Who answers when the code fails? A consultant's perspective in the age of 100% autonomous accounting.
From my position as a builder of my own tools and a technology consultant, my job is to help companies implement innovation, streamline workflows and adopt automation.
Every day I see the tremendous potential of models like Claude or ChatGPT: they process data at lightning speed, extract information from invoices in milliseconds and automate tasks that used to take days.
And the question my clients ask me most often is: why don't we switch on full autopilot? Why don't we let Artificial Intelligence handle the accounting and taxes from start to finish, with no human intervention?
As a consultant, my responsibility is not only to sell the technology, but to analyze the strategic, contractual and operational risks of the solutions we recommend.
When we read the fine print of the law, insurance policies and regulatory compliance in the Netherlands and the European Union, the answer is clear: blind autonomy in financial and tax processes is a massive risk trap.
This is my technical and strategic reflection on the questions any consultant or project lead should ask before automating the accounting function 100%.
1. The illusion of delegation: AI has no legal personality
When a client suggests replacing human oversight with an autonomous AI agent, I ask a risk-management question: if the algorithm makes a systematic error classifying VAT for a whole year, who does the tax authority notify?
The operational reality is unforgiving: tax authorities (like the Belastingdienst in the Netherlands) don't talk to software.
Artificial Intelligence has no legal personality; its assets can't be seized, it can't be fined and it can't be held liable for tax. In the eyes of the law, the taxpayer (the company) and its legal representatives are the only ones responsible.
There is no "the tool hallucinated" defense.
The company will pay the debt, the interest and the fines out of its own pocket before it can bring any claim against the software vendor.
2. The gap in insurance policies: the risk consulting can't ignore
One of the most common mistakes when designing a process architecture is assuming that the client's professional liability insurance (Beroepsaansprakelijkheidsverzekering, or BAV) will cover any technology failure.
If, as consultants, we design and implement a process that removes human validation entirely (human-out-of-the-loop), insurers may classify the incident as a breach of the duty of care (zorgplicht) or gross negligence.
If the insurer rejects the claim, the financial impact falls directly on the company's assets, which also exposes the consultant to claims for poor process design.
3. The GDPR dilemma: confusing office tools with data pipelines
In IT consulting, I often hear this objection: "If the company already uses Microsoft 365 or Google Workspace in the cloud, what's the difference in sending invoices to Claude's API?"
The difference is one of data architecture and governance:
- Sensitive data in invoices: an invoice contains information protected by the AVG/GDPR (data of self-employed ZZP'ers, tax addresses, personal IBANs, confidential service descriptions).
- Unregulated environments: sending this data through public interfaces or APIs without Enterprise commercial agreements or a formal Data Processing Agreement (DPA) constitutes a data breach (datalek).
- Uninsurable fines: in Europe, penalties imposed by data protection authorities (such as the Autoriteit Persoonsgegevens) cannot be insured by law. No policy will compensate a company for a privacy fine caused by a poorly structured workflow.
My recommendation as a consultant: "augmented autonomy," not blind autonomy
My role is not to slow down the adoption of Artificial Intelligence, but to make sure my clients' digital transformation is sustainable and legally sound. To implement AI in financial processes safely, I recommend three design pillars:
- Deployment on regulated infrastructure (EU Enterprise): use language models through managed European regions (such as Google Cloud Vertex AI or Amazon Bedrock in the EU) with DPA contracts guaranteeing that data isn't transferred outside the European Economic Area or used to retrain models.
- Human-in-the-loop architecture: design the system so that AI handles extraction, reconciliation and anomaly detection, while keeping checkpoints where a human professional reviews and authorizes critical operations or validates random samples.
- Governance and traceability (logs): configure auditable records of every decision the model makes, to support any tax inspection or compliance audit.
Artificial Intelligence is the most powerful copilot we can integrate into a company, but responsibility for the flight and control of the controls remains a strategic human commitment that no code can replace.
← All articles