AI in payments: what’s live, what’s a pilot and what’s still a pitch deck

A plain look at where artificial intelligence is genuinely deployed in payments, where it is being trialled and where the claims run ahead of reality.

Artificial intelligence (AI) is already in use across payments, but not in the places the headlines suggest. It is live in fraud detection, transaction monitoring, back-office automation and customer support. It is being trialled for onboarding, cash forecasting and agentic checkout, and one of the most visible trials has already been pulled back. Fully autonomous payment agents, sometimes called agentic payments, meaning AI that decides and pays without a person involved, remain mostly presentations.

The gap between the two is wider than most coverage admits. More than three-quarters of organisations in the United States experienced attempted or actual payments fraud in 2025, yet just 17% use AI to combat it, according to the Association for Financial Professionals 2026 Payments Fraud and Control Survey. That is not a technology that has swept the industry. It is a technology most teams are still deciding how to use.

Here is our read on where the line sits.

In this article, we cover:

1. What is AI in payments used for today?
2. Which AI payment use cases are still pilots?
3. What happened when agentic checkout went live?
4. Which AI payment ideas are still just a pitch deck?
5. Why is AI adoption in payments slower than the headlines suggest?
6. What does AI mean for cross-border payments?
7. How should a payments team choose AI use cases?

What is AI in payments used for today?

Today, AI in payments is used mainly for risk work: spotting fraud, screening transactions and cutting false alerts. It is also widely used in back-office automation and customer support. These are the uses running for real customers, on real transactions, where the results affect real decisions.

Fraud and financial crime are where AI has been earning its keep for years, long before the current wave of AI tools arrived. In research from American Banker, fraud detection and mitigation was the AI use case bankers ranked highest for impact, named by 53% of respondents. Back-office automation and customer service were tied on 39%, followed by risk management and compliance monitoring on 30%.

None of that is glamorous, and all of it is real. Machine learning has been scoring transactions, matching names against sanctions lists and cutting false alerts in monitoring queues for a long time. Inpay works with the AI-driven vendor ComplyAdvantage on transaction screening and monitoring, which we set out in more detail in how we manage compliance.

The other place AI is genuinely live is internal. Handling documents, working through exception queues, chasing payments that do not match up, answering first-line support questions and writing code are all being automated at scale. The cost of a mistake is limited and it is easier to show your working afterwards.

Which AI payment use cases are still pilots?

A pilot is a real deployment with a narrow scope, close supervision and no commitment to scale. Most of the agentic work sits here: Know Your Customer (KYC) and Know Your Business (KYB) onboarding checks, help with credit decisions, forecasting cash positions and handling disputes.

The number worth watching is how many pilots make it into everyday use. In a Wolf & Company survey of United States community banks, 40% had pilots under way but only 25% had moved a small trial into everyday use, and only 5% had scaled it successfully. That is a small sample of smaller banks, so treat it as a signal rather than a hard finding, but it matches what most people in the industry will recognise.

The blocker is rarely the AI itself. It is messy data, old systems that are hard to connect and the governance work needed before an automated decision can touch real money.

What happened when agentic checkout went live?

The clearest test so far ran in public and did not go to plan.

OpenAI launched Instant Checkout in ChatGPT on 29 September 2025. It supported single-item purchases from Etsy sellers in the United States, with over a million Shopify merchants promised soon after. A month later PayPal announced it would adopt the Agentic Commerce Protocol, the open standard behind the feature, put its wallet inside Instant Checkout and bring its merchant network into ChatGPT through 2026.

By March 2026 OpenAI had pulled it back. The company said the first version had not been flexible enough, so merchants would use their own checkout while OpenAI concentrated on helping people find products instead. Reporting at the time pointed to difficulty onboarding merchants, showing accurate product data and supporting multi-item baskets or loyalty schemes.

The lesson is not that agentic commerce failed. The Agentic Commerce Protocol is still in use even though the feature that launched it is not. People proved happy to research in a chat window and unwilling to pay in one, and keeping prices, stock and delivery options in step across millions of merchants in real time turned out to be the hard part. Anyone who has integrated a payment method will recognise the shape of that. The demo is the easy bit.

It is also a useful correction to the table above. Agentic checkout is not a slide. It shipped, ran for about five months at scale and then retreated. Expect more of that pattern before anything settles.

Which AI payment ideas are still just a pitch deck?

AI running a whole payment from start to finish, with no person involved, is not here yet. Neither is AI negotiating exchange rates, nor AI agents buying from each other at scale across borders.

That is not a criticism of the ambition, and it is not the same as saying nothing has shipped. Plenty has, as the Instant Checkout story shows. What has not arrived is genuine autonomy, along with the standards and liability rules that would make it safe. The mood in the industry matches. In the same American Banker research, agentic payments was named by 13% of bankers among the technologies poised to change payments, well behind stablecoins on 28% and real-time payments on 24%.

The sensible position is neither to dismiss it nor to commit early. For now, the focus should be on watching and understanding it.

Why is AI adoption in payments slower than the headlines suggest?

Three reasons, from our perspective:

Adding AI to an old process changes little. The Grant Thornton 2026 Banking AI Impact Survey found that when banks layer AI on top of existing onboarding, transaction monitoring or fraud workflows, they see small returns. The ones seeing real change redesigned the process around automation, which means changing how the work is done rather than buying a tool.

Regulated payments have to be explainable. Catching more fraud is only useful if you can explain a single decision afterwards, to a regulator or to the customer it affected. That requirement shapes what can ship, and it is why risk teams move slowly on purpose. It is also why the groundwork matters as much as the technology, something we learned when building our risk function.

The data is not ready. Payment data is often spread across systems and poorly organised, which limits what any AI can do with it. Newer standards that carry more detail with each payment will help, but that is a multi-year fix rather than a switch.

What does AI mean for cross-border payments?

Cross-border is where AI has the most to offer and the least room for error. Screening, sanctions checks, exception handling, payment repair, meaning fixing the details so a stuck payment can go through, and reconciliation are all high-volume tasks that need judgement. They suit automation well, and they are the wins available now.

What AI cannot do is make a slow rail fast. A model that decides in milliseconds still waits on settlement that takes days, a mismatch we touched on in our review of 2025. The same logic applies to the long shift away from card-based flows and toward account-to-account payments. Smart software sits on top of the rails. It does not replace them.

For financial institutions in particular, the question is not whether to adopt AI. It is which part of the process is actually holding things up.

How should a payments team choose AI use cases?

A short test we find useful:

1. Is the task high-volume and does it need judgement? If yes, it is a candidate. If it is rare and serious, keep a person on it.
2. Can you explain a single decision afterwards? If not, you will struggle to defend it to a regulator.
3. Would you have to redesign the process, or just bolt something on? Bolting on gives you a demo. Redesigning gives you a result.
4. Is the delay actually in the decision? If the hold-up is in settlement, no amount of AI will fix it.
5. What breaks if it gets it wrong? Start where the damage is limited.

The honest summary: AI in payments is real, useful and further along than many people think in risk and operations, and further behind than the announcements suggest everywhere else. Both things are true at once.

Frequently asked questions

Is AI actually used in payments today, or is it mostly hype?

Both. It is genuinely in use for fraud detection, transaction and sanctions screening, back-office automation and customer support. Claims about AI making payments on its own at scale are ahead of what has actually been built.

What is the most common AI use case in payments?

Fraud detection and mitigation, named by 53% of bankers in American Banker research as the highest-impact use case, ahead of back-office automation and customer service on 39%.

Has anyone actually launched agentic payments?

Yes. OpenAI launched Instant Checkout in ChatGPT in September 2025 and PayPal signed up to support it a month later. OpenAI scaled the in-chat version back in March 2026 and moved buying to merchants’ own sites. The Agentic Commerce Protocol behind it is still in use.

Why do so many AI pilots fail to reach everyday use?

Usually messy data, old systems that are hard to connect and governance, rather than how good the technology is. Adding AI to an unchanged process also tends to produce small returns, which makes the business case hard to defend.

Can AI make cross-border payments faster?

It can speed up the decisions around a payment, including screening, exception handling and repair. It cannot speed up settlement on a rail that is inherently slow.

Should we wait for the technology to mature?

Not for the risk and operations uses, which are proven. For agentic payments, watching closely and staying flexible is a reasonable position while the standards settle.