WW/PROCUREMEN
Procurement AI Needs a Credit Limit, Not Just Human Approval
A $9,000 sole-source part can out-risk a bigger routine buy. Researchers propose bank-style signature authority for procurement AI agents, with four-factor scoring caps and earned expansion.
- By
- Elena Vasquez
- Filed
- Length
- 700 words
- Read
- 4 min

Key points05
- 89% of small banks allow first-level sign-off on $25,000 loans vs 43% at $250,000 (FDIC 2024)
- Walmart's Pactum pilot reached 64% of targeted suppliers against a 20% goal
- 53% of 261 procurement leaders report risk-based autonomy limits by category (Forrester/Zycus)
- Framework scores agent actions 1-4 on value, criticality, reversibility and policy sensitivity
- Daily $5,000 orders aggregate to ~$110,000 over 22 business days, raising the authority band
Banking regulators' data shows why: 89% of small banks let a first-level decision-maker sign a $25,000 loan, but only 43% allow it at $250,000, according to the FDIC's 2024 Small Business Lending Survey. Two supply chain researchers argue procurement organizations should apply the same signature-authority logic to AI buying agents before those agents commit company money.
Dr. Irita Mishra and Dr. Omar Farooq of the University of Wisconsin, River Falls, propose a scoring system in Supply Chain Management Review that caps an agent's autonomy based on its riskiest attribute — not its spend alone. Their core warning: a $9,000 order for a sole-source spare part can carry more operational risk than a larger routine purchase, because spend is not equivalent to exposure.
AI agents already negotiate renewals, trigger replenishment and, at some firms, release payment. The standard safeguard — a human in the loop — breaks down at scale. When reviewers face hundreds of AI-generated orders a day, approval becomes routine confirmation rather than substantive oversight, the authors write.
What does the framework actually change?
Each AI action receives a score from 1 to 4 on four dimensions: value (financial exposure), criticality (operational consequence), reversibility (cost to undo) and policy sensitivity (contractual and regulatory exposure). The highest-risk factor sets the authority ceiling, mirroring binding credit limits.
Ratings come from systems of record that people own — supplier master data, the contract repository, category managers' criticality ratings and the approved-supplier list. The agent reads the data but never scores itself.
Two rules close gaming loopholes. Commitments to a single supplier aggregate over a rolling 30-day window, so splitting orders cannot lower the band. Daily $5,000 replenishment orders fall individually into Band 1, but aggregate to roughly $110,000 across 22 business days, placing them in Band 2. Orders with unapproved suppliers or terms outside the pre-approved envelope are barred from autonomous execution entirely.
Automatic escalation triggers include:
- A counteroffer pushing any factor into a higher band
- A term leaving its approved range
- An aggregate cap being reached
- Missing data
- An anomalous transaction pattern
Escalated decisions route to a designated owner under the company's delegation-of-authority structure — buyer, category manager, CPO or compliance owner. In the worked example, the $9,000 single-source part scores Band 1 on value but lands at "Recommend only" because of supply concentration. The cost is minimal: the agent prepares the order and a buyer approves it in minutes. With no buffer stock on hand, the line-down risk would justify human leadership.
How much autonomy exists today?
Bounded autonomy is already in the field. Walmart's pilot with Pactum, documented in a Thunderbird School of Global Management case, let buyers set goals, trade-offs and deal-breakers before the bot negotiated with tail suppliers for non-resale goods such as store equipment. The pilot reached 64% of targeted suppliers against a 20% goal. In a Forrester survey of 261 procurement leaders commissioned by Zycus, 53% cite risk-based autonomy limits by category — though the authors note that category alone is an incomplete proxy for exposure.
Banking supervisors offer a further precedent. The interagency model-risk guidance SR 26-2, which replaced SR 11-7, expressly excludes agentic AI, but its principle holds: oversight should tighten as materiality rises.
How should limits evolve?
Authority should be earned, not granted. The authors recommend tracking overrides, step-ups, blocked attempts and supplier disputes as evidence of calibration. Tighten limits when losses or disputes trace to an agent or overrides exceed tolerance. Expand only after a defined incident-free period and a clean audit sample, followed by approval from the designated authority — and monitor low-band categories as aggregate exposure grows.
For implementation, the authors suggest starting with a low-band category such as tail spend: define approved suppliers, term ranges and dollar caps, log all agent actions, and assess performance after one quarter. Their bands are illustrative and untested; each organization must calibrate against its own delegation-of-authority schedules and risk appetite. As agents take on more execution, the trajectory points toward earned, evidence-based expansion of autonomy rather than maximized independence from day one.
Source: scmr.com
More from Elena Vasquez
Show full bio
News editor covering people, launches and day-to-day practice in Ocean and air freight, container shipping at Waybill Wire.
996 articles
Related05
Five barriers stand between AI pilots and trusted supply chain planning
U.S. Steel offers 19.25% raises, $4.5B investment in tentative union deal
Vietnam Cuts SME Corporate Tax and Land Costs From July 2027
German firms now run 620 projects worth $3.7bn in Vietnam
Kenya names Frontier consortium top bidder for $62B Mrima Hill