AI can cut ops time

A Harvard team tested a “stopping agent” that watches live sales conversations and decides when it’s better to end a call, and on a dataset of 11,627 calls the system raised expected sales by more than 30% while identifying likely failures within the first minute of a call. (d3.harvard.edu) Separate work measuring generative-AI tools in security operations found about a 30% reduction in the average time to resolve incidents, and McKinsey’s State of AI survey shows many companies are experimenting with agent software but only about 23% have scaled an agentic system in at least one function and at most ~10% of respondents report scaling agents within any single business function. (arxiv.org) (forbes.com) “AI agents” here means software that can autonomously carry out narrow tasks — for the Harvard experiment the team fine-tuned a large language model (a pre‑trained text model adapted by additional training on task-specific examples) so it could read call transcripts and choose “quit” or “wait” decisions. (d3.harvard.edu) The Harvard group framed the problem as an “optimal stopping” task — the system learned, from hindsight, the moment that would have maximized reward by testing different hypothetical quit times — and reported that human sellers typically let low‑probability calls run far longer than the agent, with the dataset showing an average call length of 169 seconds and a failure rate above 94%. (d3.harvard.edu) Taken together, these results explain why founders are being told to automate routine workflows: narrow, well‑defined agent workflows can cut the time spent on repetitive operations by substantial margins (Harvard and the security-ops study both show ~30% effects), but industry reports stress that firms capture value only when they redesign processes around those agents and monitor outcomes as they scale. (d3.harvard.edu) (in-parallel.com) (gartner.com)