Today on The Merchant Desk: as we've been tracking over the past week, the major card networks are not waiting for agentic commerce to happen to them. We look at the disintermediation threat Visa and Mastercard's divergent AI payment rails pose to traditional banks, along with the execution guardrails needed to keep these new autonomous systems from double-spending.
Following the divergent agentic payment architectures we've been tracking from Visa and Mastercard—where Visa is embedding via OpenAI and Mastercard is building infrastructure for machine-to-machine settlement—a clear disintermediation risk is emerging for traditional banks. As fintechs like Coinbase and Robinhood also make their platforms 'agent-agnostic', the primary customer interface could shift from banking apps to AI agents.
Why it matters
This is a significant strategic inflection point. The two largest card networks are actively building the payment infrastructure for an AI-native economy, moving the point of transaction into conversational interfaces. For banks, particularly in markets like South Africa, this isn't a future-state problem; it's a current architectural threat. If the primary relationship for initiating payments and managing finances moves to an AI agent, banks risk being relegated to a back-end utility, losing out on customer data, cross-selling opportunities, and deposit liquidity.
AI shopping assistants are prone to 'hallucinating' incorrect product information or unsafe advice because models prioritize conversational fluency over factual accuracy. A new analysis from Saturday argues that model improvements and prompt guardrails are insufficient. The only robust solution is a dedicated verification layer that checks every AI-generated answer against a ground-truth product knowledge graph before it reaches the customer.
Why it matters
This moves the conversation about AI risk from a theoretical problem to an operational requirement. For any operator deploying customer-facing AI, this is critical. Hallucinations aren't an edge case; they are a systemic failure mode that can lead to lost sales, brand damage, and legal liability (as seen with Air Canada's chatbot). This argues that the real investment isn't just in the AI model, but in the structured, AI-ready data and verification architecture needed to make the model safe for commercial use. It's a call to action to treat data infrastructure as a prerequisite for AI deployment, not an afterthought.
At its Payments Forum on Wednesday, Visa announced a suite of new capabilities for agentic commerce, including 'Agent Score' for readiness assessment, an 'Agentic Directory,' a 'Large Transaction Model' for fraud detection, and enhanced token data. This rollout complements its OpenAI collaboration and comes as reports show Visa's stablecoin settlement run rate has reached $7 billion annually.
Why it matters
Visa is assembling a full-stack toolkit to facilitate agent-initiated commerce. This isn't just about the payment itself, but the entire trust and discovery lifecycle: how do you know an agent is legitimate (Agent Score)? How do you find it (Agentic Directory)? How do you protect the transaction (Large Transaction Model)? This shows a deep understanding of the operational challenges, positioning Visa not just as a rail, but as a governance and security layer for the emerging agent economy.
As Mastercard rolls out the 'Agent Pay for Machines' (AP4M) stack we've been tracking this week, a new technical analysis flags a critical 'execution gap' around 'exactly-once execution safety.' Because the protocol doesn't inherently prevent an AI agent from retrying a transaction after a crash, it can lead to duplicate payments in production environments. The analysis argues for a 'claim before execute' pattern to ensure transactional integrity.
Why it matters
This is a crucial, operator-level critique. While the high-level strategy for agentic payments is taking shape, the low-level technical details required for robust, reliable deployment are still being worked out. For anyone building on these new platforms, this highlights a significant operational risk. Financial errors like duplicate charges erode trust and create costly reconciliation nightmares. It's a reminder that for autonomous systems, execution safety is not an implementation detail—it's a core requirement for commercial viability.
Visa is introducing a new virtual and tokenized payment credential specifically designed for AI agents like ChatGPT. The system allows agents to initiate and complete purchases on behalf of users, while incorporating user-defined controls like spending limits and merchant category restrictions, all secured by Visa's existing real-time authorization and fraud monitoring infrastructure.
Why it matters
This is a concrete implementation of the 'agentic commerce' vision, built on top of existing payment rails. By creating a specific credential for agents, Visa is solving the critical trust and control problem. It allows users to safely delegate purchasing power to an AI without giving it unchecked access to their primary account. For merchants, it means AI-driven sales can flow through the standard, secure Visa network. This is a powerful move to ensure the card networks remain central even as the user interface for commerce radically changes.
Mastercard introduced 'Agent Pay for Machines' (AP4M) on Wednesday, a new service designed for high-speed, programmatic payments between AI agents and machines. Over 30 partners, including Adyen, Stripe, and Coinbase, are supporting the initiative, which aims to provide the credentialing, permissioning, and settlement infrastructure for a new class of high-volume, low-latency automated transactions spanning cards and stablecoins.
Why it matters
This is Mastercard's architectural answer to the rise of autonomous commerce. It's not about replicating a human point-of-sale transaction but enabling a fundamentally new type of commerce that is always-on and software-driven. For operators, this signals the development of a necessary, parallel payment infrastructure. The inclusion of stablecoins alongside traditional card rails is a key strategic decision, acknowledging that the future of machine-to-machine value exchange will likely be multi-rail.
South African Reserve Bank Governor Lesetja Kganyago on Saturday cited India's Unified Payments Interface (UPI) as a potential model for South Africa's own push to reduce cash dependence. Speaking at the G20 Techsprint finale in Mumbai, Kganyago highlighted UPI's efficiency and accessibility as a blueprint for developing a free, real-time national payment system.
Why it matters
This is a strong signal of the SARB's strategic direction, moving beyond PayShap towards a more comprehensive, low-cost digital payments ecosystem. For fintech operators in South Africa, this is a major tailwind. A government-backed push for a UPI-like system would dramatically accelerate the shift to digital payments, create new opportunities for payment service providers, and fundamentally reshape the competitive landscape for merchant acquiring and consumer payments. The key will be watching how they plan to address SA-specific challenges like power outages and internet penetration.
The Model Context Protocol (MCP), an open standard published by Anthropic in late 2024, has rapidly become the key integration layer for enterprise AI agents, according to analysis from Sunday. With over 10,000 public MCP servers now deployed, the protocol standardizes how AI agents interact with business systems like CRMs, databases, and other tools, enabling them to access and coordinate information autonomously.
Why it matters
MCP is solving the 'last mile' problem for enterprise AI: how to securely connect powerful models to proprietary internal data. For operators building or using AI agents, this is the architectural glue. It's analogous to how APIs enabled the web and mobile app ecosystems. Understanding MCP is now critical for designing scalable AI commerce systems, as it defines how agents will access product catalogs, process orders, and interact with payment infrastructure.
As AI inference costs become a primary concern for developers, two operational tactics for managing Anthropic's Claude models are gaining traction. A guide published Saturday highlights the Claude Batch API, which offers a 50% discount on token costs for asynchronous tasks. A separate analysis on Sunday focuses on Claude's context compression, a server-side feature that can reduce token usage by 40-60% without sacrificing quality.
Why it matters
This is about the unit economics of AI. The initial hype of building with LLMs is now meeting the financial reality of running them at scale. These architectural patterns are no longer nice-to-haves; they are essential for any business building a sustainable AI-powered product. For any operator with an AI component in their stack, leveraging these cost-control features can be the difference between a profitable service and one that bleeds cash on every API call.
Monnify, operated by TeamApt under Moniepoint Inc., processed N25 trillion (approx. $17 billion) in transactions in 2025, a 38% increase from 2023. This growth, reported on Saturday, highlights an intensifying battle among Nigerian fintechs to control the underlying payment infrastructure, shifting focus from consumer apps to the foundational payment rails that serve over 100,000 merchants and 27 banks.
Why it matters
This is a sign of market maturity. The land grab for consumer-facing apps is giving way to a more strategic, and potentially more profitable, battle for the core infrastructure. Monnify's success shows that the real moat is in the regulatory licensing, direct bank integrations, and reliability of the payment processing layer itself. For pan-African commerce, this deepens the infrastructure in its largest market, making it easier for others to build on top, but also raises the competitive stakes for new entrants.
Fintech Kora announced on Sunday its integration into the International Air Transport Association's Financial Gateway (IFG). The move aims to standardize and secure airline transactions in major African aviation hubs like Nigeria and Kenya, providing a single access point for airlines to accept diverse local payment methods, including mobile money and local cards, and reducing booking failures.
Why it matters
This is a significant infrastructure play for a high-value, high-friction sector in Africa. Fragmented payment systems have long plagued the African aviation industry, leading to lost revenue for airlines and frustration for travelers. By plugging into IATA's global standard, Kora is effectively creating a payment switch for African airlines, simplifying cross-border settlement and making it easier to do business across the continent. This is a classic example of fintech solving a structural, B2B problem.
Buy Now Pay Later (BNPL) is driving significant growth in South African retail, now accounting for almost 10% of sales at major retailer Edgars, according to reports Saturday. The payment method is attracting younger shoppers and increasing basket sizes, prompting providers like PayJustNow to expand into new sectors such as travel and healthcare.
Why it matters
This is a hard data point confirming BNPL's transition from a niche offering to a mainstream payment method in South African retail. For large merchants like Edgars and TFG, it's now a material part of their sales mix. For payment operators, this signals a durable shift in consumer behavior and a clear opportunity. The strategic question is no longer *if* merchants should offer BNPL, but *how* it should be integrated into their POS and e-commerce checkout flows to maximize conversion and loyalty.
The Agentic Commerce Land Grab Visa and Mastercard are aggressively positioning themselves as the default payment layer for AI agents, partnering with OpenAI and launching dedicated machine-to-machine payment services. This is creating a disintermediation threat for traditional banks, forcing them to consider how they integrate with LLMs or risk losing the primary customer interface.
AI Needs Guardrails: Verification & Execution Safety As AI agents move into production for commerce, the focus is shifting to operational risks. Two key themes emerge: the need for a 'verification layer' to prevent AI shopping assistants from hallucinating incorrect information, and the 'execution gap' in payment protocols to prevent issues like duplicate transactions.
African Fintech Infrastructure Matures The focus in African fintech is moving from consumer-facing apps to the underlying infrastructure. This is visible in Monnify's N25 trillion processing volume in Nigeria, Kora's integration with IATA to standardize airline payments, and Nigeria's push to digitize millions of MSMEs. The battle is now for control of the rails.
Technical Levers for AI Cost Control As AI moves into production, managing inference costs is becoming a critical business discipline. Developers are now focusing on specific architectural patterns like using Anthropic's Claude Batch API for a 50% cost reduction on non-real-time tasks and implementing robust spend governance to prevent runaway API expenses.
The UPI Model Goes Global South Africa's Reserve Bank is now openly looking to India's Unified Payments Interface (UPI) as a blueprint for its own real-time, low-cost national payment system. This signals a strategic direction for emerging markets to leapfrog legacy card infrastructure.
What to Expect
2026-06-15—Finextra editorial on European payments sovereignty and the future of Visa/Mastercard in the region goes live.
2026-06-17—SAP hosts a webinar on Agentic B2B Commerce, focusing on practical applications for profitable growth.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
308
📖
Read in full
Every article opened, read, and evaluated
114
⭐
Published today
Ranked by importance and verified across sources
12
— The Merchant Desk
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste