Science1 publisher2 min readPublished
Middle managers told Notre Dame-IBM workshops they carry AI rollouts' untracked 'glue work'
Two July workshops asked midlevel implementers and senior leaders the same questions about AI adoption. The answers diverged over who does the validating, translating and training that no performance review records.
The Scientist · Science desk
What happened
- The Notre Dame-IBM Tech Ethics Lab convened two workshops in July 2026 with the nonprofit All Tech Is Human to ask how AI adoption is changing the experience of work.
- Both cohorts said speed and productivity were becoming the most important values at work, with overall standards declining and "good enough" work becoming the norm.
- The researchers named the untracked tasks "glue work": validating and reviewing AI outputs, translating high-level strategy into practice, cross-functional coordination and employee training.
- Middle managers frequently reported taking on that extra work, and some participants described the new labor as falling disproportionately to women across their organizations.
- Executive leaders questioned whether AI-enabled organizations would still need a middle-management tier, while worrying about the loss of mentoring and knowledge transfer they called "distributed deskilling."
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- contradiction The two executive worries point opposite ways. Cutting the tier whose utility they questioned removes the mentoring and knowledge transfer they said they feared losing, because the workshops found that tier already doing it, largely unseen.
- constraint Work that sits outside job descriptions, performance reviews and organizational metrics cannot be staffed in next year's plan, because nothing in the measurement system will show it was done.
- decision Participants put a set of choices before the rollout: which parts of work stay human, whether people keep room to fail and learn, and whether efficiency gains come back to workers as time.
The mirrored question set is the useful part of this design. Both cohorts answered the same questions [4], one group implementing AI on their teams and the other deciding on adoption across the organization [3]. Because the instrument was held constant across the two levels, a gap between the answers cannot be blamed on a difference in what was asked [20]. That leaves a comparison with one variable named.
A rate would need more than two workshops. The phys.org account does not report how many people attended either session or which industries they came from [18]. The finding about women comes from participants describing their own organizations; nobody counted who performed which task [8]. And people who give up a day to a workshop convened with a technology-ethics nonprofit are self-selected [2]; the manager who finds the rollout unremarkable is less likely to be in the room. Testing the claim takes task-level data inside a firm: who reviews the model's output, for how many hours, at what grade.
Three of the four named glue-work categories are coordination or teaching work, where the result shows up in someone else's performance [19]. The researchers report that experiences of adoption differed by position in the organization and by identities and roles beyond the workplace [22].
Megan McDermott, Notre Dame director of the lab, said she and IBM's Sara Berger wanted a reading of the present while longer-term work continues [14]. "What we learned about things like grief, invisible labor and this fundamental gap between how leaders and workers describe what's going on in their organizations are insights that matter and are important to illuminate in the present moment," McDermott said [13]. The findings and the recommendations are published as a white paper [12].
Berger, the IBM director of the lab, addressed the leaders' side of it. "Nothing is decided for us. There are many futures in front of us, as a collective. Leaders, however, do shape these futures, whether they acknowledge it or not," Berger said [15].
Some participants described what a better rollout would look like, centered on greater organizational coherence and policies supporting human agency [21].
What to watch
- Whether the white paper itself reports participant numbers, sectors, and any counts behind the finding about women.
- Whether any employer publishes task-level or time-use data on who validates AI output, for how long, and at what grade.
- Whether the longer-term research the Lab is funding tracks middle-management headcount alongside the glue-work tasks.