Legacy Approaches

Most existing “AI agent” products stop at the surface. They rely on:

• Rules engines: If X, then Y. Works fine until a new scenario appears.
  
  

• No-code builders: Flexible, but only as good as the workflows you hardcode.
  
  

• Isolated models: Great at pattern-matching single data points, blind to everything else.

These tools can automate the easy stuff. But the moment they hit exceptions - mismatched invoices, missing bank entries, timing differences - they break.

And when that happens, the “agent” stops being an agent. It just becomes another inbox of errors for accountants to clean up.

What a Context-Aware Agent Looks Like

A truly useful accounting agent must be able to do three things with context:

1. See the bigger picture. Instead of just comparing two fields, it should pull in history, related records, and external sources to understand what’s really happening.
   
   

2. Handle exceptions intelligently. If something doesn’t match, the agent shouldn’t just fail. It should look for related data, suggest potential resolutions, or escalate in a structured way.
   
   

3. Learn over time. Every time an accountant resolves an exception, that judgment should feed back into the system, so the agent handles it better next time.

Imagine reconciliation where the agent doesn’t just spit out unmatched transactions, but actually explains: “This payment doesn’t match because it cleared two days later in the bank feed. I’ve linked it here.”

That’s the difference between a toy and a tool.

Context Builds Trust

At the end of the day, this is about trust. Accountants don’t care about hype. They care about whether the system makes their work easier, and whether they can rely on it when close week pressure is at its peak.

Agents that lack context can’t earn that trust. They’ll always be second-guessed.

Agents that understand context can become true partners, not replacing accountants, but freeing them up to focus on judgment, strategy, and insight.

The Path Forward

Building that kind of agent isn’t simple. It requires rethinking how data flows across systems, how workflows adapt to exceptions, and how human feedback is built into the loop.

At Stacks, that’s exactly what we’re working on. Our agents combine data, context, and workflows in a way that makes them resilient, adaptable, and trustworthy.

This isn’t about chasing 99%. It’s about creating systems that work in the messy, real world of accounting.

👉 Next in the series: In Part 2, we’ll explore Workflows, and why most agents break the moment they hit the complexities of real accounting processes.