What should it be allowed to approve without you? AI can now cross the line from preparing a payment to actually completing it. That makes your approval policy more important than the technology.
The newest payment agent does not decide whether the invoice deserves approval
Visa, BILL and Billtrust announced their first live B2B agentic-payment pilot in October 2026.
The workflow allows an authorized BILL Payment Agent to enter a supplier environment, use tokenized payment credentials and complete an approved payment without an employee manually logging into the supplier’s portal.
That distinction matters because “AI pays the invoice” can describe two very different systems.
| Automation level | What AI does | Who decides to spend money? |
|---|---|---|
| Level 1 | Reads invoice and extracts data | Human |
| Level 2 | Matches invoice to PO / receipt and proposes approval | Human |
| Level 3 | Approves low-risk invoices inside pre-set rules | AI within delegated authority |
| Level 4 | Schedules or executes an already approved payment | Previously approved human or policy decision |
| Level 5 | Selects vendor, approves exception and moves money independently | AI |
Let AI approve the boring invoice—not the unusual payment
The safest autonomous invoice is one where the system already knows nearly everything about the transaction.
LANE 1 — AI may approve automatically
LOW RISKConsider autonomous approval only when all required controls are satisfied.
- Vendor is already approved and established.
- No vendor ownership, address or banking information changed.
- Invoice matches an approved purchase order.
- Goods or services were confirmed received.
- Quantity and price are within approved tolerances.
- Invoice is not a duplicate.
- Invoice amount is below the business’s autonomous approval limit.
- Payment destination matches previously authenticated vendor instructions.
- Payment method is already approved for that vendor.
- No fraud or anomaly alert is present.
LANE 2 — AI prepares; one human approves
NORMAL EXCEPTION- Recurring utility or subscription without a purchase order.
- Small price variance above automatic tolerance.
- Invoice is larger than normal but explainable.
- New contract period or renewal.
- First invoice under an already approved agreement.
- Invoice arrives unusually early or late.
LANE 3 — Human approval required
HIGHER RISK- New vendor.
- First payment to a vendor.
- Invoice without corresponding contract or PO where one is expected.
- Material price or quantity mismatch.
- Rush or same-day payment request.
- Unusual payment method.
- Invoice amount materially outside historical pattern.
- International or unfamiliar payment destination.
LANE 4 — Two humans / dual control
MONEY-MOVEMENT RISK- Vendor bank-account change.
- Routing-number or payment-instruction change.
- Large wire transfer.
- New bank account or wallet destination.
- Owner, executive or related-party payment.
- Unusual payment above the company’s high-value threshold.
- Override of a fraud alert.
- Change to the agent’s own approval limit or payment permissions.
Never let the same agent change the vendor’s bank account and approve the payment
This is the agentic equivalent of handing one employee the vendor master file, approval stamp and bank token.
BILL’s own control architecture offers dual control for sensitive actions including adding or changing vendor bank accounts. Under dual control, one person initiates the action and a second must approve it.
The separation should survive automation. If one AI workflow proposes the vendor change, another independent control—or preferably a human using independently verified contact information—should authenticate it before money is released.
An AI agent can execute a fraudulent instruction perfectly
Many payment losses do not happen because an attacker breaks into the bank.
They happen because somebody inside the company is tricked into authorizing a legitimate payment to the wrong place.
Nacha calls this credit-push fraud authorized under false pretenses. Examples include:
- Vendor impersonation
- Fake bank-detail change
- Executive impersonation
- Payroll diversion
- Urgent wire request
- Compromised vendor mailbox
- Fraudulent invoice
- Payment redirection
Nacha’s 2026 fraud-monitoring rules now require non-consumer ACH originators to establish risk-based processes intended to identify entries that may be unauthorized or authorized under false pretenses.
Nacha specifically points businesses toward controls such as anomaly detection, behavioral tolerances, pattern recognition and velocity checks.
Do not give the robot Susan’s banking password
The Visa/BILL/Billtrust pilot is useful because it demonstrates what a better architecture can look like.
| Weak design | Better design |
|---|---|
| Agent shares an employee’s supplier login | Agent has a recognized non-human identity |
| Permanent card or bank credentials stored in workflow | Tokenized or limited-purpose payment credentials |
| Agent inherits employee’s broad permissions | Agent receives explicit payment authority |
| No way to distinguish human from agent actions | Agent activity is separately logged |
| Generic access to supplier portal | Payment path tied to the specific authorized invoice |
| Agent can change its own rules | Approval and permission changes require independent authorization |
Visa says its pilot uses a recognized agent identity, tokenized credentials, payment controls tied to the approved instruction, a payment path linked to the specific invoice and visibility into authorization and completion.
NIST is separately working on identity and authorization standards for software and AI agents, emphasizing identification, delegated authority, logging and the ability to distinguish automated identities from human users.
Before an AI agent receives payment authority, require these controls
Dedicated agent identity
The payment system can tell whether the action came from the agent or a person.
Maximum transaction limit
The agent cannot approve or execute invoices above its delegated dollar ceiling.
Daily aggregate limit
Ten $4,000 payments should not quietly bypass a $25,000 high-value review policy.
Approved-vendor list
Autonomous approval is restricted to vendors that have completed onboarding and payment-detail verification.
PO / receipt match
Where purchasing uses purchase orders, the invoice must agree with what was ordered and received.
Duplicate detection
Invoice number, vendor, amount, date and payment history are checked before approval.
Bank-change lockout
Any change in payment instructions automatically removes the invoice from autonomous processing.
Independent vendor verification
Payment-detail changes are verified through a trusted channel rather than contact information supplied in the change request.
Behavioral anomaly check
The system compares amount, timing, destination and method with normal vendor activity.
Complete audit trail
The business can reconstruct what the agent saw, what rules passed and who approved every exception.
Dual control for sensitive changes
Vendor banking, users, permissions and agent authority cannot be changed by one actor alone.
Kill switch
Someone can immediately suspend the agent’s payment capability without shutting down the entire accounting operation.
Do not copy another company’s $5,000 threshold
There is no universal SBA, IRS, Visa, BILL or Nacha rule saying an AI agent may safely approve invoices below a particular dollar amount.
A $5,000 invoice can be trivial for one company and enormous for another.
A better starter limit is tied to the company’s normal AP activity.
Consider
- Typical invoice amount
- Largest routine vendor invoice
- Monthly AP volume
- Cash balance and liquidity
- Industry fraud exposure
- Whether PO matching exists
- Quality of vendor-master controls
Then layer controls
- Per-invoice ceiling
- Per-vendor daily ceiling
- Total daily agent ceiling
- Monthly agent ceiling
- Exception threshold
- Dual-approval threshold
- Emergency suspension rule
The number is less important than the conditions around it
| Invoice situation | AI role | Human requirement |
|---|---|---|
| Known vendor + unchanged bank + exact PO/receipt match + below autonomous cap | May auto-approve | None if company policy permits |
| Known recurring vendor + no PO + amount inside normal range | Prepare approval | One human |
| Known vendor + price/quantity exception | Flag exception | One human / department owner |
| New vendor | No autonomous approval | Vendor onboarding + human approval |
| Vendor banking changed | Stop | Independent verification + dual control |
| High-value invoice above policy threshold | Prepare only | Two approvers |
| Rush wire or unusual destination | Stop | Independent verification + senior approval |
| Related-party / owner payment | Prepare only | Human approval under company governance policy |
This matrix is an INCBOOK control framework, not an official regulatory approval schedule. Businesses should tailor thresholds to their size, banking agreements, accounting systems and internal-control requirements.
This one workflow deserves a wall around it
Nacha specifically recommends validating changes to payment instructions through a trusted channel independent from the original request.
For example, if an email requests new ACH details, do not verify the change by replying to that same email or calling the new phone number contained in it.
Payment fraud is already widespread before AI gets the bank key
The Association for Financial Professionals’ 2026 Payments Fraud and Control Survey included 465 treasury practitioners across U.S. organizations.
| 2025 finding | Share |
|---|---|
| Organizations experiencing attempted or actual payments fraud | 76% |
| Organizations affected by business email compromise | 74% |
| Organizations reporting check fraud | 58% |
| Organizations reporting ACH debit fraud | 30% |
| Organizations reporting wire-transfer fraud | 25% |
| Organizations using AI to combat payments fraud | 17% |
The danger is not theoretical, and smaller companies can have a harder time recovering after a successful fraud event because they often lack the recovery infrastructure of larger organizations.
Businesses appear more comfortable letting AI watch money before letting it move money
PYMNTS Intelligence and Velera surveyed 2,474 U.S. small and medium-sized businesses about AI capabilities they would likely use from their financial institution during the following two years.
| Requested AI capability | All SMBs | $1M+ revenue SMBs |
|---|---|---|
| Track expenses and bills | 31% | 36% |
| Find suppliers / products | 22% | 25% |
| Cash-flow / credit guidance | 22% | 25% |
| Create a budget | 22% | 24% |
| Pay bills / invoices | 21% | 24% |
| Save payment credentials | 14% | 19% |
| Let AI make purchases | 11% | 14% |
Let the agent eliminate repetitive checking before eliminating human judgment
For many small businesses, the biggest efficiency gain may come before autonomous approval.
Excellent AI work
- Extract invoice fields.
- Find the purchase order.
- Match invoice to receipt.
- Detect duplicates.
- Check arithmetic.
- Compare vendor history.
- Flag unusual amounts.
- Route exception to correct person.
- Prepare the payment batch.
Keep human control longer
- Approve new bank instructions.
- Create and pay a new vendor in one workflow.
- Override fraud warnings.
- Approve unusual executive requests.
- Make large one-off payments.
- Change the agent’s own permissions.
- Turn off audit logs.
Write the rule before turning the feature on
New vendors, vendor bank-detail changes, unusual payment destinations, transactions above established limits, fraud alerts, material invoice exceptions and modifications to the agent’s own permissions require human review. Payment-instruction changes and other designated high-risk actions require independent verification and dual approval.
Where the controls come from
Important distinction: The Visa/BILL/Billtrust live pilot described in this report currently uses an AI agent to execute eligible buyer payments that have already been approved. Visa explicitly says the agent is not independently deciding which invoices deserve payment.
The autonomous-approval lanes and suggested control framework in this report are INCBOOK editorial planning recommendations, not Visa, BILL, Nacha, NIST or regulatory requirements. Actual payment authorities should reflect the business’s financial controls, banking agreements, transaction volume, accounting environment, legal obligations and risk tolerance.
Should this invoice be allowed through without you?
Enter the invoice facts. The tool assigns it to one of four approval lanes: autonomous, one-human review, human approval or dual control.
Enter the invoice facts
The suggested approval lane will appear here.
This tool is an internal-control planning illustration, not accounting, banking or legal advice. A company can reasonably choose stricter controls, including requiring human approval for every payment.
