Use case
Agentic processes at work: the tax firm example
Tax firms are liable for every statement they make, while short of staff and under constant deadline pressure. An AI agent that quotes the law from memory is no help here; it is a liability risk. The five building blocks below show how an agent can work in a tax firm without diluting the responsibility of the licensed adviser.
Building block 1
RAG for the current law
The challenge
Whether an assessment holds depends on the current state of the law (statutes, administrative guidance, case law), not on a model's training memory.
The approach
A RAG pipeline checks the facts of the case against the stored state of the law and the firm's own work instructions. The agent argues from source and date, not from recollection. Every statement carries its evidence.
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 prepares (drafts, deadline overviews, follow-up questions to clients), but it never files anything. For example, standard cases with no room for judgement are prepared automatically, and anything involving discretion 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
The firm is liable. A recommendation whose origin cannot be reconstructed cannot be defended.
The approach
Every step is logged immutably: which state of the law applied, which source was read, which licensed adviser approved. The entries are hash-chained, so later changes are detectable. In a liability question, the path to the decision is fully traceable.
Building block 4
Quality against golden sets
The challenge
Any model update can quietly degrade the quality of its classifications. In a tax firm, that means wrong drafts in series.
The approach
Continuous evaluation against fixed golden sets: closed, anonymised cases with a known correct outcome. Every model change and every prompt change is measured objectively before it goes live.
Building block 5
The licensed adviser decides
The challenge
Specialists fear two things: being replaced by AI, and losing control over errors. In a tax firm, professional liability comes on top.
The approach
The system works on its own but stops at precisely defined gates: tax planning questions, unusual facts, a strong deviation from the previous year. There the licensed adviser decides. Reserved professional tasks stay where they belong. The person corrects, and the system learns from it in a structured way. Assistance instead of loss of control.
Transferable
The tax firm is the example here. The building blocks are the same as in accounting and in any other process where errors are expensive. They are code, not concepts:
- local-agent-pipeline ↗ (whitelist, gates, hash-chained audit trail)
- rag-approval-blueprint ↗ (approving a RAG system with a control catalogue)