Sustainability in AI Products: What We're Doing and Why It's Hard
TailorFit runs on large language models. Every tailored resume, every job description analysis, every chat with Tai is inference on a GPU in a data center somewhere, drawing power we don't meter and cooled by water we never see. An AI company that markets itself as thoughtful owes users a straight account of that, including the parts we can't measure. This is that account, current as of mid-2026.
What we actually do
We contribute 1% of revenue to carbon removal through Stripe Climate. Removal, not offsets: the money funds technologies that pull carbon out of the atmosphere and store it, rather than paying someone else not to emit. It's in our footer and it's real money off our top line. At our current size, the absolute dollars are small. It's a standing commitment that scales with us, not a claim to present-day impact.
The larger lever is engineering, because for an AI product, efficiency and footprint are the same variable. We route models by task: our assistant runs on a small, fast model, and the big model is reserved for heavyweight generation work where quality genuinely requires it. We don't send a frontier model to answer "what stage is my Google application in." We cache aggressively, keep prompts lean, and log token counts on every AI call. That started as cost observability, but tokens are the closest thing we have to an energy meter, so the same discipline serves both.
What we honestly don't know
We don't run the GPUs. Our inference happens inside our AI provider's infrastructure, which runs inside cloud data centers, which sit on regional grids with carbon intensities that vary by the hour. The energy cost of a single request depends on the model, the hardware batch it shared, the data center's cooling design, and what the local grid was burning that afternoon. We can't see most of that chain, and neither can almost any company our size.
Published per-query energy estimates for LLMs range across an order of magnitude depending on assumptions. Nobody outside the labs knows precisely, and the labs publish selectively. We won't put a false-precision number on our footprint, "each resume costs X grams of CO2," because we'd be making it up. What we can say honestly: our footprint is dominated by upstream inference we don't control, our best proxy for it is tokens consumed, and we actively minimize that proxy.
Why we bother saying any of this
The alternative postures are worse. Silence implies the question doesn't matter. A green badge with no math behind it is marketing wearing sustainability's clothes. The truthful position for a small AI company in 2026 is uncomfortable but simple: the tools we're built on have a real physical cost, the accounting for that cost is genuinely immature, and the responsible moves available to us are efficiency, funding removal, and honesty about the limits of our knowledge. We're doing those three.
What's next
Two commitments. First, we'll keep tightening the efficiency loop. Model routing, caching, and prompt discipline are living engineering work, not a one-time audit. Second, when we're big enough to justify it, we intend to commission a proper third-party measurement of our emissions including the upstream inference share, and publish the results and the method, including whatever turns out to be embarrassing. Small companies mostly can't know their AI footprint today. We'd like to help make that less true, starting with ours.