Picture this: your refrigerator realises you're almost out of milk and quietly schedules a restock from your preferred grocer. Meanwhile, your electric car detects it won't have enough charge for tomorrow's drive, books a nearby fast-charging slot, and locks in a favourable rate, all before you've even woken up. No approvals, no passwords, no you.
By the time you start your day, your AI assistant has already handled errands you didn’t know were waiting. But these scenarios can obscure how limited today’s systems actually are. Most mainstream ‘smart’ features still rely on simple, predefined rules, and in most cases, payments are still initiated by people, not machines. This is the gap between expectation and what’s real today.
Agentic commerce promises a world where AI doesn’t just recommend actions but executes them within trusted boundaries. However, today’s AI systems lack the reliability, oversight, and compliance foundations required to make that shift safely. Even so, momentum is building. McKinsey estimates that agent-driven commerce could generate between $3 trillion and $5 trillion globally by 2030, fundamentally reshaping how value moves between consumers, businesses, and machines.
The question now is: what must evolve before AI can start paying on our behalf?
Agentic commerce represents a shift from automated tasks to autonomous decision-making. Instead of reacting to simple triggers, agents will interpret intent, weigh trade-offs, and act on behalf of individuals or businesses. This goes far beyond embedding a payment screen inside a chatbot or giving an interface a conversational layer.
“As it stands, LLMs don’t maintain the level of data oversight needed for payment execution,” says Aissam Errami, Chief Product Officer at XPP. “They lose context and struggle with complex, multistep validation.”
These limitations clarify why today’s AI-assisted checkout experiences remain fundamentally human-driven. Agentic commerce isn’t a UI enhancement; it’s a fundamental redesign of how transactions are initiated and authorised.
“Examples like shopping through ChatGPT or ordering through Alexa may feel autonomous, but they’re not AI-driven payments,” says Jesse Stolwijk, Chief Platform Architect at XPP. “The user is still the one making the decision; AI is only assisting the interaction.”
For AI agents to initiate payments safely and autonomously, several foundational capabilities must develop together. This includes:
None of these elements exist in a fully mature form for large-scale, autonomous payments yet, but the industry is moving steadily in that direction. The leap to agentic commerce will only be possible once all five advance together.
One of the biggest misconceptions is that existing payment infrastructure is ready for AI-driven transactions. In reality, it was built around human behaviour and human-paced decision-making.
And while today’s agents are not yet operating at microsecond speeds and typically run inside large datacenters, they are still forced to work through workflows designed for humans, which slows them far more than the agents’ own processing speed. As agents become faster and more autonomous, these structural bottlenecks will become even more pronounced.
Card networks use multi-second authorisation cycles, while AI agents operate in milliseconds. A single agentic workflow, like booking flights, routing shipments, or adjusting budgets, may involve dozens of tightly sequenced micro-decisions. An agent may need to run dozens of checks in under two seconds. Multi-second authorisations don’t just introduce delay; they break the sequence entirely.
Risk systems add another layer of friction. Fraud models today are trained on human purchasing patterns: infrequent, predictable, and relatively stable. AI agents behave differently by design. They may transact rapidly, shift contexts midstream, and interact across regions as part of normal operation. Legacy systems interpret this as suspicious, creating false declines and unnecessary revenue loss.
Most consumer payment flows are still built on the assumption that a human will confirm the transaction. That assumption introduces friction: passwords, one-time codes, approval prompts, all of which are fundamentally incompatible with autonomous decision-making.
“In autonomous systems, friction becomes as damaging as fraud, not because it takes funds, but because it breaks momentum,” says Errami.
For AI agents to operate independently, payments must shift from verifying human presence to verifying agent permissions. Instead of asking, “Is the user here?”, the system must be able to answer, “Is this agent authorised to take this action?”
Achieving this requires a policy-based trust model that embeds governance directly into the transaction. The result is a trust framework built for machines: continuous, embedded, and fast enough to support autonomous decision loops without sacrificing oversight or accountability.
Even with the right infrastructure, one challenge sits above all others: liability. When an AI agent overspends, selects the wrong vendor, or acts outside expectations, determining responsibility becomes far from straightforward. Today’s dispute systems are built on the premise that a human initiated the transaction. Agentic commerce dismantles that assumption entirely.
If an agent books the wrong flight, who pays? Is it the user who set broad parameters, the platform that hosted the agent, or the AI provider whose model made the decision?
“Autonomous decisions divide responsibility across multiple parties,” says Errami. “Users may set the boundaries, but the liability frameworks around those boundaries are still catching up.”
Until clear liability standards emerge, AI’s role will remain strongest in augmentation rather than execution. It can identify anomalies, detect fraud, analyse behaviour, and strengthen risk decisions without taking final action.
“Fully autonomous AI-initiated payments aren’t feasible at scale today, but AI-enhanced payment intelligence absolutely is, and that’s where businesses can benefit immediately,” says Stolwijk.
Today’s payment infrastructure isn’t built for the speed, scale, or decision logic that agentic commerce will require. However, the shift is inevitable, and the organisations preparing now will define what comes next.
Ginger, XPP’s payment gateway solution for e-commerce, has already laid the groundwork for that future. It delivers the core foundations for safe, policy-driven autonomy, including tokenised payments, single-tenant architecture, real-time data access, and an API-first design for seamless system-to-system flows. Together, these capabilities form the trust layer that supports AI-augmented operations today and prepares organisations for fully autonomous transactions tomorrow.
The future of autonomous payments won’t be defined by AI, it will be defined by the infrastructure beneath it.
.png)
.png)
