Mastercard enables agent payments as CFOs flag cost risks.
Mastercard’s AI agents make payments.
Mastercard has unveiled a system to let AI agents conduct digital transactions with explicit human-set limits. At the Cloudflare Connect event on 20 October, the company demonstrated a platform called Agent Pay for Machines that enables one software agent to buy services from another on behalf of a business ([1]). The system uses an open protocol to exchange secure payment vouchers, which the agent’s human operator can cap in advance ([2]). Mastercard has already run end-to-end pilots with Cloudflare, Coinbase and several other partners and is now opening early access to developers ([3]).
By allowing AI agents to pay for data, software or cloud computing on their own, this capability addresses a key limitation in autonomous workflows. Until now, even the most advanced agents could get stuck if a task required a purchase order or credit card, forcing a human to step in. Mastercard’s solution suggests it will soon be possible for agents to initiate certain low-value transactions on the fly, for example, paying per-use fees for an API or online tool, without compromising a company’s spending controls. The platform automatically enforces spending limits and permissions on each agent ([4]), so business leaders can give their AI assistants some financial autonomy without handing them a blank cheque
Pre-trained ‘employees’ for safer automation.
A Silicon Valley startup, Kognitos, is taking another approach to making AI agents enterprise-friendly: give them pre-defined jobs and rules to follow. On 21 October, Kognitos launched a set of 61 “AI Employees”, software agents pre-trained for common finance and banking roles ([1]). Each comes with a library of standard operating procedures (SOPs) that define exactly how it should handle tasks like invoicing, accounts reconciliation or Know-Your-Customer checks. The system also includes an agent called Quill that can draft new SOPs in plain English for a human manager to review and approve ([2]).
Kognitos is betting that clearly coded processes and built-in audit trails will help companies trust AI agents with sensitive work that has to be done the same way every time. Its CEO notes that many CFOs want AI back-office helpers “that follow approved procedures rather than making up the process as it goes” ([3]). The need for such consistency is backed by research: in one 18-month study of 15 AI models, accuracy improved over time but even the best agents still produced inconsistent results from one run to the next ([4]). Kognitos is trying to close that reliability gap by fusing large-language models with deterministic code execution, an approach in line with Lumo’s view that AI works best in well-defined workflows with oversight. Wider data supports this: an IBM global survey found that organisations embedding automated controls into their AI systems had 25% fewer incidents than those relying on manual governance ([5])
CFOs warn of runaway AI spending.
A new survey of chief financial officers is highlighting a different challenge as AI agents take on more work: the risk of unpredictable costs. In Deloitte’s third-quarter CFO Signals poll, published 21 October, 74% of large-company CFOs said it would take their organisations from several days to weeks to cut off an AI system whose usage charges were exceeding its budget ([1]). Only 7% of these billion-dollar firms could intervene in near-real time to stop an AI tool from overspending ([2]). This is a stark change from traditional software, where companies buy fixed licenses; now that generative AI is often metered by usage, “AI bills can spike quickly due to large context windows or long conversation loops” ([3]), the report notes.
This lack of cost control is more than a finance concern, it can slow adoption of helpful AI assistants. If a well-intentioned coding or research agent silently runs up a huge cloud bill, it becomes a business risk that will give CFOs pause. Deloitte’s survey found finance leaders are increasingly being tasked with AI governance ([4]), underscoring the importance of setting clear budgets and monitoring tools for any autonomous software that might otherwise operate without immediate human oversight. Under the right controls, however, the latest advances show that AI agents can shoulder more responsibility, even handling payments or complex processes, which could unlock valuable efficiencies across the enterprise