Why payments needs applied AI

In part one of this series, I made the case that payments complexity now exceeds what any vendor, SI or bank can deliver on time, in budget, defect free and to spec and that nobody is incentivised to say so. Part two explained the way out: via applied payments AI that change the economic model. This third (and final) part is about what “applied AI” means. 

AI has already changed the world. I will leave others to have the (very serious and needed) discussion on how to ensure it doesn’t end the world. Like every disruptive technology, AI might, at first, appear to fall short of the hype. But I have no doubt it will exceed anything we can imagine, as it is applied and used for good.  

“Applied” is the key word. Models need to be trained on input from experts, not on the general internet. Hierarchies of skills and tools built on years of experience and orchestrated by AI will be as good as the experts. Code that is generated from the words of experts, modelled on code templates from experts, and then reviewed by other agents using the knowledge base of experts. That is what will transform the quality, time and cost to build. AI is great at coding. Experts know payments. One without the other is a losing proposition. 

We have seen this curve before. STP (straight-through processing) went from 0% to 99%+ for the newest instant rails and remade the economics of payments processing on the way. “Built by AI” will follow the same curve. In 2025, AI delivered little more than 0% of an end-to-end payments modernisation upgrade. 2026 will see that rise to single digits. To deliver the promise of AI, we need key technical elements to surround the general models. 

The harness for AI’s power

Not all skills and tools are created equal. Over time, models will learn how to improve their own skills and develop their own tools. Today, they are only as good as the experts who build them. 

The best frontier models – Claude, Gemini or ChatGPT – are limited by the skills and tools that orchestrate them. On their own, they simply do not know enough about payments, specific systems, or the realities of payments complexity. 

Wherever AI is used in complex ecosystems (pharma, intelligence, weather, payments) the harness that directs the AI’s power is essential. Even though the best frontier models achieve near-100% success on tasks that take a skilled human under four minutes, they succeed less than 10% of the time on tasks taking more than about four hours.  

This demonstrates you can’t use general AI to manage the complexity of payments, never mind a whole payments modernization flow. For this, you require true, deep payments expertise built into harness to sustain the agents and direct the models.  

This payments expertise needs to be layered into every skill and tool built to access the OPFs, GPPs and IPFs of this world. No frontier model in the world today, or coming tomorrow, can direct its power without a whole applied payments framework to guide and implement. This, collectively, is what a payments expert agent is.

What a payments skill looks like

AI knows a pacs.008 is not a pacs.009. AI knows the scheme rulebooks, and that they change every cycle. But a skill built by a payments expert knows how a specific platform, or rails platform, needs to actually configure a flow. That knowledge does not come from the internet. It comes from experts who have done it.

These skills and tools provide immediate output to analysts, developers and testers. They are the low-hanging fruit. Meticulously built, structured and cross-related, they allow the quick wins that will lift payments modernisation from single digits to 25-35%. There are real wins to be had here. But skills and tools alone top out there. To follow the STP revolution all the way, where AI provides 70%, 80% (and one day 95%) of the rulebook updates and builds new payment flows, we’ll need more sophisticated technology.

SLMs are the special advisers

Small, fine-tuned models can be trained specifically on payments reasoning and expert judgement. When paired with a large frontier model, they act as a specialist adviser. The large language model (LLM) still does the general reasoning, but the small model steers it to think and act like a payments expert.

To use AI speak (actually, some of us used short sentences before Claude)…The division of labour matters. The large model orchestrates. The small models specialise. Reasoning and thinking through a scheme change like a payments expert. Teasing out the issues. Looking around the corner.

We can now capture how an expert actually thinks through a problem – understanding their reasoning and instinct for what the check next – and turn that into structured knowledge. Reasoning algorithms process that knowledge, then store it in a small model that sits alongside the large one. This isn’t theoretical, it’s possible today.

Two years ago we started taking RedCompass Labs experts out of consulting and instead developed methods to capture their knowledge, reasoning and experience. So instead of helping one client at a time, they put their time into sharing their expertise with hundreds.

Workshop agents and knowledge bases

We know from our chat log analysis that even with the latest RAG (a common knowledge retrieval method), analysts, developers and testers only access a very narrow percentage of the payments knowledge contained in rulebooks, system documentation and project documentation.

Retrieval finds what you request. But it won’t surface what you didn’t ask for.

That tells us two things. First, the data needs deep enrichment, so edge cases are drawn out early. Second, the edge cases live in the 95% of knowledge that goes unaccessed. So how do you make sure that, when relevant, the 95% is still fed into the process?

This requires real-time workshop analysis. Agents that sit in the workshop. That listen the way a senior BA (business analyst) listens. That ask the clarifying questions before the requirement leaves the room. That hold the spec in mind and carry it, intact, into the documentation, the code and the tests. Using SLM reasoning, and accessing knowledge bases and priority knowledge stores (workshops from previous projects, knowledge from SMEs etc).

This is what kills the game of telephone from part one. The requirement never gets retold, because the same intelligence that heard it writes the code and the tests.

The effort to build this is what takes Built-by-AI rates up to 66-80%. And it has a benefit beyond speed: retirement, job changes and new hires no longer bring risk.

Payments flows built by AI

None of this is exotic. Every piece exists today. Frontier models. Fine-tuned SLMs. Expertly built skills. Enriched knowledge. Workshop agents. What is scarce is the payments expertise to build the harness.

Which is where this series started. Nobody wants to say payments is too complex or too hard. It turns out complexity was never the enemy. Pretending it was simple was the problem.

Take one flow. Workshop it. Document it. Code it. Test it. Maintain it. Upgrade it. And measure what percentage was built by AI. That number is about to become the most important KPI in payments modernisation.

Want to know more about Payments Expert Agent?

We’ve built an AI platform that brings a new approach to payments modernization. Payments Expert Agent combines our deep payments knowledge with intelligent technology to deliver change safely and at scale for global banks.

Find out more here.