AI does not have to burn a coal plant to answer a support email. Somewhere right now, a data centre is consuming enough power to light a small town, running a 175-billion-parameter model to do something a 9-billion-parameter model could do in your office, on a single GPU, without a single byte leaving your site. The AI energy debate is real. The answer is not to slow adoption. It is to stop deploying AI lazily.
A real problem, and the lazy default
A scientific nanotechnology equipment manufacturer came to us with a genuine problem. Customer support cases were taking up to four months to resolve. The MD was drowning in manual triage. Engineers were being pulled off R&D to answer emails. There were more than 1,800 historical cases with no pattern recognition across any of them.
The industry default would have been to connect a large cloud model, write a clever prompt, and call it AI-powered support. We did not. We started where every good engineer starts: with the problem itself.
Decompose first, then reach for a model
We mapped the entire support journey. Every step, every decision point, every handoff. Every output is validated before it becomes the next input, because noise in is still noise out, no matter how sophisticated the model. What we ended up with was a 90-step workflow covering the full support lifecycle, and AI is used at exactly six of those steps. Not because we could not use it more. Because we did not need to.
When you decompose a complex process with engineering discipline and validate each output before it feeds the next, the AI problem shrinks from something enormous and expensive into a series of small, precise, tractable tasks. And small, precise tasks do not need a data centre the size of a football stadium to solve them.
The result: roughly 100x less energy, as a consequence of engineering
The system runs on a single self-hosted GPU in a local office. No hyperscaler, no data centre on the other side of the globe, no aquifer. Energy consumption came out roughly a hundred times lower than the cloud API equivalent, not as a secondary benefit, but as a direct consequence of engineering the problem properly before touching a model. Email to structured ticket in under two minutes. Triage accuracy above 90%. Root-cause patterns surfaced across the entire installed base for the first time. This is the thinking productised in Cortex, the same engineering-first discipline we bring to every build. There is more on the hidden carbon cost of AI if you want the ESG angle.
The technology does not have to cost the earth. We just have to stop pretending that prompting a giant model is the same thing as engineering a solution. Decompose first. Validate every step. Use the smallest model that does the job properly.