Operators
When it becomes too much
We’ve written at length about getting teams to start using AI. We were pleased to have two conversations this month that were past that point, but it’s creating a new problem:
“Do you ever just feel like AI might be using too much time? I’ve got Claude running on two accounts, Codex on another window, and that’s just me. Part of me feels like [our company] is just doing AI for the sake of AI.”
- Founder, Professional Services Company Focused on Fundraising
“What I’m trying to get across to my team is ‘hey, does [using AI] accelerate the project schedule or is it just a nice to have?’ And if it doesn’t accelerate the schedule, they should just do it by hand.”
- CEO, Specialty General Contractor
Whether it's self-policing from the first quote or managing a team from the second quote, we see that the use of AI is not infinite. After seeing a couple wins, there’s a risk that teams start using it as a one-size-fits-all crutch.
Wide-range adoption when getting started is a different story, so we don’t want to discourage that. However, the experimentation returns taper off. Eventually whether or not to use AI becomes task dependent.
A few weeks ago we wrote about where the machine should stop and the human should start (Declaring Victory). While that post was about end-to-end workflows, this time we draw the same line for individual tasks.
In broad strokes, here are three signs that AI usage may have crossed the line:
Prompt time > work time: regardless of your typing speed, good prompts often take multiple sentences and several minutes and that time is rarely factored into the AI clock.
Revision time > work time: rushed or weak prompts easily lead to more time than expected on the back end as people rewrite prompts or revise outputs.
Manual time not considered: if people stop estimating how long it will take them to complete a task by hand, then the right decision is hard to make.
Here’s something to do in the next week:
Before starting any task with AI, write down how long it would take you to complete without it. Then time yourself with AI, including prompt drafting and output revision, and log the result. At the end of the week compare the two columns to see where AI lost.
There’s no way to know what is or isn’t faster without running an individual test. And while this might be a fun exercise for a particular week, the timing tradeoff should eventually become a question asked naturally before every action.