Most of what is sold as AI for the enterprise stops at a chat window. It can read your data and write a nice paragraph about it. It cannot release the purchase order, post the document or update the opportunity, so the work still has to be done by a person in the system of record.
We set a different bar. Before a Fidevra agent is allowed to touch SAP, IFS, Salesforce, HubSpot, Acumatica or QuickBooks, it has to pass five tests.
1. It holds its own credentials
Each agent gets its own scoped service account in each system, with the minimum roles the job needs and nothing else. If an agent misbehaves, you revoke that account and the rest keep running. There is no shared super-user behind the scenes.
2. It uses the system's own doors
Agents call the platform's supported APIs: OData and BAPIs in SAP, projections in IFS Cloud, REST in Salesforce, HubSpot and Acumatica. They go through the same validations, workflows and authorisations a person would. Nothing is written straight to a table, and nothing bypasses the approval flow your auditors already signed off.
3. It knows when to ask
Every job has a brief that says what the agent may do alone and what needs a person. Amount thresholds, role checks and four-eyes rules are enforced by the platform, not by each prompt, so a clever phrasing cannot talk an agent past them. When a rule says ask, the agent packages the evidence and asks the right person where they already are: email, Teams or Slack.
4. It leaves a trail an auditor can read
Every read, write, decision and the evidence behind it is recorded and queryable in plain language. "Show me every purchase order this agent released over fifty thousand, and why" is a one-line question, not a project.
5. It can be switched off
Pause one agent or every agent in a click. Per-agent limits on actions per hour and amounts per day mean a bad brief becomes a bad hour, not a bad quarter.
What it looks like in practice
Take a late purchase order in IFS Cloud. The agent reads the supplier's email saying the delivery slips nine days, matches it to the open PO, checks which customer orders depend on that line, and finds that one promise date is now at risk. Within its limits it confirms the supplier's new date on the PO. Because a customer promise is affected, it does not change the sales order itself; it drafts the re-date, costs the expedite alternative, and sends both to the planner for a decision. The planner approves in Teams. The agent commits the change in IFS, sends the customer an approved update, and logs the whole chain.
That is the difference between an assistant and an agent. One tells you about the problem. The other brings you the decision, already prepared, and handles everything around it.