quality_assurance · saas · workflow
Dialpad achieves 20% labeling quality improvement and 41% cost reduction with Labelbox
Dialpad's legacy crowdsourced annotation provider delivered inaccurate and missing labels that fell short of quality standards, causing data scientists to grow hesitant to request training data at all because of the time and effort involved.
How it works
Common implementation structure
How this type of workflow is generally built, generalized across documented cases — not tied to any one vendor's stack. Click any stage to read what happens there. Specific products that implement these stages appear in “Tools commonly seen” below.
Stage 1 · Training data need identified
Building and maintaining NLP and LLM models requires large amounts of training data for custom models.
Tools used
LabelboxLabelbox Boost
Outcome
After a year with Labelbox, Dialpad saw a 20% improvement in labeling quality and a 41% reduction in labeling costs, with data scientists now proactively requesting training data and able to scale AI development faster.
What failed first
The previous labeling provider used crowdsourced labeling whose quality steadily declined; extra time and resources spent redesigning labeling projects still could not meet quality requirements.
Results
Volume20%
Cost replaced41%
Grounding & classification
Source type: vendor customer story
27 fields verified against source quotes.
sentiment analysisspeech to textsummarizationfailure mode describedmetric backednamed customerproduction runtime claimedsource backedtools describedworkflow describedsoftwareaccuracy improvementcost reductionemployee productivitytime savedvendor customer storyback office opsquality assurancehuman review queue