Under SOX Section 404, management assesses internal control over financial reporting, and for accelerated filers the external auditor attests to it. The SEC’s interpretive guidance names four IT general control areas that automated controls depend on: program development, program changes, computer operations, and access to programs and data. An AI agent that reads or writes NetSuite, SAP or Salesforce is a new actor inside the access area.
That is where projects stall. A finance team wants Claude or ChatGPT to pull aging reports and draft reconciliations, and the first question from IT audit is simple: what account does the agent use, what can that account do, and who reviews it? KPMG’s 2026 paper on agentic AI in financial reporting puts it plainly, asking whether agents are assigned specific user IDs with roles and permissions or run with a shared service account, and warning that shared service accounts weaken the audit trail.
Why the usual shortcuts fail an access review
- A personal API token pasted into an agent carries everything that person can do, including posting.
- One integration user shared by every script and agent shows up in the ERP log as one name, so nobody can tie an action to a person.
- Long-lived keys sit in prompts, notebooks and config files, outside any access review.
- New agents appear without an owner or an inventory record, which ISACA flags as a core agentic AI audit risk.
What least privilege looks like for an agent
Treat the agent like any privileged system actor: give it a defined set of actions, on an account whose ERP role matches the job, with every action traceable to a person who can be asked about it. engini is the layer where that happens. Your ERP keeps its own roles and logs; engini decides which of its tools an agent can call and on which connection, and keeps the record.
NetSuite
SAP S/4HANA
Salesforce