The Three-Layer Architecture
1 AI Classification
Our agents use large language models to read and classify financial data — invoices, receipts, bank transactions, expense reports. The AI understands context: it knows the difference between a business lunch and a personal one, between a software subscription and a one-off purchase, between standard-rated and zero-rated supplies.
This is where AI excels. Classification is inherently fuzzy — it requires understanding context, reading unstructured documents, and making judgment calls. LLMs are built for this.
2 Deterministic Computation
Once transactions are classified, the tax math is handled by Accora's deterministic engines. No LLM touches the computation. VAT, income tax, social security contributions, payroll — all calculated by rules engines built on codified law.
Why? Because tax computation isn't a prediction problem. The VAT rate is 20%, not "probably around 20%." The tax bracket threshold is €35,300, not "approximately €35,000." Deterministic engines produce the same result every time for the same inputs. They can be audited, tested, and verified. AI predictions can't.
Accora maintains these engines across 30+ jurisdictions, updating them when legislation changes. Our agents call Accora's API to compute — we don't build or maintain tax logic ourselves.
3 Accountant Review
Every output is reviewed by a qualified, warranted accountant before it's finalised. They check classifications, verify computations, and sign off on filings. They carry professional liability.
This is the part that makes the system trustworthy. AI classifies. Engines compute. Humans verify. The accountant doesn't do the work — they quality-assure it. Over time, the review rate drops as the AI learns the client's patterns, but it never reaches zero. There are always edge cases that require professional judgment.
Why This Matters for Your Firm
Most "AI accounting" products use AI for everything — including the math. That means your clients' tax returns are being produced by a system that predicts what the right answer looks like, rather than calculating it.
Our agents are different. The AI does what AI is good at (reading, classifying, understanding context). The math is done by engines that are right every time. And a qualified person checks the output before it goes anywhere.
That's why our clients see 100% VAT accuracy, not 95% or 98%. The architecture doesn't allow for "close enough."
Built on Accora
Accora is the accounting infrastructure layer that powers our agents. It provides:
- Deterministic tax engines across 30+ jurisdictions
- API and MCP (Model Context Protocol) access for agent integration
- A network of warranted accountants for review and filing
- Jurisdiction-specific classification rules updated when laws change
Learn more at accora.ai. Or read how we built it for the full backstory.