- AT&T classifies support calls at 91% accuracy and 35% of the prior cost of its previous large-model classifier, using models fine-tuned with H2O Cascade.
- AT&T cut the time to process a day of support-call transcripts from 15 hours to 4.5 hours.
- H2O Cascade distills a large model's skill on a narrow task into small models you own, so classification work does not have to run on an expensive third-party model at scale.
- You test each candidate model on your own transcripts and compare accuracy and cost with the model you run today before any traffic moves.
Every call has a reason, and you pay to find it
Every call your agents take ends in a transcript, and every transcript holds a reason: a billing dispute, a device fault, an outage, a plan change, a threat to cancel. You need that reason tagged the same way every time to route follow-ups, see trends early and fix what drives repeat contacts.
Asking agents to tag calls by hand is slow and inconsistent, and no person applies a large category taxonomy the same way twice. Sending every transcript to a frontier model works in a pilot, but at full volume the bill grows with every call, the overnight batch runs long, and customer conversations leave your environment.
How support-call classification works
Step 1: Start from what you already have
Your current frontier-model results on historical transcripts become the teacher signal, so the project does not begin with a manual labeling queue.
Step 2: Curate the training data
H2O Cascade generates and balances training examples for each category, so rare call reasons are not drowned out by common ones.
Step 3: Fine-tune small models
Fine-tuning in H2O Cascade adapts an open-weight small language model to your transcripts and your category codes, with support for classification and several fine-tuning methods.
Step 4: Split the taxonomy where it helps
Classification across a large category list does not have to live in one model. AT&T distilled a large model into several smaller fine-tuned models, splitting the taxonomy across them.
Step 5: Evaluate against the incumbent
Test each candidate on your own held-out transcripts and compare accuracy and cost per call with the model you run today, before any traffic moves.
Step 6: Serve on your own infrastructure
Serve the tuned models, or move the weights to your own serving platform, and route calls to them.
AT&T classifies support calls at 35% of the prior GPT-4 cost
AT&T is an H2O customer and replaced a GPT-4 support-call classifier with fine-tuned small models. The result is 91% accuracy at 35% of the prior GPT-4 cost, classifying support calls across its full call taxonomy. The change also moved the processing window: a day of transcripts now takes 4.5 hours instead of 15. This is a vendor-published result from one customer; your accuracy and cost depend on your taxonomy and transcripts, which is why testing on your own data comes first.