- H2O Driverless AI trains natural language processing and multiclass classification models that tag each complaint or query with your own categories, in real time.
- The same models that classify a contact can score the caller's risk of leaving, with a reason for every score.
- AT&T, a telecommunications customer, cut the time to process a day of call transcripts from 15 hours to 4.5, at 91% accuracy and 35% of the prior GPT-4 classification cost.
- The models run in four deployment options: on premises, air gapped, private cloud and cloud, so customer conversations stay inside your organization.
Turn every call into a chance to keep the relationship
Your customers use the app until something goes wrong. Then they call: a disputed charge, a blocked card, a payment that did not post, a fee they do not understand. That call is often the moment a customer decides whether to keep the relationship.
Your contact center is measured on handle time and cost per call, but the money is in what happens next. A complaint that is misrouted, answered twice or closed without a fix shows up later as a closed account or a product never bought.
Most of the signal is already in your hands. Call transcripts, chat logs, complaint records and product holdings sit in separate systems, so your agents answer the question in front of them without seeing the customer behind it.
How call center optimization works
Step 1: Bring your contacts together
Join call transcripts, chat logs and complaint records to the customer's accounts, products and recent activity.
Step 2: Classify the reason for every contact
H2O Driverless AI trains natural language processing and multiclass classification models that tag each complaint or query with your own categories, in real time.
Step 3: Suggest the resolution
A second model maps each contact type to the resolution that closed similar cases, so agents get a suggested next step and routine contacts can move to self-service.
Step 4: Flag the relationship at risk
The same models score the caller's risk of leaving, with reason codes, so an agent knows when a routine call is really a retention call.
Step 5: Extend to generative AI where volume justifies it
H2O Cascade fine-tunes a small language model on your own transcripts, so high-volume classification runs at a cost you control.
AT&T's call center results, from telecommunications
AT&T, a telecommunications customer, replaced a GPT-4 support-call classifier with fine-tuned small language models on H2O Cascade. The result: 91% accuracy at 35% of the prior GPT-4 cost, and the time to process a day of call transcripts fell from 15 hours to 4.5.
This is a telecommunications result, not a banking one. It shows what the classification step does at contact center volume. Your result depends on your own categories and data.