
Hey there,
This fall I was in discussions with a mid-sized software company about its AI work, and one of its leaders walked me through what had happened so far. Early this year, the company did what a lot of companies say they want to do with AI. It pulled together three or four of its most capable people into an AI working group. They built real tools: Claude skills for the sales team, aimed at work the reps did every day. Some reps picked them up. Many didn't.
Then the CTO made a ruling. Using AI is great, he told the group, but it's not as important as your main job. Nobody announced anything after that. The group simply stopped meeting.
Notice that nobody in this story did anything truly wrong. The people in the working group really were too busy. Leadership really didn't want to pull its best people off their real jobs. And the company really did want to get more out of AI. All three things were true at once, and the only way to make them fit was to let the group fade.
When an AI effort stalls, the usual assumption is that the firm didn't know what to do. So it goes looking for a better plan: a framework, a maturity model, a roadmap from someone who has done this before.
But the plan isn't much of a secret anymore. Find out what people are already using. Set guardrails based on how bad a mistake would be. Pick a few workflows that matter and get them working. Get leaders using AI themselves, early and visibly. You can find some version of that list in a dozen places, and I'd guess most firms that stall could recite it.
What they don't have is someone with the time and standing to work through it while the regular work keeps coming. The software company's story is the simplest version of that problem. It isn't the only one. I've seen two others just in the last few months.
At one architecture firm I work with, the AI steering group was capped at seven people to keep it manageable. That made sense on paper. In practice, some members had workloads so heavy they couldn't give the group real time, while other people who were already doing useful AI work outside the meetings had no way in. The firm's fix was to make the group "more porous," and they're still working out what that means. A fixed roster is a bet on who will have time, and that bet goes stale as soon as workloads shift.
At another firm, the AI group is productive. About five people have built somewhere between five and ten tools, and they keep a dashboard of them. The leader of the marketing team likes the tools. Her team doesn't use them, because they go from one deadline to the next and have no time to learn something new or to work with the group on making it fit. Here the builders have the hours. The people meant to use what they build don't.
"Here the builders have the hours. The people meant to use what they build don't."
There's new research that puts some evidence behind what these stories suggest. Arvind Karunakaran of Stanford, with Roshni Raveendhran of the University of Virginia and Tami Kim of Dartmouth, studied AI deployments at a law firm, an ad agency, and an IT services firm, interviewing 184 employees. At the law firm, they followed two closely matched divisions for nearly two years. Both got the same AI tool.
In the first division, the manager pitched the tool to paralegals as a way to draft documents faster and shorten turnaround times. The write-up describes no training, mentorship, or time to explore behind that message. The paralegals used the tool for the tasks they already had, and some said it made them feel like "just another cog in the machine."
In the second division, the manager started by asking paralegals which tasks they found most tedious and what they had long wanted to do but never had time for, then asked whether the tool could take the routine work off their plates. This team also received a lot of support: formal training, weekly knowledge-sharing lunches, Q&A sessions with the vendor, informal mentorship on harder work like legal research, and time set aside each week for self-directed exploration. Those paralegals used the tool 58% more than the first division and experimented with it 70% more. They started using it for legal research and case analysis, and some even lobbied to sit in on legal strategy meetings with attorneys.
To be fair to the research, the two divisions differed in both the message and the support, so the study can't say which one did the work. Karunakaran's own reading is that you can't separate them. Telling people AI will make their jobs better is a promise, and resources are what make the promise believable. "You need to put this scaffolding in place to make the commitment feel credible," he told Stanford's Institute for Human-Centered Artificial Intelligence. Without it, people hear the message as talk.
He also described what usually happens instead: "managers see the time savings and raise expectations for how much of the same work people should do." I wrote about this pattern in Follow the Money, Part 2: when AI compresses the time a task takes, organizations tend to fill the gap with more of the same work rather than with the better work they promised. The software company's CTO was doing a version of the same thing. He never said no to AI. He just never took anything off anyone's plate to make room for it.
The research comes with a cost that I want to state plainly. The paralegals who used AI for deeper legal research and case analysis produced higher-value work, and it took more time, so they handled fewer cases. By the old measure, the division that adopted AI better looked less productive. A firm that makes room for this work should expect some of its familiar numbers to dip, and decide ahead of time whether it's willing to defend that.
I've felt the fragile version of this in my own work. Recently, the internal champion at one of my clients left before the next phase of our work was signed. The executives there were using AI well, and that turned out not to be enough. Until enough people across a firm are engaged, the effort depends on one or two people who have the time and the standing to keep it moving. When one of them leaves, everything they were carrying is suddenly at risk.
So, what can you do about it? I should be upfront that helping firms with this work is part of what I do, so I have a bias. But none of the recommendations below require hiring anyone. Here are three things you can start on today.
Start by naming who owns the AI work, and write down what comes off their plate to make room. If nothing comes off their plate, you haven't given them a job. You've given them a side project, and side projects lose to deadlines every time. The manager in the second law division did a version of this when the tool rolled out, by asking paralegals what they'd gladly stop doing and what they had wanted to do instead.
"If nothing comes off their plate, you haven't given them a job. You've given them a side project."
Then budget time for the people on the receiving end, not just the builders. A tool that a busy team never has an hour to learn is a cost, not an asset. In firms that run on billable hours, that means deciding in advance that some time won't be billed, and protecting it the way you'd protect time for a client.
Finally, decide which number you're willing to let slip. If AI is supposed to buy better work, some measure of volume will probably fall for a while. Leaders who pick that number in advance can defend it when it drops. Leaders who skip this step usually end up making the same call as the software company's CTO, and pulling back before the work has had a chance to pay off.
The software company never decided to give up on AI. It made a reasonable call about priorities, and the working group ended. Many stalled AI efforts end the same way, through a string of sensible decisions that never make room.
So the question I'd put to any leadership team isn't whether you have the right plan. It's who owns this work, and what did they stop doing so they had time to dedicate to it?
Break a Pencil,
Michael
P.S. Who owns AI at your firm, and what came off their plate when they took it on? Reply to this email, or if you're reading on the web, write me at [email protected]. I'm curious how often the honest answer is "nothing."
P.P.S. If someone you know just got handed AI as a side project, share this with them.




