- Agents must act through the ERP’s rules, never around them.
- Drafts before posting, approvals as today, and an audit record for every agent action.
- Raise autonomy per process, only when the measured error rate allows it.
The fastest way to lose a finance team’s trust in AI is to let it post a wrong journal. The fastest way to get no value from AI is to forbid it from touching anything. Human in the loop is the middle path, and it is a design discipline more than a slogan.
Principle 1: agents see what the user may see
An agent acting for a user must be checked against that user’s permissions, not given its own wide access. Otherwise the agent becomes a way around the access rules your auditors rely on.
Principle 2: act through the application
Your ERP’s ledger is trustworthy because of its rules: balanced journals, document states, closed periods, approval workflows. An agent that writes directly to database tables skips all of them. An agent that uses the same services as a person inherits every validation and every approval.
Principle 3: drafts before posting
Agents create documents in draft. Completing and posting remain separate, authorised steps. This single rule removes most of the risk, because a person sees every transaction before it affects the books.
Principle 4: an audit trail for every action
For each action the system should record which agent acted, for which user, using which data, with what result, and who approved it. Auditors should be able to reconstruct any agent decision as easily as a human one.
Principle 5: grade autonomy per process
Autonomy is not on or off. A useful scale runs from answering questions, to advising, to preparing work for approval, to acting within policy limits. Different processes sit at different levels: an agent might match routine bank lines on its own while large purchase orders always wait for a person. Our research article on the science of agentic ERP sets out the full scale and the human factors research behind it.
How to decide when to raise autonomy
Measure the share of agent proposals accepted without change, and the cost of the errors that were caught. When acceptance is consistently high and the errors caught are small, the process is ready for the next level. When it is not, the measurement tells you what to fix.
See how eVitalyst applies these principles in architecture and governance.
Want the full picture? Download the agentic ERP white paper for leadership teamsor see eVitalyst AI on your own data.