The Change Management Playbook Procurement Already Had
MIT’s new research on AI adoption confirms what a decade-old change management framework predicted but that most rollouts still ignore.
In Part 1, I argued that procurement’s slow AI adoption is only half a technology story. The other half is the same change management problem I described in 2018, before generative AI existed. It hasn’t gotten easier just because the technology got better. This piece explains what closes the gap, drawing on two sources that don’t usually appear in the same paragraph: an MIT research team’s recommendations for managing generative AI at work, and the practitioner-level change management playbook procurement has needed long before anyone talked about copilots.
The MIT report doesn’t stop at describing the bottleneck, cafeteria, and learning curve problems. It offers six practical recommendations for organizations trying to get the upside without the damage, and each one maps onto a procurement failure mode I’ve watched play out.
Minimize the drudgery people actually dislike, not the tasks that happen to be easiest to automate. A category manager who enjoys building a negotiation strategy but hates reformatting the fourth version of a comparison matrix wants the matrix automated, not the strategy work handed to a model. Ask people what they’d cut first before deciding for them.
Promote learning and watch for what MIT calls mental offloading. A junior buyer who lets an AI draft every negotiation position without ever building the underlying judgment ends up dependent on a tool rather than developed by one, and MIT calls this out as a real risk worth designing against. Procurement teams need AI use that still asks juniors to do the reasoning some of the time, even when the AI could do it faster.
Preserve teamwork and mentorshi p. The cafeteria problem is real, but the cafeteria conversation was never just an information transfer. It also built the relationship that made the next ten favors possible. An AI that answers the question the expert would have answered still leaves that relationship unbuilt, so procurement functions that let AI fully replace cross-functional check-ins will save time and lose something they’ll miss later.
Design interfaces for situational awareness, not just output. A recommendation an AI hands a buyer with no visibility into how it got there is a black box, and buyers who don’t trust the box either ignore it or, worse, follow it without checking. Interfaces need to show their work.
Keep domain expertise in the loop to validate what the AI produces. This is the least glamorous recommendation and the one most likely to get skipped under budget pressure, because keeping an expert reviewing the output makes it look like the AI didn’t actually save the headcount it promised. It’s also the recommendation that prevents the expensive failure: a wrong AI output that nobody with the expertise to catch it ever reviewed.
Get Bertrand Maltaverne’s stories in your inbox
Join Medium for free to get updates from this writer.
Build accountability, so AI use doesn’t become a shortcut around judgment. If nobody’s job depends on the quality of the decision, and the AI made the recommendation, responsibility diffuses to nobody in particular. Procurement functions need to decide, explicitly, who owns an AI-assisted sourcing decision before the first one goes wrong, not after.
The fourth piece in my 2018 series, “ The Reality of Change,” made an argument that maps almost exactly onto what procurement needs for generative AI: don’t try to convince everyone, because you won’t.
Roughly a quarter of any organization resists a given change regardless of the evidence you bring them, and roughly a quarter embraces it early regardless of how rough the first version is. The group that actually decides the outcome is the passive half in the middle, who will follow whichever side looks like it’s winning. That makes the single most useful move for an AI rollout something other than a company-wide training session aimed at converting skeptics. It means finding the early embracers, turning them into visible champions and super users, and letting the passive middle watch people they already trust use the tool successfully before being asked to use it themselves.
The same piece applied the Kübler-Ross change curve, originally a grief model, to technology rollouts: denial (“this won’t work for a category like ours”), anger, bargaining, depression, and eventually acceptance, usually only after someone actually tries the tool and finds it survivable. That curve runs on its own timeline, and no amount of executive enthusiasm compresses it. Change management for an AI rollout has to run before the tool goes live, when expectations get set and champions get picked; during it, with visible support and fast fixes for the failures that will happen; and after it, reinforcing the new behavior long enough that it survives contact with a busy quarter.
The competence progression from the same piece matters just as much here as it did for eSourcing: people move from not knowing what they don’t know, to knowing what they don’t know, to a competence that takes visible effort, to one that finally feels automatic. Skipping straight to expecting automatic competence, which is what most AI rollout timelines quietly assume, guarantees the rollout will look like it failed when it was actually just early.
MIT’s six recommendations describe what a well-designed AI-augmented procurement job looks like: less drudgery, more supervised judgment, expertise still in the loop, accountability that doesn’t evaporate. The change management playbook describes how an organization full of people with the psychology I described in Part 1 actually gets there, and it isn’t through a launch announcement. It’s through champions, a curve that takes time, and preparation that happens before the pilot rather than after it stalls.
Neither half works alone. A well-designed AI tool rolled out without change management runs into the same wall eSourcing hit for a decade. A textbook change management program wrapped around an AI tool that dumps unreviewable output on already overloaded buyers won’t survive contact with the first bad recommendation. Procurement has the second half of that equation written down already, a decade before generative AI needed it. The organizations that get ahead on AI adoption will be the ones that finally use the playbook they already had, not the ones waiting on a better model.
Sources: MIT IPC, “ Humans in the Loop: The Evolution of Work in Early Experiments with Generative AI” (Armstrong and Shah, April 2026); my own “ Procurement’s Technological Insanity” series on Medium (my previous blog), particularly The Reality of Change and Repeating the Same Mistakes Over and Over.
AI disclosure: AI was used to research the topic and for editing purposes.
Originally published at https://procurementtidbits.substack.com on October 16, 2026.
