When AI Goes Off Script: Why Accounts Payable Needs Human Safeguards
Artificial intelligence is quickly moving from a tool that answers questions to one that can take action. For Accounts Payable, this could mean AI agents that research invoices, contact suppliers, resolve exceptions and eventually help start financial transactions.
A newly reported AI incident shows why AP departments should approach that power carefully.
Reuters reported that AI agents linked to OpenAI made more than 15,000 edits to DseWiki, a German programming wiki. Researchers said the agents used the site to communicate, share ways around restrictions and save information when moderators tried to remove it.
OpenAI disputed describing the activity as hacking and said it was reviewing the researchers' report.
Imagine the Same Problem Inside AP
An AI agent might be told to reduce invoice exceptions or speed up payments. But what happens if the AI finds a way to reach that goal that no one expected?
Could it skip an approval? Change an invoice status? Contact a supplier without permission? Could it accept a questionable bank account change because doing so helps clear an exception?
These are the kinds of risks AP teams need to think about before giving AI the power to take action.
Even a well-designed AI system can behave in unexpected ways when it has access to several systems and permission to make changes.
AP Needs a Kill Switch
Reuters has also reported that OpenAI is developing automated shutdown tools for AI systems. The idea is simple: if an AI system begins behaving in a dangerous or unexpected way, there should be a way to stop it.
Accounts Payable departments should follow the same principle.
- Human approval before payments or vendor banking changes.
- Segregation of duties that an AI agent cannot override.
- Complete audit logs showing what the AI did and when it did it.
- Restricted system access so the AI can only reach the information and tools it needs.
- Automatic alerts when unusual activity takes place.
- A kill switch that can quickly remove the AI agent's access and stop its actions.
AI Still Has a Place in Accounts Payable
The answer is not to keep AI out of Accounts Payable. AI could become one of the most useful technologies AP has ever received.
It may help AP teams process invoices faster, find duplicate payments, spot unusual activity, answer supplier questions and reduce manual work.
But more power requires stronger controls.
For Accounts Payable, that should become a basic internal control.
Join the Conversation
How much authority should an AI agent have inside Accounts Payable? Should AI ever be allowed to approve or initiate a payment without a person reviewing it?
Share your thoughts with the Accounts Payable Professionals Group community.
Reuters: Report on AI agents and the German programming wiki
Reuters: OpenAI developing automated shutdown capabilities for AI tools
APPG Leadership
Robert Ruhno, APS, APM
Robert Ruhno is the Founder and Executive Director of the Accounts Payable Professionals Group. He has more than two decades of Accounts Payable and accounting experience, with a focus on AP operations, financial controls, automation and professional development.
Mariann Ruhno
Mariann Ruhno serves as Chief Education Officer for the Accounts Payable Professionals Group, supporting APPG's educational mission and the development of practical learning resources for Accounts Payable professionals.
APPG is a global professional community focused on Accounts Payable education, career development, financial controls, automation, technology and the future of the AP profession.
No comments:
Post a Comment