Can coworkers and designated 'AI champions' teach people AI skills as well as a manager coaching them one-on-one?

This explores whether peer learning and AI champions can replace direct manager coaching when people are learning to work with AI. To be clear up front: this collection has no studies of workplace AI champions or coaching programs. What it does have is a surprisingly close parallel. AI researchers have spent years asking whether a model can learn from its peers instead of an expert supervisor. Their answers point to what peers do well, and to what a coach still provides.

The encouraging finding is that peers can stand in for an expert judge, as long as the peer group is diverse. When models score each other's answers instead of relying on a human-labeled answer key, a mixed group of different models works much better than a model grading itself, and it often matches training with the answer key Can peer models replace external judges for reward signals?. Grading yourself tends to collapse into reinforcing your own blind spots. The same pattern shows up in setups where one model invents harder and harder challenges while a neutral judge gives verdicts Can language models learn skills without human supervision?. Those setups only work if something stops the group from drifting into its own private game. A survey of these 'co-evolving' systems finds that improving alone stalls, while groups that push each other keep progressing Can agents evolve beyond the constraints humans engineer?. In workplace terms, a peer network beats people figuring it out alone. A champion who is one more voice in a varied group is healthier than a single champion everyone copies.

The caution comes from what kind of feedback actually moves learners forward. Models stuck at a performance plateau didn't get unstuck from more pass/fail scores. They improved when given a written critique explaining *why* an attempt failed Can natural language feedback overcome numerical reward plateaus?. Learning only from expert demonstrations has a similar limit: it caps competence at what the demonstrator happened to show Can agents learn beyond what their training data shows?. That is roughly the risk of champion-led learning built on 'here's my workflow, copy it.' People learn the champion's use cases, not how to diagnose their own failures. Arguably, specific critique of someone's own work is the part of coaching that is hardest to spread across peers.

The one human study here adds a useful twist. People evaluating strategic decisions with an LLM considered more factors but were no more accurate. They also felt more overloaded and less ownership of their choices Does using LLMs actually improve strategic decision making?. That gap isn't about learning features. It's about judgment, workload, and accountability, which sit squarely in a manager's job and are hard for a peer to take on. Separately, the best predictor of success on long AI-driven tasks was persistence through repeated try-check-revise loops, not how good the first attempt was What predicts success in ultra-long-horizon agent tasks?. A structure that keeps people iterating, whoever provides it, may matter more than the job title of the person providing it.

Taken together, the analogy suggests that diverse peers can replace much of the 'is this right?' checking a manager would do. Champions spread practices efficiently. But neither replaces critique that explains why, or ownership of decision quality. These are model findings carried over to people by analogy, not workplace evidence. Answering the organizational question directly would need research the collection doesn't yet hold.