When a technical team reports "latency" to the boardroom, they almost always mistranslate it: they frame it as an infrastructure topic. It isn't. In a payment gateway, every extra millisecond is a sale at risk, a customer who hesitates, and — in the worst case — a double charge someone will have to reconcile. Latency is a P&L line item; the problem is almost nobody reads it that way.
Latency is measured in money
A payment that lags isn't "slow": it's friction at the exact moment the customer decides. At checkout, every point of abandonment is revenue that doesn't come back. And unlike a visible error, latency doesn't trigger a red alert: it bleeds quietly, in the conversion rate, until someone asks why sales dropped when "nothing" changed.
Where the P&L bleeds
The damage isn't a single number; it's a sum of leaks that rarely land in the same report:
- Checkout abandonment: the user who watches the spinner and leaves.
- Timeout → retry → double charge: a slow response that gets retried can charge twice. The cost isn't just the refund — it's the team's time reconciling and the customer's trust.
- SLA penalties: if you promised a partner a p99, missing it has a number in the contract.
- "Just in case" over-provisioning: buying hardware to paper over a design problem is CapEx that shouldn't exist.
p99, not the average
Here's the most expensive mistake: looking at the average. The average says "almost everything is fine" — and it's true, almost everything. But the money is lost in the worst 1% of transactions: the p99. A system with a 50ms average and 3 seconds at p99 doesn't have a small problem; it has one you only see once you've already lost the customer. The right question for your team isn't "how long does it take on average?", but "how long does the worst 1% take?".
Mitigation with a return
The good news: almost none of this is fixed with more servers. It's fixed with architecture decisions, most of them low-cost:
- Latency budget: deciding how much time each step may spend. It's a financial decision disguised as a technical one.
- Idempotency: making a retried operation not execute twice. Kills the double charge at the root.
- Circuit breakers: when something degrades, contain the bleeding instead of dragging down the whole system.
- p99 observability: you can't manage what you only look at as an average.
The CFO should watch the payments p99 the way they watch churn: a business metric that, ignored, collects itself. Latency isn't an infrastructure topic. It's a money topic solved with engineering.
