Directing Intelligence Digest | November 19, 2026
If your team is starting to hand work to AI agents (software that works through multi-step tasks by itself), OpenAI just showed what that looks like at full speed. Inside its own research group, agents now succeed on their own at 46% of tasks that would take a person two to four hours, up from 23% in March. Yet its researchers say the agents still have “poor taste” in ideas, code, data and writing, and the work needs close human involvement. A vocational classroom in China shows the same thing from the learning side: students who copied the most from an AI assistant remembered the least three weeks later. For a manager or a dean, the job is deciding what to hand off and keeping people able to judge what comes back.
One question to carry through today’s brief: in your organization, who decides which work goes to AI, and who still knows enough to check it?
The short version
Agents doubled their solo success at OpenAI: on 2-to-4-hour research tasks, success with no human help rose from 23% to 46% from March to October, by OpenAI’s own count. (OpenAI)
Saved time mostly goes back into work: AI users in an LSE survey of 3,001 professionals save 19% of the week and put about 72% of it into more work. (LSE and Protiviti)
Copying tracked forgetting: among 272 vocational students, those who pasted the most AI output into practice scored lowest on a retest three weeks later. (Computers)
AI at work: inside OpenAI, agents do more of the research and people still judge it
What was studied. OpenAI published its second report on how AI agents are changing its own research work. It draws on usage logs, estimates of how often handed-off tasks succeeded, a staff survey and ten interviews. It’s a company measuring its own people using its own products, released yesterday. OpenAI report.
What it found. On tasks that would take a person two to four hours, agents succeeded without help 46% of the time in October, up from 23% in March.
The gains are jagged. From August to mid-October, agents succeeded on their own at 79% of tasks watching over model-training jobs, but at only 25% of tasks writing the code that serves models to users. OpenAI notes the trend may also reflect researchers changing which tasks they hand off.
In interviews, researchers said agents got lost in unneeded steps and made invalid analysis assumptions. One researcher, burned when agents got a task almost right, changed the routine: “Go make one. I’ll review it and make edits. Then go do the rest.” People still came up with most of the key ideas behind OpenAI’s latest models, though one researcher credited the agent with key insights.
How it fits with other work. On Tuesday, a London School of Economics team and consulting firm Protiviti released a survey of 3,001 professionals worldwide. AI users save 19% of the week, and about 72% of that time goes back into more work. How well people thought their performance was rated showed no consistent link to AI use. It tracked qualities like resilience and willingness to keep learning. “Leaders need to be clear about where AI adds value and how the time it frees should be used,” said LSE’s Grace Lordan.
Where it falls short. These are elite researchers at a well-funded lab, measured by the company that sells the agents. Its survey of 101 staff is self-reported. Asked what they’d trade to keep their agents for three months, the median answer was 40% of team headcount, a measure of how much they value the tools, not a staffing plan.
What to watch: whether the weak spots close in OpenAI’s next report.
Read it: OpenAI report | LSE and Protiviti report | On X: OpenAI’s post
Who’s Directing? If I ran an operations team, I’d copy that researcher’s routine: have the AI do one example, review and fix it, then let it run the rest, because OpenAI’s people got burned skipping that step. I’d also sort our tasks by where agents are strong, so nobody hands off the 25% work as if it were the 79% work. And I’d decide where the freed hours go, since the LSE team found that’s mostly left to each employee.
AI in learning: students who copied the most AI work remembered the least
What was studied. Shunping Li of Chongqing College of Finance and Economics randomly assigned eight classes (272 students) in a required data-analysis course at a vocational institution in China, in a study published November 13 in the peer-reviewed journal Computers. Some practiced for three weeks with an AI assistant built on the course materials, the rest with an instructor question desk. Three weeks after the first test, everyone retook a similar one with no AI. Li, Computers.
What it found. Having the assistant during the first test added 7 points on a 100-point exam, and practicing with it added about 5. Three weeks later, students who had practiced with it scored an estimated 4 points lower than the others. The author calls that cost “unresolved,” because with only eight classes the strictest statistical checks couldn’t confirm it.
In the AI group, students pasted an average of 41% of their practice steps from the assistant. Those who copied least scored 67 on the later test, and the heaviest copiers scored 60. Students chose how much to copy, so copying may not be the cause.
How it fits with other work. In a survey by the news site Inside Higher Ed of 5,034 undergraduates at 310 mostly public two- and four-year colleges, out Tuesday, 16% say they rely on AI too heavily to think or write on their own, and 65% say responsible AI should help them learn, not do the work for them.
Where it falls short. It’s one course at one institution, with one author and eight classes. The three-week cost wasn’t confirmed by the strictest tests, and the copy measure is new.
What to watch: the semester-long replication the author calls for.
Read it: Li, Computers | Inside Higher Ed Student Voice
Who’s Directing? If I were a community college dean, I’d ask faculty to split AI use by purpose, as the author suggests. Allow it where graduates will use AI on the job, and protect the practice time where students build the skill in their own heads, because that’s where the cost showed up. I’d also ask instructors to watch one signal during AI-assisted practice: how much of a student’s work is pasted straight from the assistant.
Quick hits
Saved minutes don’t prove value. Analysts at research firm Forrester wrote Wednesday that only about a fifth of US companies with 250-plus employees that use AI are building new workflows around it (a blog summary of a paid report). Forrester
Triangulation: signal and noise
Where the evidence agrees. An AI lab, a workplace survey and a classroom point the same way: AI takes on more of the doing, and results depend on what people do with the output. OpenAI’s researchers review an agent’s first example, and the students who copied least scored highest later. That’s Directing Intelligence: deciding what the AI does and checking what it gives back.
Checking only works if people still know the work. That’s Cognitive Coverage, which means having the depth and breadth of expertise the work demands, so you can judge AI’s output and grow from it, and practice is where that depth gets built.
What it can’t tell us yet. OpenAI is grading itself, the LSE data is self-reported, and the classroom cost is unconfirmed.
What I’d do this week. Pick one task your team hands to AI and add a first-example review, the routine OpenAI’s researchers settled on. And ask one instructor to measure how much student work gets pasted from AI, because heavy copying tracked weaker recall.
Until tomorrow, that’s the intelligence. How it’s directed is up to you.
Todd





