Use case
Agentic processes at work: the accounting example
An AI agent working in a business-critical process makes mistakes that cost money straight away. Accounting is the hardest example: strict rules, audit obligations, no tolerance for invented figures. Conventional AI only reads documents here. The five building blocks below show how a tool that reads out documents becomes a system you can trust with work. Each one runs in production for me or sits in a public repository on GitHub.
Building block 1
RAG for policies
The challenge
A document is only as good as its context. Whether a booking is correct depends on internal policies, contracts and tax rules, not on the document itself.
The approach
A RAG pipeline reads the document and, in the same step, checks it against internal company policies, contracts and tax rules. The agent argues from the sources, not from memory.
Building block 2
Action-space whitelisting
The challenge
An agent that may do anything is a risk. An agent that may do nothing is useless.
The approach
The agent may propose journal entries and trigger approval workflows, but only within clearly predefined limits. For example, documents up to €500 are booked automatically and anything above goes to the human-in-the-loop gate. Any attempt outside the whitelist is discarded and logged.
Building block 3
Hash-chained audit trails
The challenge
In finance, no auditor accepts an AI decision that cannot be traced.
The approach
Every step is logged immutably: which document was read, which prompt ran, which person approved. The entries are hash-chained, so later changes are detectable. The result is fully traceable and documented for audit.
Building block 4
Quality against golden sets
The challenge
In financial reports, one invented figure leads straight to wrong decisions. And any model update can quietly degrade quality.
The approach
Continuous evaluation against fixed golden sets: real, correctly coded, anonymised data. Every model change and every prompt change is measured objectively before it goes live. Coding accuracy never drops unnoticed.
Building block 5
Human control in the right place
The challenge
Specialists fear two things: being replaced by AI, and losing control over errors.
The approach
The system works on its own but stops at precisely defined gates: a discrepancy in the cash discount, an unclear cost centre, a new supplier. There it asks a person to approve. The person corrects, and the system learns from it in a structured way. Assistance instead of loss of control: autonomy grows with proven quality, not the other way round.
Transferable
Accounting is the example here. The pattern carries over to any process where errors cost money: purchasing, contract review, maintenance, customer service. The building blocks are code, not concepts:
- local-agent-pipeline ↗ (whitelist, gates, hash-chained audit trail)
- rag-approval-blueprint ↗ (approving a RAG system with a control catalogue)