AI moved the fintech moat.

Code got cheap. Licences, rail connections and ledgers that balance did not. Here is where the advantage went.

Read
3 minutes
Last updated

For a decade, the hard part of fintech was the software. A team that could ship a wallet or a card app fast had an edge.

That edge is shrinking. The parts of a financial product that AI cannot write are now where products win or fail. Licences, rail connections and a ledger that balances are those parts.

Code got cheap

Running a GPT-3.5-class model fell from $20 to $0.07 per million tokens between November 2022 and October 2024. Stanford’s AI Index calls it a more than 280-fold drop.

Startups noticed. In early 2025, Y Combinator said about a quarter of its latest batch had codebases 95% written by AI.

The productivity evidence is real but uneven. In one controlled task, developers with an AI assistant finished 55.8% faster. A Google trial measured about 21%. A 2025 study found experienced open-source developers were 19% slower.

The direction is clear even where the size is not. Writing the app is no longer the scarce step.

What did not get cheap

The layerWhy AI does not shortcut it

The licence

A regulator decides who holds customer money. In the EU, the median e-money or payment licence takes 9.5 months to approve.

In the Philippines, new non-bank e-money licences were closed from December 2021 to December 2024.

The rails

InstaPay, PESONet, QR Ph and the card networks admit participants, not repositories.

Connectivity is contracted, certified and tested. It is not generated.

The ledger

Every balance must trace to entries that match the bank’s records.

A model can draft the schema. It cannot make last month’s books agree.

The operations

Onboarding review, disputes, monitoring and reporting to the regulator, every day.

These are obligations with named owners, not features.

The cost of getting the ledger wrong

In 2024, a US middleware company between fintech apps and their banks collapsed. The CFPB later alleged its records did not match the banks’, with a $60–90 million shortfall.

Consumers could not reach their money for weeks or months. The apps on top worked. The layer underneath did not.

That is the new failure mode. Shipping faster raises the stakes on the parts you cannot see. What a ledger must do

Where to spend your advantage

Spend your cheap code on the product your customers see. Buy the regulated layer from someone who already runs it.

That is the choice TNKR is built around. Run it on your licence if you hold one. Run it on ours if you don’t.

Either way, payouts, ledgers and onboarding run on rails PayMongo already operates. See Ledgers

Talk to an expert — tell us what you’re building, and we’ll tell you which layer to buy.

Frequently asked questions

It makes the software cheaper. Licences, rail access, correct ledgers and compliance operations still take time and capital.

Studies disagree. One controlled task ran 55.8% faster and a Google trial found 21%. A 2025 study found experienced developers 19% slower.

The regulated layer: the licence, the rail connections, a ledger that reconciles, and the operations that keep it examined.

Selectively. A moratorium on new non-bank EMIs ran from December 2021 until it was lifted in December 2024.

Originally published

Build on one ledger that already runs.