Incident sheet · /security
- System
- Security & Risk
- Filed
- Read
- 4 min · 862 words
- Severity
- Routine
AI Agents Now Move Money: Who Authorized the Transaction?
Visa, Bill and Billtrust now let authenticated AI agents execute invoice payments. Discovery at one firm found 48,000 agent versions where 7,000 were assumed — and logs point to tokens, not people.

Event log
- Visa, Bill and Billtrust launched a live pilot in October for AI-agent-executed business-invoice payments.
- Saviynt discovery found roughly 48,000 distinct agent versions at a firm that estimated it had 7,000 agents.
- Agent activity trails often end at service tokens rather than accountable employees, per Saviynt.
- Google's agentic AI guidelines require explicit human approval for financial transactions.
- Authorization records must show why a transaction was allowed, not just that it occurred.
By Daniel Okafor
Authenticated AI agents began executing business-invoice payments across buyer and supplier systems in a live pilot launched in October by Visa, the payments company Bill, and Billtrust. The agents do not decide which invoices get paid. They execute payments that businesses have already approved. That distinction is doing a lot of work. It is also the entire governance problem.
The pilot forces a question CIOs cannot defer: when an agent creates a financial obligation, how does the enterprise establish, enforce, and prove the agent was authorized to act? The answer requires more than application access controls. IT, finance, and legal must agree on what agents may do, when a human must approve, and who answers for an agent that exceeds its authority.
Access does not equal authority
An agent's identity, the systems it can reach, and its authority to perform a specific action are three separate things. Vibhuti Sinha, chief product officer at identity security firm Saviynt, makes the case bluntly. An agent might hold access to a payment system that permits approving a $2 million invoice. Whether it should depends on who delegated the task and the business context at the moment of the transaction.
The problem compounds across applications. "Access builds up with every hop, and nobody reviews the total," Sinha said. "By the third application, the logs show a service token instead of the person who actually asked for the work."
One data point should unsettle every security team. A global analytics firm estimated it ran roughly 7,000 AI agents. Saviynt's discovery across the environment counted closer to 48,000 distinct agent versions — nearly seven times the assumed figure. Copies, redeployments, and versions with modified prompts or tools had acquired their own credentials, many accumulating cross-application access that nobody had reviewed collectively. Worse, the trail of agent activity frequently ended at a service token, not an accountable employee.
"Agent access rarely gets out of hand in one dramatic step. It drifts, one hop and one copy at a time, until no one can say who's accountable," Sinha said.
What enforcement looks like
Permissions alone will not stop an unauthorized transaction. Sinha's position: enforce authorization at the moment the agent attempts the transaction, not at login. An authorization gateway between agent and application can evaluate the requested action against business rules — spending limits, approval requirements, recipients. Routine requests proceed. Some get blocked outright. Others pause the agent and notify the responsible person via Slack, Teams, or another channel for approval or denial.
Google's agentic AI guidance aligns with this, recommending explicit human approval for consequential actions, including financial transactions, before agents execute them. Google also states that customers bear responsibility for granting agents appropriate permissions. Note what is absent: no standard, no audit framework, no liability allocation. Guidelines only.
Five questions before an agent moves money
CIOs should demand answers to five questions before authorizing transactional agents:
- Who delegated the task? Identify the person or business function accountable for the agent's work.
- What can the agent do? Define permitted transactions, systems, recipients, and spending limits.
- When must a person approve? Set thresholds and exceptions requiring human review before execution.
- What evidence is retained? Connect agent identity, transaction, governing policy, and approvals across systems.
- Who handles a mistake? Assign responsibility for investigation, reversal, and escalation before an incident occurs.
The critical distinction: a transaction record shows what happened. An authorization record explains why it was allowed.
Who pays when an agent goes rogue?
Technical controls cover half the problem. Veronica Glick, a partner in Mayer Brown's cybersecurity and data privacy practice, frames the legal exposure in one line: "Generative AI says things; agentic AI does things. It can send messages, change records or move money."
Contracts can specify permitted actions, reachable systems, and approval triggers. But when an agent exceeds its lane, the cause may be the customer's instructions, the underlying model, or a third-party tool — making customer misuse hard to separate from vendor failure.
"A workable approach aligns responsibility with control: the vendor stands behind what it controls, such as its model, its guardrails and its tools, and the customer owns its own instructions, configurations and approvals," Glick said.
Enterprises should scrutinize liability caps, indemnities, audit rights, incident definitions, and notification requirements in vendor agreements. Glick's caveat deserves weight: even large enterprises hold limited leverage with major technology providers. Prioritize higher-risk agent uses. Revisit protections at renewal, not after a breach.
Proving authorization after the fact
"The problem is that logs tell you what happened, and an auditor wants to know why it was allowed," Sinha said. "Five disconnected logs aren't evidence."
Companies must trace each agent action back to the delegating person — documenting the access rights used, the policy permitting the action, and every approval along the chain. Glick adds that approval requirements should scale with the consequences of the action and whether it can be reversed, and that activity records, including third-party records where feasible, must arrive fast enough to investigate.
A payment record proves the transaction cleared. It does not prove anyone authorized the agent to make it.
via techtarget.com (Source)
More from Daniel Okafor
Show full bio
Staff writer covering technology and product news in enterprise information technology and the leaders who run at CIO Monitor.
203 articles
The newsroom digest
One email a day with the stories that matter. No noise.
Cross-references
- Citi shifts AI-agent hack losses to customers under new terms
- Microsoft, Google Move AI Agent Governance Into Enterprise IT Mainstream
- UK Regulators: Enterprises Own the Legal Liability for Their AI Agents
- Gartner: 70% of Firms Running Agentic AI Face Material Incidents by 2029
- AI-Enabled Sanction Evasion Becomes New IT Governance Headache