Since the spring I have built a set of small tools, each aimed at one job in payments or fintech where AI might do something useful. Some I use in client work to show what can be done. A few may become products. One is already live, and one I stopped.
I build them to find out which uses of AI hold up, and which only look good in a demo. The answer is usually narrower than the pitch.
Built to answer a question from client work, or to find out whether something can be done at all. Used in engagements and demos.
There is a clear user and a problem they would pay to solve. Being tested with a small number of them.
Live, with its own site and customers. Most tools will not get here, and that is fine.
Tested and set aside. Kept on the page because the reason it stopped is usually the most useful part.
Built from nothing to production this year. Listed on the FreeAgent integrations directory, with Xero, Stripe, HubSpot and Companies House also connected.
Likely to graduate as a service rather than as software. It is how a fixed-price payments cost review gets done in weeks rather than months, independent and with no commission on any switch.
The design question is whether the team's behaviour changes. A dashboard nobody acts on is the usual end for this kind of tool, so the loop from signal to action to result is built in.
The first worked example is predictive netting for FX. It sits behind the AI Economics engagement.
[Confirm: two sentences on what it does and what it showed]
[Confirm: two sentences on what it does and what it showed]
Runs a defensibility assessment from several points of view, product, sales, finance, an investor, and shows where they disagree. Exports the result for a board or investment committee pack.
A scoring framework for payments APIs, applied the same way across vendors so the scores can be compared.
[Confirm: two sentences on what it does and what it showed]
The prediction worked well enough. The market did not. At the end of the market I could reach, there were too few businesses with enough invoices to make it worth paying for, and the ones that had the problem were already stuck for reasons a prediction does not fix.
The tools feed the engagements. Stratum sits behind AI Economics, and the card fee analyser is how a payments cost review gets done quickly. When a client needs a model or a prototype, I build it rather than commission one.
Building them is also why I can say, with some confidence, what these systems do and do not do well. The inference bills, the data joins and the model choices in a review are ones I have dealt with myself.
Most of these are easier to understand in fifteen minutes on a call than on a page. Email me and say which one.