AI FinOps & Unit Economics, Part 2
The SLM Right-Sizing Play: Stop Using Supercomputers to Categorise Emails
By: Rahul Kumar, Regional Director, Experis Europe
I was reviewing an enterprise architecture diagram with a lead engineer recently when I spotted an expensive habit:
Every inbound customer support email was being routed to an unconstrained, multi-hundred-billion-parameter foundation model simply to classify whether the sender had a billing question or a password reset issue.
They were spending 2.5p per classification across two million transactions a year. That is approximately £50,000 annually for a task that an open-weight, 3-billion-parameter Small Language Model (SLM) could handle on a sub-£200-per-month cloud instance, with near-zero latency.
That is the equivalent of dispatching a transatlantic jet to deliver a letter across town.
During the initial rush of the generative wave, enterprise teams standardised on frontier foundation models for convenience. One API endpoint solved every problem. But convenience at low volume becomes economic erosion at enterprise scale.
According to research from MIT Sloan Management Review, specialised small models running targeted domain workloads achieve comparable or superior task accuracy to massive general-purpose models, while slashing inference costs by up to 90%.
The goal is not to abandon foundation models. The goal is architectural discipline.
To break free from bloated inference billing, technology leaders must right-size their model portfolio:
Triage workloads by cognitive depth: Reserve frontier foundation models for ambiguous, multi-step logical reasoning and executive synthesis. Route repetitive classification, entity extraction, and sentiment routing to compact, dedicated SLMs.
Host domain-distilled models on private instances: Fine-tune compact 3B to 8B parameter models on your proprietary business records. Hosting them on standardised compute eliminates unpredictable per-token API spikes and locks in fixed operational costs.
Implement an intelligent model router: Build a policy-based gateway upstream of your inference layer. The gateway evaluates prompt complexity and automatically routes queries to the lowest-cost model capable of meeting accuracy thresholds.
At Experis, we help organisations decouple their technology stacks from runaway hyperscaler meter rates.
Our Project Services teams don’t just advise on theoretical cost reductions. We work side-by-side with your engineering teams to build practical, right-sized architectures focused on sustained cost efficiency. Supported by our cross-border delivery hubs, our squads refactor pipelines, distil domain models and deploy lean SLM architectures designed to optimise production compute spend without sacrificing accuracy.
Are you paying frontier prices for basic classification tasks, or has your architecture evolved to match workload reality?
Contact Experis to discuss how we can help optimise your AI architecture, reduce compute costs and improve performance.
References:
