A year ago, the machine waited for you
Twelve months ago the AI conversation inside most banks was about prompts. You asked, it answered, you edited. The system sat still until somebody touched it. That was the whole shape of the thing, and it was straightforward to plan around, because a human stood at every step.
Cameron Hedrick, founder of Hedrick Advisory and former Chief Learning and Culture Officer at Citi, put the pace problem plainly in conversation with Sarah Clarke, CEO of Intuition Americas: people overestimate what happens in a month and underestimate what happens in a year. Now look at your capability plan for 2026. Most of them were written for the version of AI that waits.
Agents do not wait. They hold a goal, take a sequence of actions, call other systems, and change approach when they hit something unexpected. In labs and in a handful of firms this is already ordinary. In most institutions it is arriving faster than the planning cycle allows for.
"Once this thing takes off with agents and proxies, it's going to be something," Cameron said. "And with all of that comes a change in one's self-conception and leadership style and the way we train and the way we do culture."
That last clause is the one L&D should read twice.
Three things most L&D functions are not ready for
The speed, once it turns. Adoption curves do not feel gradual from the inside. The signals are visible now, in pilots and vendor roadmaps, and they are easy to file under "not yet." Then the knee of the curve arrives and the gap between what your people are permitted to use and what they have been trained to use opens in a single quarter.
Their own insecurity. This is the one nobody puts on a slide, and it is the one that will quietly decide whether your rollout works.
Financial services rewards depth. The credit analyst who knows the covenant structures cold. The compliance lead who can recite the reporting calendar from memory. The product specialist everybody calls before a client meeting. These people built careers on being the person who knows, and their standing in the room comes from it.
A system that retrieves and synthesizes faster than they can is not a productivity story to them. It is a status event. Cameron named it directly: if you have built your career on being a technical master of something, and a machine is clearly faster at the knowledge part, you have to deal with your own insecurity before you can do anything useful with the tool. Teams do not say this out loud. They express it as tool skepticism, as data quality objections, as "we tried that." Treat those as a training gap and you will train the wrong thing.
Psychological safety across humans and machines. Human-machine teaming needs the same conditions any high-performing team needs: clarity about who owns what, permission to be wrong, and a manager who knows what good looks like. Most institutions have never had to design for it.
The job is changing into something closer to architecture
Cameron's thesis about the function itself is worth arguing with, because it does not flatter anyone. L&D is not going away. It is going to be responsible for training agents as well as people, and it will only work if it sits much closer to two groups it has historically operated beside rather than with.
Recruiting, because recruiting is about to be sourcing every form of intelligence in the building, not only the human kind. And culture, because culture governs the conditions under which any of that intelligence actually gets used.
He is blunt about the alternative. An old-school function drops content into an environment knowing the content will have evaporated from the learner's mind by that afternoon. His words: it's like just throwing away money. That is not a critique of instructional design. It is a critique of designing for delivery instead of designing for conditions.
There is a second shift underneath it. You are a systems architect now, whether that is in your job description or not. Not a programmer. Nobody needs the CLO writing code. But you should be able to answer three questions about your own stack without calling anyone: What senses the skills in this organization, and how accurate is it? What creates the content, and what data is it standing on? What gets the right thing to a person at the moment they need it? If those questions feel like somebody else's, that is the gap to close first.
What to do before budget season closes
Six months is enough time to do two things properly. It is not enough time to do eight things badly.
Make experimentation genuinely safe, and mean it. Not an AI academy with a completion certificate. Not a governance committee that meets monthly and approves use cases quarterly. Real permission to use the tools inside real work, with room to get it wrong. Cameron does not soften this one: if you don't, you're dead before you start. For a regulated firm the honest version is bounded permission, published clearly, with the boundaries trained rather than emailed. Unsanctioned use is untrained use, and untrained use is a control gap, not a productivity story.
Build structured reflection into the work itself. This is the cheapest high-value thing on the list and almost nobody does it. When someone brings you AI-assisted work, ask why they made that choice. Ask what the implications are across one month and across three years. Ask what they checked, and what they decided not to check. Cameron calls it grooving the wires. It is also the closest thing anyone has to a repeatable method for building judgment, and judgment is the thing you cannot buy a license for.
Both of these are free. Neither needs a platform decision. Both start on Monday.
The measurement problem you will hit in month four
Completion data will not survive the first serious question from a CFO or a risk committee about whether any of this changed behavior. If you are going to ask for budget in September, agree the two business metrics you will be judged on before the program starts, in writing, with the sponsor who owns them. Baseline capability first. Re-assess after. That sequence is unglamorous and it is the difference between a program that gets renewed and one that gets absorbed into a cost review.
The agentic shift is not a scenario to plan for in the abstract. It is the next chapter, and the L&D leaders who matter in it are already asking a different set of questions.
2026 Free L&D Maturity Assessment
Score your learning function across five dimensions, see how it compares with peer institutions, and get a gap analysis you can take into a planning conversation.