Skip to content
All case studies

CASE STUDY / Banking intent classification

An 8B specialist. Better accuracy. A fraction of the cost.

Isoquant raised banking intent accuracy from 74.7% to 91.8%, while cutting inference cost by 91.7%.

91.7%lower inference cost
91.8%intent accuracy
36.8%lower mean latency

The task: understand what the customer needs.

A short banking message can hide an important distinction. A pending transfer, a missing transfer and a transfer to the wrong account require different responses. This workload asks a model to classify each request into one of 77 banking intents.

The GPT-5.4 mini baseline correctly classified 2,299 of 3,079 requests. The goal was to reduce inference cost without giving up baseline task quality.

A specialist built for the work.

Isoquant evaluated a Qwen3-8B specialist trained on labeled banking examples against the original general-purpose configuration. Specialist training taught the smaller model the distinctions that matter for this workload.

The final selection cycle reused the trained checkpoint and ran a fresh, matched comparison. It did not require an additional training run.

Lower cost. Higher accuracy. Faster completion.

The specialist answered 526 more requests correctly, lifting accuracy from 74.7% to 91.8%. Estimated inference cost fell by 91.7%, and mean completion latency fell from 931 ms to 588 ms.

That is a win across all three dimensions: cost, quality and latency. The specialist became the saved recommendation for this workload.

MATCHED COMPARISON

By the numbers.

BASELINEGPT-5.4 mini
WITH ISOQUANTIsoquant Qwen3-8B specialist
Baseline and optimized results on 3,079 requests
MetricBaselineWith Isoquant
Intent accuracy74.7%91.8%
Correct / total2,299 / 3,0792,825 / 3,079
Inference cost / 1,000 tasks$0.373$0.031
Mean completion latency0.931 s0.588 s
YOUR WORKLOAD. YOUR SUCCESS CRITERIA.

Let’s build your next performance advantage.

Talk to us
Coding and repair64.4%lower inference costRead the case study Returns-policy decisions97.5%decision accuracyRead the case study Multi-step order resolution98.7%task successRead the case study