AI
The 1:1 ratio of talk time to paperwork
Contact centers break under stress. Over 50% cite hiring, training, and productivity as critical barriers. High stress is the number one reason operators leave. After-call work (ACW) takes 6.3 minutes per call. That matches the average call duration of 6.5 minutes. Operators spend as much time typing notes as talking to customers. Summaries depend on memory and writing skills, so data quality is inconsistent. Fujitsu engineering targeted ACW directly: mechanize summarization and data extraction to cut post-processing time by 50% or more.
Four-stage low latency pipeline
The pipeline starts with voice capture: stereo audio is split into agent and customer channels, noise filters remove back-office chatter, and buffer management masks credit card numbers and PII before the data reaches the LLM. The speech-to-text (STT) engine must maintain above 90% accuracy using domain-specific dictionaries — for example, distinguishing “term life” from “term right” in insurance calls. Inverse text normalization converts spoken “five thousand dollars” to $5,000. The generative AI core uses few-shot prompts to output separate bullet lists for customer inquiry and operator action, plus a reasoning layer that classifies intent from a predefined list of call reasons.
Results: 50% less ACW time, consistent data
After deployment, average ACW dropped from 6.3 minutes to 3.1 minutes. Across 500 seats handling thousands of calls daily, that reclaims dozens of full-time headcounts. Data entry quality moved from subjective operator memory to a standardized, logic-based output. Call reason tagging became consistent. By removing repetitive administrative typing, cognitive load dropped, which directly combats the stress that drives turnover.
Constraints and the three-phase roadmap
Three constraints remain. STT accuracy on heavy accents or poor audio quality limits the LLM. Initial API token costs for long 20-minute transcripts are high during scaling. PII masking adds latency and architectural overhead. The roadmap addresses each. Phase one adds explainable AI that coaches operators on soft skills and empathy after calls. Phase two feeds categorized intent data into time series analytics to forecast call volume spikes and optimize shift scheduling. Phase three builds low latency sentiment and acoustic analysis to detect customer verbal abuse and automatically alert supervisors or transfer the call to an AI voice agent.
Notable Quotes
operators are spending almost as much time doing administrative data entry as they are actually talking to the customers Thor Schaeff · ▶ Watch (5:41)
Key Takeaways
- A four-stage pipeline cuts after-call work from 6.3 to 3.1 minutes.
- Stereo channel separation and PII masking are prerequisites for reliable LLM summarization.
- The roadmap targets operator coaching, predictive staffing, and abuse detection.