Payments is too complex. We need a new model

Last week, in part one of this series, I made the case that payments has hit a strategic inflection point. The complexity of payments now exceeds what any vendor, SI, or users of payment flows / banks can deliver on time, in budget, defect-free, and to spec. Nobody is incentivized to say so. There is technology that could change it, but you can’t bet it all on something that’s not fully proven. And doing nothing means slowly drowning in the mess we already have. 

That’s the diagnosis. This is the cure. 

If we are at an inflection point, then getting through it will take more than a quick deck chair shuffle on the Titanic. So, what has to change?  

There are several approaches. But they all rely on moving far beyond whatever Claude, OpenAI, or Gemini alone can provide, however powerful their generic models are.  

AI promises to lower costs for clients and give higher margins to payment platforms or service providers. It’s not that straightforward. AI can’t just be pasted onto your current payments provider (bank or payments company), vendor (in-house or black box platform), large systems integrator’s (SI) broken model, or the Software Development Lifecycle (SDLC) models of today.  

Ok, so how can we turn this promise of AI into reality? 

Changing the model

We see three options:

1. Vendors

Share prices are down because AI is a threat to software companies. It enables clients to build and capture the value these vendors have always provided.

The sprint to out-of-the-box platforms was about the pain and cost of managing complexity. Vendors tried to solve every client’s needs on unfinished platforms – which is why professional services teams got bloated, and most implementations were verging on bespoke.

Now, the more vanilla your payments engine, the easier it is for a client to replace what you do with an in-house, AI-built one.

So vanilla won’t work. But you can still own the relationship. You can provide your expertise. NVIDIA doesn’t make the chips. TSMC does. TSMC doesn’t make the machines that make the chips. ASML does. You don’t need to do it all to win. You can be Apple. You can be Nike.

All you need to do is fulfill the original brief: deliver a working payments engine that fits a client’s needs and keep it compliant and fit for purpose.

The old world of the license vs. services ratio for quarterly earnings needs to change. If AI is eating the software companies’ lunch, they need to fight back. It’s vital that they combine the best of software (repeatable income) with the best of services (bespoke solutions). It’s the same with payment platforms such as OPF, GPP, or ACI. They need to use AI to quickly design, capture, code, and test, and then maintain and upgrade client flows.

Like industrial 3D printers that no longer need die-casting lines, payment platforms need to use applied payments AI to design and print the flow and provide the platform as a service.

Rolls-Royce used to only make jet engines. Now, 69% of their revenue comes from servicing them. They monitor, plan maintenance and keep the engines spinning, selling an ongoing service not just the one-off product. Banks will do the same with payment flows: happily outsource the design, build, and upkeep, as long as it stays compliant, fits perfectly into the bank, and they can trust it will be done right.

Vendors’ revenue won’t come from treating customers like hostages – it will hinge on how well you deliver, and how well you use the tools. TSMC makes more cutting-edge chips than AMSL’s other customers because it focuses relentlessly on client needs. The solution is to earn the client every day. “There is real gold in thar hills!”

The dream of vanilla software license money (build it once and sell it many times) is over. AI has crushed that. Applied payments AI that understands payment complexity lets you build, maintain, and sell a platform customized to the client, capturing more of the value chain.

Fable, or Opus, or GPT-6 can’t do that. But it can be done with deeply applied payments AI models, payment expert agent tools, expert skills, and access to the deepest knowledge. Where workshops are captured and turned into specification-driven development cycles, testing runs continuously and human experts remain to sign off and develop edge-case analysis.

2. Payment flow providers/banks

Go back to how this started. If you don’t want to outsource complex payment flow management, then build your own flows. This doesn’t mean dabbling in Teams recordings, Copilot, or generic OpenAI or Anthropic models.

To build payment flows, you will need AI infrastructure that thinks and reasons like a payments expert. It also needs the knowledge of every scheme in the world, and all your in-house knowledge and external expert knowledge. This requires large language models (LLMs) to orchestrate, but smaller language models (SLMs) for specific tasks like reasoning and edge-case tools.

This means you’ll be able to go from people in a room talking to code that’s ready to test in hours. No single generic model gets you there on its own. But you can with a relatively simple assembly of:

• An LLM
• Payment agents
• A dedicated payment SLM
• Platform tools
• Expertly built skills
• Reference to all your payments documentation, enriched with expert interpretation and analysis

3. System integrators

Use the same payments AI technology as vendors. Master it. Use it across 10 or 20 banks. Prove you’re the TSMC (i.e., you can get more out of the applied payments AI than others). Build with it so many times, in so many ways, that you become the expert. Use the applied payments AI to build flows for vendors and banks using the same tools they will need. Maintain them.

You can own solutions. You can manage them, just not with the processes and armies you have today. You can’t win with your AI that covers everything broadly and nothing deeply.

It doesn’t matter how many airport ads you pay for or athletes you sponsor. To win, you must change the model. That new model should be deeply vertical AI that works, thinks, and delivers like the best 1% of payment experts in the world. But with the right tools, your brand can win the contracts that allow payment platforms to be implemented and bought.

Impressive claims on expensive billboards can be proven true. And you can win. Keeping the status quo, with general AI fluffed up by inflated success stories, could buy you an extra year or two, but it can’t last forever. No amount of lipstick will cover up what you have and what you deliver. AI is changing that.

Why now?

Remember why separate flows died the first time around. It wasn’t the architecture. An AS400 flow was magnificent at running reliably. It died on cost. Every change needed rare specialists, long timelines, and painful money. The knowledge lived in a few heads and walked out the door when they did. And testing every change was slow and terrifying. Platforms won on economics, not because one giant interconnected system was ever a good idea.

Applied payments AI with access to the best tools, fine-tuned payments SLMs, skills, experts, and knowledge changes those economics. It can compress the SDLC and remake it with speed and feedback loops that will make the very idea of agile seem slow. Two weeks for a sprint? This technology reduces it to two hours. It can compress the timeline and stop the game of telephone.

Payments AI has a payments memory. Every workshop, meeting, defect triage call, every decision point, every remark an SME makes in a room can be remembered forever and pulled into every project that follows. Knowledge flight risk? gone. . Reinventing the wheel every time consultants rotate? Over.

Payments AI is a developer that’s also your in-house SME. It’s an industry expert. Oh, and it can code too. And thanks to the efforts in the wider AI industry, it codes very well when working with the applied payments AI for specifics. And it will only get better: the cost of generating and changing a flow is falling every few weeks.

Payments AI can test. Over and over. 24/7/365. It can find edge cases before your client does.

Knowledge, change, and verification. Those are the three costs that killed the separate-flow model. Payments AI, with a few experts, attacks all three. This means individual flows on common services stop being nostalgic. For the first time since the mainframe era, inside a new-style payments enterprise platform product/service, they become the economically rational answer.

One ingredient doesn’t come from this machine: payments expertise. Someone has to evaluate what AI decides, feed what AI analyses, and steer what AI codes. The expertise isn’t replaced. It becomes the scarcest thing in the room. But what changes is where you once needed 10 experts, now you only need 1 or 2.

No other option

We didn’t leave the mainframes because we wanted to. We left because we couldn’t afford to stay. The same logic is now running in reverse: we can’t afford the complexity we’ve built, and for the first time, there’s a credible way out.

Take one flow and use payments AI agents to workshop it. Document it. Code it. Test it. Upgrade it. Maintain it.

We’re still those polar bears drifting out to sea on melting icebergs I mentioned in part one. There’s really only one decision left. Do you jump to applied payments AI and build a new economic model, or stay on the ice and go down holding on?

In Part 3 we will dig deeper into the specifics and technicality of the Applied Payments AI architecture. Looking at the limits of generic AI, the power of fine-tuned SLMs, how all skills are not equal, the need for platform- and payment-specific tools, as well as what it means to capture knowledge (both industry and expert). And how this can all work together to support the workshop, documentation, coding, testing, maintenance, and upgrading.