Weekly Experiment
This Week's Practical AI Bet
One concrete, time-boxed experiment per week, chosen for its balance of low effort and clear business value.
Week of Jul 6, 2026
Prompt caching cost benchmark on VoC summarization
Hypothesis
Enabling Bedrock prompt caching on our repeated-context summarization prompts will cut input token cost by 60%+ with no quality regression.
Estimated effort
3 engineer-days
Tools required
AWS BedrockProvider abstractionEval harness
Expected value
Direct COGS reduction on the highest-volume AI feature; hard number to justify the Bedrock migration.
Previous experiments
Week of Jun 29, 2026
RunningComputer-use agent for CX ticket triage (sandboxed)
A UI-operating agent can correctly categorize and route 80% of a sample of support tickets in a fenced sandbox.
Anthropic computer-useSandbox VMTicket sample set
Week of Jun 22, 2026
CompleteVoice agent-assist latency spike
The Realtime API can deliver suggested responses under 400ms end-to-end in a live call simulation.
OpenAI Realtime APITwilio simRetrieval index