Case Study — Cutting AI Call-Scoring Costs 75% — Straight Up AI
Case Study · Healthcare AI · Cost Reduction

Quality maintained,
for 25% of the cost

A pharmaceutical cleint developed a system to score sales conversations against 25 KPIs. The first build calculated each KPI independently. That meant sending the same 30–60 minute transcript 25 times, and paying for it 25 times. We grouped the KPIs that were semantically related and calculated them together. Six requests then did the work of 25. Quality stayed the same, and the client could add new KPIs without scaling the cost.

Client
Healthcare
Service
ML / GenAI Engineer
Focus
AI Cost Reduction
Stack
Python · OpenAI · Databricks · MLflow
The Challenge

Paying for the same transcript 25 times over

The 25 measures come in two kinds. Some are simple: did the rep set out an objective before the call? Others need judgement: did they spot a chance to grow the account? The obvious way to build this is to ask the model each question on its own. But every one of those questions needs the whole transcript to answer it. So a single conversation gets sent, and charged for, 25 times.

Each transcript is a 30–60 minute conversation, and the system was meant for every rep in the United States. The bill grows with the number of measures, the length of the call and the size of the sales force, all at once. Most of that money is spent on the model re-reading words it has already read. The obvious build treats 25 related questions as 25 unrelated ones, then repeats that waste on every call in the country.

The Approach

Group what's related, ask it once

Before
25 ×
~200,000 tokens · ~$250 per 1,000 calls
After
6 ×
~48,000 tokens · ~$60 per 1,000 calls
The Results

Same scores, a quarter of the bill

−75%
Fewer tokens used on each conversation.
$250 → $60
What it costs to analyse 1,000 conversations.
No additional LLM calls
Provided new KPIs are semantically related no additional LLM calls are needed.

AI bill growing faster than your usage?

See if we're a fit →
Previous
Previous

Negative Sampling

Next
Next

Image Consistency