Paul Roetzer opened MAICON 2026 in Cleveland with a question I'm still turning over a day later. If AI now does the work people used to learn from, how does an organization develop its next managers and executives?
Every session I attended after his keynote answered a piece of that question. The last one, Andrew Yang 's, asked it again at the scale of the whole economy and left it open.
Paul's keynote was inspirational and amazing. It set the agenda for the rest of my day.
He walked through the early years of his career, writing press releases and pitches billed at $125 to $150 an hour. That work was how a young professional learned the craft. AI can now do much of it for about $20 a month. Nobody decided to pull out the bottom rungs of the career ladder. They are disappearing anyway.
His answer is a new team design built around three roles. The architect designs the systems and the governance. The orchestrator runs them every day and directs the agents. The apprentice replaces traditional entry-level production. That person reviews AI output and watches experienced people work alongside it, building the context and judgment the job used to teach through sheer repetition. Paul wants that learning to happen in the open, with agent interactions in shared channels where newer staff can watch how decisions get made.
I fully agree with the apprentice model. I made a version of the same argument in my last article about law firm marketing and BD teams: put the AI draft next to the version a senior colleague would have sent, have the senior person explain why one is stronger, and get the reasoning out of private calls and into channels the team can see.
Paul also described a reaction he has seen in the legal industry. The first instinct is to cut associate hiring. The second thought is that today's associates are tomorrow's partners. Law firm marketing and BD departments face the same arithmetic with their coordinators and specialists. If those junior roles get cut now, nobody will be ready to run a practice group's business development or lead the department a few years from now.
Claire Potter-Schneider's session was incredibly valuable. She has led marketing AI transformation at two companies, first at Motorola Solutions and now at Dynatrace , where she is Head of Marketing AI Transformation. A blueprint from someone who has done this twice is about as practical as conference content gets.
At Motorola, 30 marketers saved 12,000 hours and generated $1.5 million in pipeline in the first six months. Adoption was still at 80% two years later. She was clear that those results came from a program with no end date. Pilots end.
Her governance model swaps approval gates for guardrails. Leadership publishes clear rules and tiers use cases by risk, so nobody waits for approval to summarize their own meeting notes and nobody skips review on work that touches client data. For law firm marketing teams still waiting on IT and the general counsel, that tiering offers a way to start on internal work now.
Then she laid out an enablement program that begins with assessment: an anonymous benchmarking survey to find out where people really are before anyone designs a course. I fully agree. A survey shows you who is already building on their own (often with tools nobody approved) and who hasn't started. Training designed before that survey is a guess about its audience. I build AI training for law firm marketing and BD teams, so I have a stake in this. I would still put the survey first every time.
Keith's presentation was good as usual. I loved his session last year and I loved this one, for the same reason. It is always practical, with resources and downloads you can put to work the next morning.
His approach starts at the bottom. A role breaks into job functions, functions into tasks, tasks into subtasks. You build an agent or a plugin for one subtask, then the next. Automate enough subtasks and the task follows, then the job, then the function. At his firm, L2 Digital, that approach cut ongoing client work from 20 to 30 hours per client each month down to 5 to 8.
His method starts with the subtask you dislike most. Do it by hand five times with a stopwatch running. Strip out every step you can before automating anything, so you are not baking in bloat. Then roll it out slowly, on what he calls a crawl-walk-run schedule. Review every output at first and step back only as the outputs earn it. Never let an agent with write access or a public connection run unsupervised.
I like this approach because it matches how adoption happens. Nobody automates a function. You automate the Tuesday afternoon chore of pulling the weekly media mentions report or reformatting attorney bios for a pitch. Then you automate the next one.
It is also the bottom-up half of what Claire described from the top down. Her guardrails define what people can build. Keith's method shows them how.
Holly Chen , VP of Growth Marketing at Samsara , gave an incredible presentation, the best session I have seen so far this week.
Her marketing organization went from about 3% of staff using AI tools to more than 90% of marketers building their own workflows and agents in roughly a year, with over 100 agents in use. Pipeline per marketing dollar rose 30% year over year. Pipeline per marketing head rose 50%. An eight-person team in Mexico built 14 of the growth marketing group's 40 agents. You do not need to sit next to a tech hub to do this work well.
The point that resonated most with me was about hiring. Don't go looking for AI-native people. Make the team you already have AI native.
That is Paul's apprentice idea applied to the whole department. The BD manager who knows which partner signs off on what and the specialist who knows which clients each practice group is chasing both carry knowledge no new hire walks in with. Teaching them to direct agents is faster than teaching a newcomer everything they already know.
Holly was also candid about what breaks at scale, especially teams building the same tool twice and prototypes with no owner. One landing page agent stopped working when its creator went on leave. Her fix was a three-level structure for personal, team and org agents, each with its own review standard. Claire and Holly work at very different companies. Both built their governance on tiers.
Her quality bar: if a deliverable looks AI-generated or falls apart under light scrutiny, it isn't finished.
Andrew Yang closed the day, and I walked out unsettled.
He is a funny speaker and he told a lot of jokes. But from where I sat, his vision of the future was downright depressing, and the jokes did not match it. It felt like one of those songs with a bouncy, upbeat melody wrapped around a lyric about losing everything.
He walked through the five most common job categories in the country, from clerical work to truck driving. Together they account for roughly half of American jobs. He showed how exposed each one is to automation. He also predicted that 25% of colleges will go out of business within three years. That one I'm still sitting with.
He told a story about a Brown University economics course. When students could use AI on a take-home exam, class averages jumped well above their historic norm. When the professor required an in-person, pen-and-paper final, 30% of the class dropped the course. The students who stayed posted the lowest scores in its history. That is Paul's broken ladder, inside a classroom.
Yang's remedy is to shift economic value toward health and fitness, art and creativity, and taking care of people, paid for in part by taxing automation instead of human labor. I like those ideas. I just don't know how we actually get there. He didn't lay out a path from here to there.
Paul asked how an organization grows its next leaders when AI does the entry-level work. The three sessions in the middle of the day gave me real answers for the law firm marketing and BD teams I work with. Yang asked the same question about the whole economy. Nobody on that stage had an answer for that one.
On to Day 3. I'm looking forward to it.