Who is responsible when AI makes mistakes in your ERP?

Accountability for AI in your ERP always lies with people, never with the agent: functionally with a named process owner who spot-checks results and decides edge cases, technically with a role that maintains permissions and approval levels, and entrepreneurially with the management that sets the frame of use. Legally, an AI agent is a tool. Whoever lets it post without named accountability has not introduced automation but a responsibility gap.

Why the question must be answered before the first agent

As long as Copilot suggests text, the question is academic: a human accepts the suggestion or not. It becomes real the moment an agent acts on its own, for instance creating orders from emails or pre-coding incoming invoices. From that point on, documents are created, and every document needs someone who stands behind it: towards the auditor, the tax office and your own customer.

The good news: Business Central is built for this. Agents work under their own user context with their own permissions, their actions are distinguishable from human ones in the log, and approval workflows apply to agents just as they do to colleagues. The tools are there. What companies must bring themselves is the decision about who operates them.

The three roles that suffice

1. The process owner: spot-checks the agent results of their process, decides edge cases and reports patterns. This is not a new position but a named responsibility of today's process owner.
2. The permissions owner: maintains what each agent may see and do, as narrowly as possible. An order agent does not need general ledger access. This role usually sits with the person who manages user rights today.
3. The management: sets the frame: which processes may be automated, up to which value limits an agent works without human approval and when to escalate. One page of paper, reviewed once a year.

The midmarket needs no more, and every additional committee slows down exactly the decisions the automation was built for. What matters is that all three roles are filled before the first productive agent, not after the first incident.

Approval limits: the practical core

The most effective steering instrument is unspectacular: value limits. The agent creates orders up to a defined amount on its own; above that, the transaction moves to human approval. The same logic applies to new business partners, deviating prices or incomplete documents: everything unusual goes to a human, everything uniform flows through.

This preserves the relief without losing control. And the limits are not dogma: after a quarter of clean spot checks, the value limit may rise. Accountability does not mean permanent mistrust, it means trust with evidence.

What accountability has to do with the data foundation

Most agent errors are not AI errors but data errors at a new pace: the agent posts to the duplicated customer or picks the outdated item because both exist in the system. Without a clean data foundation an agent multiplies the chaos, and no approval limit catches that permanently.

That is why the accountability order includes an owner per data object: who in the company is responsible for ERP data qualityindirectly also decides how safely every agent works. This is exactly the order we help build, senior-led: governance and delivery management, with steering on your side.