The token bill your P&L never sees — and why cross-border operators should care
Every cross-border seller I know runs a quiet second business: an AI-assisted content and ops factory. Listing copy in five languages, ad variations for TikTok Shop, supplier emails, review responses, Shopify product descriptions, Amazon A+ modules — most of it now passes through Claude Code, Codex, or Cursor at some point. And almost nobody I talk to can answer a simple question: what does that actually cost us per month, and what does it cost the planet? That’s the gap Unprompt is poking at, and while it’s a consumer Mac/iOS utility rather than an e-commerce tool, the underlying discipline — measuring the invisible line item — is exactly what separates operators who scale margins from operators who scale vibes.
What Unprompt actually does, and why the framing matters
The maker, Joëlle, built three apps solo and admits she “had no idea what all my AI tokens added up to. In dollars, or in water.” That’s the origin story, and it’s refreshingly honest. The product has two halves:
- On Mac: a free companion app reads the local logs of Claude Code, Codex, and Cursor, converts token counts into dollars and liters of water. Conversation content is never read. Dollars reflect API prices, not your subscription bill. Water is derived from electricity at roughly 4.9 liters per kWh (Berkeley Lab, 2024).
- On iPhone: a companion reads Screen Time to count how often you open AI apps, separating work hours from personal. It blocks nothing.
Her own 30-day number: 272M tokens. Today’s tokens and cost are free; the 30-day water card sits behind Premium. No account, no API keys, no proxy — totals sync through your own iCloud, and friend comparisons are opt-in with trends shown by default rather than exact numbers.
For a cross-border seller, the interesting part isn’t the water guilt-trip. It’s the metering reflex. If you’re running a DTC brand on Shopify with an Amazon FBA arm and a TikTok Shop storefront, you already obsess over CAC, ACOS, and 3PL fees. You probably don’t obsess over your AI token burn — and that’s a blind spot worth closing before it becomes a line item someone in finance asks about.
Why Amazon sellers should care more than Shopify ones
Shopify operators tend to be lean: one storefront, one theme, a handful of apps, a small content team. Amazon sellers live in a different cost universe. Between Seller Central listing refreshes, Helium 10 keyword research, A+ content localization, and the endless churn of review responses and Brand Registry takedown drafts, AI usage compounds quietly across a dozen workflows. Multiply that by five marketplaces and the token count stops being a rounding error. The same logic applies to Etsy sellers running long-tail listing variants and eBay sellers rewriting titles at scale — the more SKUs and locales you touch, the more your AI spend behaves like a variable cost, not a fixed subscription.
How it differs from the incumbents you’d actually compare it to
Here’s where I’d push back on treating Unprompt as a novelty. The obvious comparisons are:
- Native dashboards. OpenAI’s usage dashboard and Anthropic’s console show token spend for API users, but they don’t unify Claude Code, Codex, and Cursor, and they say nothing about energy. If you’re on a flat subscription, they’re nearly useless for budgeting.
- FinOps-style SaaS trackers. Tools in the SaaS spend management category track seat licenses, not per-token inference. They’ll tell you what you pay for Cursor, not what your Cursor usage costs in equivalent API terms.
- Carbon and energy calculators. Electricity Maps and similar tools give grid intensity, but they’re not tied to your actual token logs. Unprompt’s edge is that it sits on the local logs — no API keys, no proxy, per the maker’s reply to Leon Jakob Kastelic — which is a genuinely smart privacy posture for anyone handling supplier contracts or unpublished product copy.
The differentiation, in one sentence: Unprompt is the only tool I’ve seen that turns your local AI logs into a unified dollar-and-energy estimate without asking for credentials. That’s a narrow but real wedge.
Where the math breaks
Gal Dayan, who builds Dial, raised the sharpest critique on the launch thread: the 4.9 L/kWh figure is a fixed average, but real water use swings wildly by datacenter, cooling design, and season. Joëlle’s response is worth quoting because it reframes the claim honestly: the number is “mostly not cooling. It’s 0.36 L on site plus 4.52 L for generating the electricity,” and the app “presents water as an order of magnitude, not a measurement.” That’s the right framing — but it also means the water figure is directional, not auditable. If you’re the kind of operator who reconciles every 3PL invoice to the cent, treat the liters as a vibe check, not a KPI.
What cross-border sellers can borrow from this
Three transferable lessons, in order of how fast I’d act on them:
1. Meter your AI workflows like you meter ad spend. You don’t need Unprompt specifically, but you need something. If your content team burns tokens across Claude, ChatGPT, and Cursor, the equivalent-API-cost number is a useful benchmark: it tells you whether a $200/month seat is a bargain or whether you’re effectively spending $1,400 in API terms and should renegotiate. Pull the logs, do the math once a month, and put the number next to your Klaviyo bill.
2. Adopt the “no keys, no proxy” privacy pattern for your own tooling. If you’re evaluating any AI vendor that touches supplier pricing, unpublished listings, or customer PII, the question “does this read local data or require my credentials?” should be a filter, not an afterthought. Unprompt’s design — local logs, iCloud sync, opt-in sharing — is a template worth copying in your internal tooling reviews.
3. Treat energy as a procurement signal, not a moral one. European marketplaces and increasingly US retailers are pushing sustainability disclosures down the supply chain. Knowing your own AI footprint — even roughly — puts you ahead of the curve when a marketplace or a B2B buyer asks. It’s also a cheap differentiator in DTC brand storytelling if you can back it up.
The uncomfortable question for FBA brand owners
If you’re running a private-label brand with a five-person team, your AI token spend is probably smaller than your Fulfillment by Amazon storage fees. But your content volume is the thing that scales fastest as you add SKUs and locales. The operators I’d worry about are the ones running 200+ ASINs across Amazon, Temu, and SHEIN with a two-person content team and no metering at all. That’s where the invisible bill hides.
Where my judgment says it falls short
Three honest reservations:
- Platform coverage is narrow. Claude Code, Codex, and Cursor cover a slice of the developer-adjacent AI stack. A cross-border content team living in Jasper, Copy.ai, or Midjourney gets nothing from this. That’s a real ceiling for the tool’s usefulness to sellers specifically.
- The dollar figure is an API-price fiction. Joëlle is upfront that it’s “what your tokens would cost at API prices, not your subscription bill.” Useful for benchmarking, misleading if you paste it into a budget as actual spend. Don’t let anyone on your team confuse the two.
- The water number is a single point estimate. Dayan’s critique stands: a range would be more honest than one number presented with the same confidence as the dollar figure. The maker’s “order of magnitude” framing is correct, but the UI (as described) doesn’t obviously signal that asymmetry. If you’re going to show both numbers side by side, they should carry different confidence labels.
None of these are dealbreakers for a free Mac utility. They are reasons not to over-index on the outputs as accounting-grade data.
What I’d watch / test next
This week, three concrete moves. First, if you’re on a Mac and use any of the three supported tools, install the free companion and let it run for seven days — then compare the equivalent-API-cost number against your actual subscription invoices. The delta is your real “AI efficiency ratio,” and it’s the kind of number that should live in your monthly ops review next to Google Ads spend and 3PL fees. Second, if you’re not on Mac, do the poor man’s version: export your OpenAI and Anthropic usage CSVs, sum the tokens, and multiply by current API rates. It takes twenty minutes and will probably surprise you. Third, add one question to your vendor evaluation template: “Does this tool require my API keys or read my data remotely?” If the answer is yes and the vendor can’t explain why, that’s a flag — the local-first pattern Unprompt demonstrates is increasingly the standard sophisticated buyers should expect. Watch whether the maker ships a range-based water estimate and broader app coverage; both would move this from a curiosity to a genuine line item in the cross-border operator’s dashboard.
