AI & Agentic Payments

Two waves are colliding inside embedded payments: AI inside the rails (fraud, underwriting, optimization, personalization) and AI on top of the rails (autonomous agents that transact). Platforms that design for both will define the next decade. Those that don’t will get disintermediated by something that did.

22 min read

Why this chapter, and why now

Every previous chapter of this guide treated payments as static infrastructure: cards, networks, processors, PayFacs, and the operational levers that compress economics and grow attach rate. That framing has been correct for the last 15 years. It is now actively incomplete.

Two structural changes are happening in parallel. The first — AI inside the payment rails — has been underway since at least 2018 (Stripe Radar, Visa Advanced Authorization) and is now the default. Every serious processor, network, and lending product runs on machine learning. If your platform is not benefiting from those models, you are paying a hidden tax in fraud loss, declined transactions, and missed credit decisions.

The second — agentic payments — is genuinely new. In the span of 12 months across 2025 and early 2026, every major payments network and infrastructure provider shipped some version of an agent-enabled commerce stack. Visa announced Intelligent Commerce (April 2025). Mastercard launched Agent Pay (April 2025). Google released the Agent Payments Protocol (AP2) in September 2025 in partnership with American Express, Mastercard, PayPal, and others. Stripe shipped the Agent Toolkit, then partnered with OpenAI to enable in-ChatGPT checkout for hundreds of millions of users. These are not pilots — they are production systems.

The combined effect: payments is becoming both more intelligent (AI inside) and more accessible to autonomous software (agentic on top). Platforms that design for that double shift will keep pace. Platforms that don’t will discover, in the next 24–36 months, that a generation of buyers (and increasingly, AI agents acting on behalf of buyers) has routed around them.

AI inside the rails — the current state

AI has been quietly transforming payments for years. The use cases below are not aspirational — they are deployed at scale across the networks, processors, and embedded finance leaders.

Real-time fraud detection

Fraud detection is the most mature AI application in payments. Stripe Radar processes hundreds of billions of dollars annually across millions of businesses, using machine learning trained on network-wide transaction data to assign a risk score to each payment within milliseconds. Visa’s AI-driven authorization platform prevents billions in attempted fraud annually. Bank of America’s AI fraud system flags suspicious card activity and triggers immediate alerts.

For vertical SaaS platforms, the advantage of AI fraud is vertical specificity: a model trained on your vertical’s transaction patterns can flag anomalies that a general-purpose model misses, with fewer false positives. The Worldpay/Payrix Vertical SaaS Benchmarking Study confirms vertical SaaS platforms achieve a 0.04% all-client chargeback rate — well below general industry benchmarks — partly because of this vertical specificity (see Chapter 6 for the full chargeback table).

Payment routing and authorization optimization

Modern processors use AI to dynamically route transactions to the optimal acquiring bank, retry failed authorizations with adjusted parameters, and time submissions to maximize approval rates. Stripe Adaptive Acceptance, Adyen RevenueAccelerate, and Checkout.com Intelligent Acceptance each report 1–3 percentage points of incremental authorization rates — at meaningful TPV, this is a multi-million-dollar lever with no merchant-side cost.

For platforms running flat-rate PFaaS configurations, this lift is captured by the processor, not by you. For platforms running IC+ or managed PayFac with auth-rate-shared economics, the math changes — improved approvals flow to platform GMV and to merchant retention. This is one of the underweighted reasons to migrate off flat-rate above $50M GMV (see Chapter 9).

Embedded credit underwriting

AI-driven underwriting has rewritten what is possible in SMB lending. Affirm has been “AI-driven from inception” in its consumer credit decisions, using machine learning on transaction and behavioral signals to underwrite point-of-sale loans in seconds. Sezzle’s “Prophet” credit model cut default rates from 3.2% to 1.8% for approved borrowers over one year. Shopify Capital analyzes over 70 million data points on sales and merchant performance to determine optimal loan amounts and eligibility — over $5 billion extended on the back of this model. Square Loans, Toast Capital, Parafin, and YouLend all run analogous models on platform payment data (see Chapter 8 for the data-moat mechanics).

The strategic point for vertical SaaS executives: if you are evaluating embedded lending partners (see Chapter 10), the underwriting quality of your partner — and your platform’s ability to feed them transaction-level data — is the difference between a marginal product line and a $10M+ revenue contributor.

Personalization and dynamic pricing

AI enables hyper-personalized financial offers in embedded contexts. Tesla’s embedded auto insurance adjusts premiums monthly based on AI analysis of driving behavior captured by the vehicle — a real-time pricing model unimaginable in traditional insurance. Affirm’s 2025 launch of AdaptAI personalizes loan offers (APR, term length, eligibility) for individual consumers based on real-time behavioral and transactional signals. OpenAI, working with Stripe, deployed an AI-autofill feature in DALL·E checkout that resulted in 40% faster payment completion — a direct conversion uplift on a non-financial flow.

For vertical SaaS platforms, the application is straightforward: AI on top of your payment data lets you tailor merchant pricing tiers, surface contextual financial products (working capital at the right moment, insurance at the right vertical), and reduce friction in your checkout flows. The Charge Forward Payments Revenue Calculator quantifies the revenue impact of conversion uplift at your platform’s specific GMV and ticket size.

Operational automation

AI also handles the unsexy work that drains operational margin in PFaaS and PayFac models: KYC document parsing, transaction monitoring, dispute response, settlement reconciliation. J.P. Morgan’s COIN platform reduced 360,000 hours of contract review to seconds using NLP. Equivalent automation in PFaaS operations means your compliance team can scale 10x faster than the merchant base, instead of growing linearly with it. Below $25M GMV this is a curiosity. Above $100M GMV it is a structural cost advantage.

Charge Forward Insight

The most common AI-in-payments gap we see at advisory engagements is not a missing model — it is a missing data pipeline. Platforms have rich transaction data sitting in their processor’s portal, but it never makes it into the platform’s own data warehouse. Without that flow, every AI-driven product your team wants to build (smart routing, custom fraud rules, embedded credit underwriting, personalized pricing) is permanently blocked. The fix is the data architecture in Step 2 of the Chapter 8 framework. Get that right, and every AI use case in this chapter becomes available to your platform. Skip it, and you outsource the data advantage to your processor for the life of the relationship.

Agentic payments — the new layer

If “AI inside the rails” is the current state, “agentic payments” is the next layer. An agent is software that can take actions on behalf of a user or business — discovering products, comparing options, negotiating terms, authenticating, and transacting. Until recently, agents could browse and reason but could not reliably pay. That changed in 2025–2026.

What shipped in the last 12 months

A condensed timeline of the agentic-payments stack going live:

The cumulative implication is significant. For the first time, an AI agent can be given a goal (“book me a flight under $400 and a hotel under $200”), reason through options, and execute the payment — without a human ever clicking a checkout button. The infrastructure is in production; the merchant adoption curve is the variable.

What changes for vertical SaaS platforms

Three structural changes follow from agentic payments going mainstream.

First, your merchants become agent-discoverable — or they don’t. Agents need machine-readable product catalogs, prices, availability, and credentials. Static HTML checkout pages will continue to work for humans for a long time; they will not work for agents. Platforms whose merchants surface structured product data via APIs (or via emerging standards like AP2) will be the merchants that agents transact with. The platforms that don’t will be invisible in the new layer.

Second, your checkout has to support delegated authentication. An agent transacting on a user’s behalf needs to prove the user authorized the transaction, prove the merchant authorized the agent, and execute on a payment instrument that hasn’t been physically presented. Network tokenization (Visa Token Service, Mastercard MDES) is the foundation. Agent-specific tokenization layers (Visa Intelligent Commerce, Mastercard Agent Pay) are the new layer on top.

Third, your data model has to handle agent counterparties. Disputes, refunds, fraud investigations, and customer support all assume a human is on the other end. Once a meaningful share of transactions are originated by agents, your reporting, your dispute response workflows, and your fraud rules need to know the difference. This is not yet a 2026 problem for most platforms. It will be a 2027–2028 problem for the platforms moving fast and a 2029+ problem for everyone else.

Where the agent commerce volume will land first

Agentic transaction volume will not arrive uniformly. The early pockets are predictable:

• Travel: high-intent, comparison-driven, parameter-rich (dates, destinations, budgets). Agents are already booking flights and hotels in pilot deployments.

• E-commerce SKU purchasing: B2C and B2B catalog buys where product data is structured. ChatGPT Shopping (Stripe-powered) is the live consumer example.

• B2B procurement: supplier discovery, quote comparison, contract execution. The unsexy, large-dollar use case where agentic ROI is highest.

• Subscription management: agents that audit, renew, and renegotiate recurring subscriptions on a user’s behalf. Privacy and authorization frameworks here are non-trivial.

• Vertical SaaS workflows: a contractor’s AI assistant booking equipment rental, ordering materials, paying subcontractors, paying insurance — all from inside a single agentic interface.

That last bullet is the one most relevant to readers of this guide. If your platform’s merchants are field-services contractors, restaurants, or property managers, their AI assistants of the next 36 months will increasingly transact through embedded rails. Your job is to make sure your rails are the ones the agent finds.

Charge Forward Insight

When advising platforms today on agentic readiness, three concrete tests are useful. (1) Can an external service programmatically retrieve a structured product/service catalog for one of your merchants? (2) Can an external service initiate and complete a payment to a merchant on a credentialed customer’s behalf, with delegated authentication and tokenized credentials? (3) Does your reporting distinguish agent-originated transactions from human-originated transactions? Almost no vertical SaaS platform answers “yes” to all three today. The ones that get to “yes” on the first two by end of 2026 will be the ones agents transact with in 2027.

How AI changes the monetization math

The economics chapters of this guide (5, 6, 10) treat monetization as a function of attach rate, take rate, and product-line expansion. AI inside the rails — and especially agentic on top — alters each lever.

Conversion uplift on existing GMV

AI-optimized checkout, dynamic authentication, and agent-readable flows together can move authorization and conversion rates by hundreds of basis points. The OpenAI–Stripe DALL·E result (40% faster payment completion) is a leading indicator — agentic flows have less friction than human flows, not more, because the agent has perfect memory of credentials, addresses, and preferences. On a $250M GMV platform, a 1-point conversion improvement is $2.5M in incremental flow — captured by the platform if the economics model is IC+ or PayFac, captured by the processor if the model is flat-rate.

New revenue from agent-aware infrastructure

Platforms that expose agent-readable APIs, structured catalogs, and delegated-authentication endpoints will have a service surface that is monetizable on its own. Charging agents (or agent operators) for verified access, premium catalog placement, or transaction-quality signals is the agentic-era analog of marketplace listing fees. Too early to size with precision. Almost certainly real by 2028.

Embedded finance acceleration

Every embedded finance product — working capital, insurance, banking, payroll — gets better with AI underwriting and worse without it. The platforms that ship AI-enabled lending in 2026–2027 will see meaningfully higher attach rates and lower default rates than the platforms that ship the same product with rule-based underwriting in 2028. The difference compounds: better data → better models → better unit economics → more product expansion. (See Chapter 10 for the embedded finance sequencing framework.)

Subscription monetization for agents themselves

If your platform’s merchants serve consumers, an emerging monetization layer is the AI agent that represents the consumer. ChatGPT, Claude, Gemini, and consumer agentic apps will increasingly route transactions to your merchants — for a fee, a referral, or a placement model not yet standardized. Platforms that build the relationships and API surfaces today will be the ones with monetizable agent-traffic distribution tomorrow.

Embed finance or embed AI?

A question that comes up in every CEO/CFO conversation about embedded finance in 2026: “should we be putting our energy into AI features, or into embedded finance?” The honest answer is: this is a false choice. The leading platforms are doing both, in part because each makes the other more valuable.

AI thrives on data and context. Embedded finance generates both, at higher quality than any third-party data source. Embedded finance, in turn, benefits from AI for underwriting, fraud, personalization, and operational efficiency. The compounding effect is the data flywheel described in Chapter 8: more attach rate → more data → better AI models → better financial products → more attach rate.

The competitive evidence reinforces the synergy. Toast, the canonical vertical-SaaS embedded-finance story, uses ML extensively for its fintech products (lending risk, fraud, scheduling) and is now layering generative AI on top for restaurant operator tools. Affirm, an AI-first fintech, just launched AdaptAI to make its embedded credit product more personalized. OpenAI, an AI-first company, partnered with Stripe to enable monetization — agentic commerce being the natural extension. The pattern in all three: AI and embedded finance reinforce each other, and the firms with both are pulling away.

For vertical SaaS executives running this trade-off, the cleanest framing is:

• If you do not yet have embedded payments at meaningful attach rate, that is still the highest-ROI move — start there.

• If you have embedded payments live but no AI inside your data flows, that is the next gap to close — fraud, routing, underwriting, personalization.

• If you have both AI and embedded payments live, the agentic readiness work (catalog APIs, delegated authentication, agent-aware reporting) is the next strategic moat.

• If you are AI-first and don’t yet have embedded payments, you will need to. The OpenAI–Stripe partnership is the proof point: even the most AI-native company in the world needed embedded payments to monetize at scale.

The Charge Forward Embedded Payments Maturity Framework places the AI-and-agentic work explicitly inside Stages 3–5. At Stage 1 (Capability) and Stage 2 (PFaaS Transition), focus is on attach rate and economics. From Stage 3 (Margin Expansion) onward, AI-inside-the-rails and agent-readable infrastructure become the lever that separates leaders from peers in the same GMV band.

What can go wrong

Every chapter in this guide has been bullish on the direction of embedded payments and embedded finance. This chapter ends with a list of failure modes specific to AI and agentic payments — not because they will dominate, but because being aware of them makes the bullish case easier to execute on.

• Black-box underwriting. AI credit models that cannot explain individual decisions face increasing regulatory scrutiny (CFPB, state regulators, ECOA adverse-action rules). Vendor selection should include model-interpretability standards.

• Bias and adverse selection. Models trained on historical data inherit historical biases. Embedded credit programs that systematically under-serve specific populations face both regulatory and reputational risk. Periodic third-party model audits should be in the operating budget.

• Agent authorization fraud. As agents proliferate, so do agent-impersonation attacks. Visa Intelligent Commerce, Mastercard Agent Pay, and AP2 each address this, but the security model is novel and untested at scale.

• Network bypass via stablecoins and direct rails. Agentic commerce protocols are not network-locked. Some early agentic flows use stablecoins or A2A rails that route around the card networks. This is opportunity for issuers building stablecoin programs and risk for platforms whose economics depend on card-network interchange share (see Chapter 5 for the take-rate decomposition).

• Privacy and consent friction. AI agents transacting on a consumer’s behalf require explicit, granular, revocable consent. Building this correctly is non-trivial. Building it incorrectly invites regulatory action.

• Vendor lock-in to closed AI ecosystems. If your only agentic-commerce surface is OpenAI or Anthropic or Google, you are renting your distribution. Multi-channel agent readiness — AP2-compliant, multi-network-tokenized, multi-LLM-integrated — is the hedge.

Charge Forward Insight

The agentic payments stack is moving fast enough that any specific protocol named in this chapter may have been superseded by the time you read it. The right operational stance is not to chase the latest acronym — it is to build the underlying capabilities (machine-readable catalogs, delegated authentication, network tokenization, real-time data pipelines, AI-inside fraud and underwriting) that every credible standard will require. The standards will converge. The capability requirements won’t change.

What’s next

This is the final chapter of the Embedded Payments Knowledge Hub. The arc of the guide — from market thesis (Chapter 1) through operational economics (Chapters 5–6) through adoption (Chapter 7) through data and embedded finance (Chapters 8 and 10) and now to AI and agentic payments — should leave you with a coherent operating model for building a payments-and-finance business on top of vertical software.

Three things to do this quarter, regardless of where your platform sits on the maturity curve:

1. Confirm your embedded payments data is flowing into your own data warehouse (Chapter 8, Step 2). Without this, every AI and agentic use case in this chapter is blocked.

2. Identify the one AI-inside-the-rails capability you are not yet capturing — fraud, routing, underwriting, or personalization — and assign an owner. Even one of the four moved at your specific GMV usually justifies six months of focused work.

3. Audit one merchant workflow against the three agent-readiness tests in the Charge Forward Insight above. If you fail all three, you are not alone — but you are also not on the path to being agent-transactable.

To benchmark your platform against the framework introduced in this chapter, take the Charge Forward Embedded Payments Fit Assessment. To identify your current stage and the right next move, download the Embedded Payments Maturity Framework. For the full benchmark dataset referenced throughout this guide, see Chapter 11. To model the revenue impact of any of the AI and agentic use cases discussed above, run the Charge Forward Payments Revenue Calculator.

SOURCES & REFERENCES

Visa — Intelligent Commerce launch and developer documentation (April 2025); Mastercard — Agent Pay launch and partner announcements (April 2025); Stripe — Agent Toolkit and OpenAI Instant Checkout (2025); Google — Agent Payments Protocol (AP2) specification and partner announcements (September 2025).

Affirm — AdaptAI launch (2025); Sezzle — “Prophet” credit-model disclosures; Shopify — Shopify Capital data-model disclosures; Square Loans — Block annual reports.

Stripe Radar, Adaptive Acceptance, and Adyen RevenueAccelerate technical documentation; Visa Advanced Authorization and Visa Token Service documentation; Mastercard Digital Enablement Service (MDES) documentation.

Charge Forward research synthesis: “AI’s Impact on Embedded Finance: A Strategic Analysis” (Charge Forward Research Backup, 2025); Adyen / BCG — Embedded Finance Report (2024); Flagship Advisory Partners — The Massive Embedded Finance Opportunity (2025).

Public Charge Forward tools referenced in this chapter: Embedded Payments Fit Assessment, Embedded Payments Maturity Framework, Payments Revenue Calculator, Vendor Database. All available at chargeforward.io/tools.

By Jane Podbelskaya · Updated June 20, 2026