AI Agents Can’t Pay With M-Pesa Yet: The Agentic Gap

AI agents can now buy things — through Mastercard’s Agent Pay, Visa’s Intelligent Commerce, and OpenAI’s Instant Checkout in ChatGPT — but every one of those systems settles on card rails, and none of them can complete a payment on M-Pesa or MTN MoMo. The $1.4 trillion that flowed through Sub-Saharan Africa’s mobile money wallets in 2025 is, for practical purposes, invisible to the agentic commerce stack being standardized in San Francisco and Purchase, New York. That gap is the largest open opportunity in African fintech this decade — and the builders and merchants who close their side of it first will own the next commerce layer.

Key Takeaways

• Agentic commerce went from concept to production in eighteen months: Mastercard announced Agent Pay in April 2025, OpenAI and Stripe shipped Instant Checkout in ChatGPT with the open Agentic Commerce Protocol in late 2025, Google’s AP2 launched with 60+ partners, and Mastercard’s Agent Pay for Machines arrived in June 2026 (1)(2)(3)(4).
• Every major protocol — ACP, AP2, Visa Intelligent Commerce, Agent Pay — is built on card-network primitives: tokenized credentials, signed mandates and shared payment tokens. None natively supports mobile money (3)(4)(5).
• Meanwhile the rails Africans actually use processed $1.4 trillion in 2025 — $806 billion in East Africa alone, the most of any region on earth — across 1.2 billion registered accounts (6).
• The developer tooling gap is documented: until independent builders intervened, the Model Context Protocol ecosystem contained zero African fintech integrations — no M-Pesa, no Africa’s Talking, no USSD (7); community-built Daraja MCP servers only began appearing in 2025–26 (8).
• McKinsey-cited projections put AI-agent-driven transactions at $1 trillion in the US alone by 2030 (2); whoever connects that behavior to mobile money rails captures an African analogue nobody is yet positioned to own.
• The gap is not one bridge but three — authorization, rails and trust — and each can be started now by Safaricom-scale incumbents, PAPSS-scale institutions, or a two-person team with an API key.

What is agentic commerce, and why is it arriving so fast?

Agentic commerce is what happens when the customer’s AI assistant stops recommending and starts transacting: searching, comparing, negotiating, placing the order and paying — under instructions and limits the human set in advance. It moved from white paper to checkout button with startling speed because the world’s payment giants decided, almost simultaneously, that it was existential.

The timeline tells the story. In April 2025 Mastercard unveiled Agent Pay, issuing “Agentic Tokens” — an extension of the same tokenization service that secures contactless cards — so that verified AI agents can transact on a consumer’s behalf (1). Visa answered with Intelligent Commerce, scoped tokenized credentials and machine-authentication frameworks, integrated with the major LLM platforms (5). In late 2025 OpenAI and Stripe shipped Instant Checkout inside ChatGPT, powered by the open-source Agentic Commerce Protocol (ACP): a Shared Payment Token lets ChatGPT complete a purchase from Etsy or Shopify merchants without ever seeing the buyer’s card details (3). Google convened more than sixty organizations — Mastercard, PayPal, Adyen among them — around AP2, the Agent Payments Protocol, whose core primitive is the mandate: a digitally signed statement of what an agent is authorized to buy, for how much, until when (4). By June 2026 Mastercard had extended the stack to Agent Pay for Machines, settling machine-to-machine micropayments as small as fractions of a cent at machine speed (2). Analysts frame the stakes with a McKinsey projection of $1 trillion in US agent-driven transactions by 2030 (2).

Notice what every layer of this new stack assumes: a card network underneath. Tokenized credentials extend card tokenization. Shared Payment Tokens are issued by card processors. Mandates terminate in card authorizations. The architecture is elegant, open — and built entirely on rails that most African consumers do not use. Some analysts argue stablecoins will become the settlement layer for AI buyers precisely because they are programmable where legacy rails are not (9) — an argument Africa should read as both warning and clue.

Why can’t an AI agent pay with M-Pesa today?

Three reasons, and only one of them is software.

First, the authorization model was designed to require a human thumb. Mobile money’s security rests on the SIM card, the PIN, and a prompt on a handset: an STK push appears on your phone, you enter your PIN, money moves. That design is why fraud stayed manageable while mobile money grew into a $1.4 trillion system — but it means there is no native way to tell M-Pesa “this agent may spend up to 50,000 shillings on my behalf this week.” Card networks had tokenization infrastructure lying ready to repurpose into agentic credentials (1)(5); mobile money has no equivalent of the mandate — yet. The human-in-the-loop is hard-coded, which protects users and excludes their agents in the same stroke.

Second, the developer tooling simply was not built. This is the documented, embarrassing part. As one Kenyan developer recorded after auditing the ecosystem in 2026: the Model Context Protocol registry — the standard way AI agents discover tools — contained zero African fintech entries. No M-Pesa. No Africa’s Talking. No USSD. He contributed the documentation and built the first MCP server connecting agents to those APIs himself, noting that the undocumented quirks (STK Push’s base64 password generation, webhook verification, Kenyan phone-number normalization) were exactly what kept agents locked out (7). Community-built Daraja MCP servers have since multiplied, letting AI systems initiate and reconcile M-Pesa transactions in sandbox and beyond (8). The raw APIs were always there — Safaricom’s Daraja exposes 22 M-Pesa APIs, and MTN’s MoMo open API has run since 2018 (10) — but an API without agent tooling is a road without an on-ramp.

Third, the institutions have not yet decided to want it. Telcos earn from human-initiated transaction fees and have regulatory duties built around human KYC. An agent layer raises hard questions — whose PIN, whose liability, whose fraud loss? — that are answerable (the card networks just answered them with verified-agent registries and signed mandates) but not yet asked loudly in Nairobi or Kampala. JEPA’s analysis of agentic AI in East African payments reaches the same conclusion: the capability conversation is global, the rails conversation has not started locally (11).

The result is an absurdity worth stating plainly: a ChatGPT agent can buy a $40 candle from an Etsy seller in Ohio tonight, but cannot buy a 4,000-shilling bag of maize flour from a Nakuru wholesaler whose till sits on the most successful digital payment system in the developing world.

How big is the prize on the other side of the gap?

Run the numbers on what is waiting to be connected. Sub-Saharan Africa processed $1.4 trillion in mobile money transactions in 2025 — two-thirds of the global total — with East Africa the world’s densest corridor at $806 billion. The region holds 1.2 billion registered accounts and the world’s largest population of monthly-active mobile money users (6). This is not a market that needs creating; it is a market that needs an adapter.

And the cross-border layer is consolidating at exactly the right moment. PAPSS, the Pan-African Payment and Settlement System, now connects 19 countries and over 160 commercial banks, settling cross-border trades in local currencies in seconds — and its February 2026 partnership with Kenya’s Pesalink reaches into banks and mobile wallets directly (12). When an AI procurement agent in Kigali can eventually pay a supplier in Accra, the settlement will run through rails like these. An agentic layer on national wallets plus PAPSS underneath is, structurally, the African answer to the card networks’ global stack — built on the payments integration the region is already assembling.

The deeper shift makes the prize larger still. Agentic commerce is one expression of the broader move from software you operate to services that execute outcomes on your behalf — and payments are where every executed outcome terminates. Whoever owns the agent-to-wallet connection does not just earn transaction fees; they become the default checkout of the AI era for half a billion people. McKinsey’s $1 trillion US projection (2) has no published African counterpart, which is itself the point: the analysis gap mirrors the tooling gap, and both reward whoever moves first.

Who might actually close the gap?

Four candidates, in descending order of firepower and ascending order of speed.

Safaricom and the telco wallets. Safaricom has the assets to make M-Pesa the first agent-ready mobile money system in the world: Daraja’s API estate, tokenization experience, and the fraud-modeling depth to underwrite verified-agent risk. The move would rhyme with history — M-Pesa won by making payments programmable for humans before anyone else believed in it. MTN’s MoMo open API platform, live across a dozen markets, gives it the same option at pan-African breadth (10). The question is incentive and regulatory courage, not capability.

PAPSS and the institutional layer. A continental settlement system adding an agent-mandate standard — even a pilot — would give every connected bank and wallet a common answer to “how does an agent get authorized?” instead of forty incompatible ones (12). Institutions move slowly, but standards are their natural product.

Fintechs and stablecoin builders. The pragmatic wedge: a fintech that issues scoped virtual cards or stablecoin wallets funded from M-Pesa/MoMo balances gives agents something they can already spend — an interim bridge that works with today’s protocols while native mobile-money mandates mature (9). Expect the first production agentic purchases “via mobile money” to actually settle this way, with the wallet as funding source rather than rail.

Independent builders. The MCP story shows the floor is open: one developer with API keys and patience put African fintech into the agent ecosystem before any institution did (7)(8). The protocols are open-source — ACP is Apache-licensed, AP2 is public (3)(4) — and nothing prevents an East African team from drafting the mobile-money mandate extension and forcing the conversation. In agentic commerce, the spec-writer’s advantage is real: the protocols being ratified now will be the plumbing for decades.

The Three Bridges: a map of what must be built

Strip the problem to its load-bearing structure and the gap is not one bridge but three. I offer this as the Three Bridges Framework — the checklist for any builder, investor or policymaker asking where agentic mobile money actually stands.

Bridge 1 — the Mandate Bridge (authorization). The missing primitive: a way for a wallet owner to grant an agent bounded spending authority — amount caps, merchant categories, expiry, revocation — with liability defined when it goes wrong. Card rails built this as signed mandates and agentic tokens (1)(4)(5); mobile money needs its own version, designed for PIN-and-SIM culture: think a sub-wallet with a hard cap that the owner tops up deliberately, the prepaid-airtime mental model applied to agent authority. Whoever ships a credible mandate primitive for M-Pesa or MoMo has built the keystone.

Bridge 2 — the Rail Bridge (tooling). Everything that lets an agent discover, call and reconcile mobile money: MCP servers, SDKs, sandboxes, webhook standards, documentation that does not require tribal knowledge (7)(8). This bridge is the cheapest and is already half-built by volunteers. It is also where a small team can contribute this quarter — every API documented for agents, every edge case captured, compounds for the whole ecosystem.

Bridge 3 — the Trust Bridge (the merchant side). An agent with money still needs merchants it can transact with: machine-readable catalogs with true prices and stock, verifiable business identity, instant programmatic confirmation, and a dispute path. This bridge is built one SME at a time — and it is the same construction as the human-facing one, because an agent-ready merchant is simply a well-run conversational commerce stack whose shelf is accurate and whose till confirms instantly. Grounded catalogs that never let an AI guess a price are the merchant’s half of the handshake.

The framework’s diagnostic use: most discourse fixates on Bridge 1 (the hard institutional problem) while ignoring that Bridges 2 and 3 are open to anyone, today, at near-zero capital cost. Build from the edges in.

What should first-movers do this year?

If you build: ship Bridge 2 artifacts now — MCP servers, mandate-wrapper prototypes using sub-wallets or scoped virtual cards, agent-readable catalog standards for WhatsApp-based merchants. Publish openly; in protocol races, adoption beats secrecy. The first credible “agentic checkout for mobile money” demo will have every accelerator, telco innovation lab and fintech conference on the continent asking for a meeting.

If you run an SME: become agent-ready by becoming conversation-ready. Accurate digital catalog, grounded automated answers, instant payment confirmation, exported customer records — every layer serves your human customers this quarter and their agents next year. The discipline is identical, and the evidence on AI returns already favors firms that build it.

If you invest or regulate: treat the mandate question as urgent and answerable. The card networks proved agent authorization can be done safely at global scale in eighteen months. A sandbox pilot — one telco, one regulator, capped sub-wallet mandates, full audit trail — would put East Africa ahead of every emerging market on earth in the next payments paradigm, on rails it already owns.

The West is building agentic commerce for the cards it has. Africa gets to build it for the rails it actually uses — rails that are younger, more programmable in spirit, and already woven into daily life more deeply than cards ever were anywhere. The gap between the two stacks is not a verdict. It is a blueprint, drawn in negative space, of the most valuable infrastructure project on the continent’s roadmap. Someone reading this will build a bridge. Start with the edge you can reach.

Frequently Asked Questions

What is agentic commerce?
Agentic commerce is buying and selling executed by AI agents acting under human-set instructions and limits — searching, comparing, ordering and paying autonomously. Production systems arrived in 2025–26 via Mastercard Agent Pay, Visa Intelligent Commerce, and OpenAI and Stripe’s Instant Checkout in ChatGPT, all settling on card networks.

Why can’t AI agents use M-Pesa or MTN MoMo?
Mobile money authorization requires a human entering a PIN on a SIM-registered handset, with no mandate primitive for delegating bounded spending authority to an agent. Until recently the developer tooling was also absent — the MCP agent ecosystem had zero African fintech integrations before independent builders created the first Daraja servers.

How large is Africa’s mobile money market?
Sub-Saharan Africa processed $1.4 trillion in mobile money transactions in 2025 — about two-thirds of the global total — across 1.2 billion registered accounts. East Africa is the world’s densest corridor at $806 billion, which is why connecting agents to these rails is such a consequential prize.

Who is best positioned to connect AI agents to mobile money?
Safaricom and MTN hold the APIs, tokenization experience and fraud models; PAPSS could standardize agent mandates across 19 countries and 160 banks; fintechs can bridge sooner with scoped virtual cards or stablecoin wallets funded from mobile money; and independent builders are already shipping the open-source agent tooling.

What should a small merchant do to prepare for agentic commerce?
Become machine-trustworthy: keep an accurate digital catalog, ground automated answers in approved prices and stock, confirm payments instantly, and own your customer records. An agent-ready merchant is the same thing as a well-assembled conversational commerce operation — preparation pays with human buyers immediately.

Related Reading

• From Chat to Checkout: The Conversational Commerce Stack East Africa Is Assembling
• PAPSS and the Integration of East African Payments
• The Service-as-Software Shift
• Does AI Pay Rent? The SME Evidence Review

Sources and Evidence

1. Mastercard Newsroom, “Mastercard unveils Agent Pay, pioneering agentic payments technology to power commerce in the age of AI,” April 29, 2025. https://www.mastercard.com/us/en/news-and-trends/press/2025/april/mastercard-unveils-agent-pay-pioneering-agentic-payments-technology-to-power-commerce-in-the-age-of-ai.html — Primary corporate announcement of Agentic Tokens built on MDES tokenization.
2. Mastercard Newsroom, “Mastercard launches Agent Pay for Machines to unlock super-fast, always-on payments,” June 2026. https://www.mastercard.com/us/en/news-and-trends/press/2026/june/mastercard-launches-agent-pay-for-machines.html — Primary source for machine-to-machine micropayments and the McKinsey $1T projection.
3. Stripe Newsroom, “Stripe powers Instant Checkout in ChatGPT and releases Agentic Commerce Protocol codeveloped with OpenAI.” https://stripe.com/newsroom/news/stripe-openai-instant-checkout — Primary source on ACP, Shared Payment Tokens and Etsy/Shopify merchant rollout.
4. Vellum, “Google’s AP2: A new protocol for AI agent payments.” https://www.vellum.ai/blog/googles-ap2-a-new-protocol-for-ai-agent-payments — Technical analysis of AP2’s signed-mandate model and its 60+ launch partners including Mastercard, PayPal and Adyen.
5. Digital Commerce 360, “Visa and Mastercard both launch new agentic AI payments tools,” October 2025, and “How Visa and Mastercard are approaching agentic commerce,” April 2026. https://www.digitalcommerce360.com/2026/04/02/visa-mastercard-in-agentic-commerce/ — Trade-press comparison of Visa Intelligent Commerce and Agent Pay architectures.
6. GSMA, “State of the Industry Report on Mobile Money 2026,” March 2026. https://www.gsma.com/sotir/ — Industry-association flagship dataset: $1.4T SSA transaction value, East Africa at $806B, 1.2 billion registered accounts.
7. Mahia, G., “Why M-Pesa, Africa’s Talking, and USSD are missing from AI agent tooling — and what I did about it,” DEV Community, 2026. https://dev.to/gabrielmahia/why-m-pesa-africas-talking-and-ussd-are-missing-from-ai-agent-tooling-and-what-i-did-about-it-56fo — First-person practitioner documentation of the zero-African-fintech MCP gap and the first remediation work.
8. PulseMCP / community registries, “M-Pesa (Safaricom Daraja API) MCP server.” https://www.pulsemcp.com/servers/jameskanyiri-daraja-mpesa — Evidence of community-built agent tooling connecting AI systems to Daraja; open-source, early-stage.
9. FinTech Weekly, “Agentic commerce, stablecoins and micropayments: the missing settlement layer for AI payments.” https://www.fintechweekly.com/magazine/articles/agentic-commerce-stablecoins-micropayments-ai-payments — Industry analysis of programmable settlement for AI buyers; cited for the stablecoin-bridge argument.
10. Safaricom, “Daraja Developer Portal” (22 M-Pesa APIs) and MTN, “MoMo Developer Portal” with Ericsson case study on the Uganda-first open API program. https://developer.safaricom.co.ke/ and https://momodeveloper.mtn.com/ — Primary API documentation establishing that the raw rails are programmable today.
11. JEPA, “Agentic AI in East Africa’s mobile payment services.” https://www.jepaafrica.com/insights/bab1o5d9t20jym41ctzhur78u8a9ff — Regional policy-analysis outlet framing the local institutional gap; treated as analytical commentary.
12. TheCable, “PAPSS now connects 19 countries, facilitates cross-border payments in seconds,” and African Export-Import Bank, “Pesalink and PAPSS unlock cross-border payments in local currencies in Kenya,” February 2026. https://www.thecable.ng/papss-now-connects-19-countries-facilitates-cross-border-payments-in-seconds/ and https://www.afreximbank.com/pesalink-and-papss-unlock-cross-border-payments-in-local-currencies-in-kenya/ — Press and institutional primary sources on PAPSS scale (19 countries, 160+ banks) and the Kenya mobile-reach partnership.