I Almost Added AI to Our Forex Platform Just Because Everyone Else Was Doing It

The pressure to add AI is real, and it's mostly stupid

When I was at Finvestfx managing our treasury and forex SaaS, I started getting the same question in every quarterly business review: "What's your AI roadmap?"

This was late 2023. ChatGPT was everywhere. Every competitor deck suddenly had "AI-powered insights" plastered on it. Our enterprise clients, CFOs at mid-sized firms, were getting pitched AI solutions left and right.

And I kept saying no.

Not because I don't believe in AI. I spent the last year at Sonic Linker building an AI-first product from scratch. I know what good AI implementation looks like. That's exactly why I said no at Finvestfx.

Here's the thing: our clients didn't need AI. They needed their forex data to be accurate and accessible faster. They needed reconciliation workflows that didn't break when their bank changed an API format. They needed reliable email alerts when exposure limits were hit.

Adding "AI-powered forecasting" would have been a distraction wrapped in a buzzword.

The three questions I ask before considering AI

1. Is the current solution actually broken?

At Finvestfx, our reporting suite worked fine. Treasury managers could pull the data they needed in under 30 seconds. Yes, we could have added predictive analytics or natural language queries, but the existing SQL-based filters were fast and people knew how to use them.

Compare that to Sonic Linker, where we were solving a fundamentally new problem. Linking fragmented data across platforms at scale, without humans doing manual matching. That workflow didn't exist before. AI wasn't an enhancement there, it was the entire solution.

If people are working around your current feature, fix the feature. If they're just not using it, ask why. But if it works and they're happy, don't add AI just to add AI.

2. Will AI actually reduce friction, or just add complexity?

I learned this the hard way. At one point, I prototyped an "AI assistant" for Finvestfx that could answer treasury policy questions. Sounds useful, right?

Turns out, our users wanted definitive answers backed by their company's actual policy docs, not probabilistic responses from a language model. They needed to show audit trails. An LLM saying "here's what I think your hedging policy allows" wasn't just unhelpful, it was risky.

We scrapped it. Spent the eng time building a better search and tagging system for policy documents instead. Way less sexy. Infinitely more useful.

AI adds a layer. Every layer is friction unless it removes more friction than it creates.

3. Can you measure success without hand-waving?

This is where most AI features die quietly. Someone adds a "smart recommendation engine" but can't define what success looks like beyond "users engage with it."

At Sonic Linker, we knew exactly what success meant: reduction in time spent on manual data matching, measured in hours per week per user. Accuracy rate of matches, measured against a labeled test set. Those were hard numbers tied to the core value prop.

If you can't define the success metric before you build the AI feature, you're not ready to build it. And "looks innovative to clients" is not a success metric, it's a marketing pitch.

When I finally said yes

Six months after those first AI questions at Finvestfx, we did add predictive cash flow modeling. Not because clients asked for AI, but because three separate CFOs told us they were spending 4-5 hours a week manually forecasting currency needs based on invoice schedules and historical patterns.

That's a real problem. Repetitive, time-consuming, pattern-based. Perfect for ML.

We built it, and it reduced forecasting time by about 60% for the clients who used it. Not because it was AI, but because it solved a specific, measurable pain point that we'd validated with actual users.

The real test

Here's my rule now: if I can't explain the AI feature's value without using the word "AI," I don't build it.

"This feature saves you 4 hours a week on cash flow forecasting" passes the test. "This feature uses AI to optimize your treasury operations" does not.

AI is a tool. Sometimes it's the right tool. Most of the time, especially if you're adding it because competitors are or because it sounds good in a pitch deck, it's not.

I've shipped AI features that genuinely transformed products. I've also killed AI features that would have been pure theater. The difference wasn't the technology. It was whether the problem actually needed that solution.

If you're adding AI to check a box, you're solving the wrong problem.