Mandate 01

The Promotion Nobody Explained

The AI mandate for knowledge workers and managers. What your organization now expects, why the announcement never came, and how to meet it from inside the job you already have.

ContentsOpening
  1. Opening
  2. Premise 1The Mandate Has Already Arrived
    1. Employees without AI skills won’t be needed. AI workers will.
    2. The work expanded. The job description did not.
    3. Do not wait for employer training. Start now.
    4. What you can do now
  3. Premise 2Expertise Is Never Finished
    1. Practice on small, real projects.
    2. Make time for play, not just practice.
    3. Learn faster with peers than alone.
    4. What you can do now
  4. Premise 3Build What Outlasts the Tools
    1. Build AI skills in stages.
    2. Learn principles over tools.
    3. Grow your thinking. Not just its application.
    4. What you can do now
  5. Premise 4Integrate AI Into the Work
    1. Add AI to the workflow you already have.
    2. Redesign workflows around AI.
    3. Develop agent creation and coordination skills.
    4. Build safeguards as fast as capabilities.
    5. What you can do now
  6. The Closing ChallengeBecome the AI-Powered Professional
  7. Source notes

Somebody down the hall is doing in an afternoon what the schedule still says takes a week.

They did not get a memo. They were not trained. They were handed the same access you were handed, and they started before anyone told them what good use looked like. Nothing about that is a verdict on you. It is information about the standard, and the standard moved before the words used to describe it did. Which is the useful part. If nobody announced it, then nobody got ahead by being told, and nobody in the building is closer to your work than you are.

A manager and a small team of scientists sat down to talk about AI. They were intelligent, highly trained professionals who reviewed scientific literature for a living. Their organization had given them access to AI tools and told them to use them.

They tried.

What came back was generic or scientifically weak. They had the tool. They had a mandate. They did not have a method.

After one practical discussion about how generative AI differs from the systems they were used to, the manager said something I have not forgotten.

“This is the most practical discussion I’ve had about AI since we were given access to the tools and told to use them.”

That sentence captured the problem. Access came first. Expectations followed. Instruction never arrived.

If people who read primary literature for a living had that gap, the gap is a condition of the moment rather than a personal failing.

There was no companywide meeting. No revised job description arrived. No leader stood at the front of the room and explained that the standard for professional work had moved.

It moved anyway.

The people who noticed had no better information than anyone else. They watched routine research, reporting, writing, and coordination compress into a fraction of the time. They learned that the first draft was no longer the work. Choosing the right work, defining what good looked like, and judging what came back became the work.

The people who did not notice kept meeting the old standard. They worked hard. They stayed competent. They remained responsible. None of those qualities became less valuable. The agreement around them changed before the words for it did.

That is what makes this moment difficult. The danger is not a machine arriving one day to take a whole job. The danger is a job expanding in pieces while nobody tells the person holding it. The standard moves quietly. The professional discovers it only after being measured against it.

Waiting feels reasonable under those conditions. A responsible employee waits for an approved tool, an official policy, and training tied to the work. Yet adoption did not arrive that way. It came through individuals who began using what was available, often before the organization had decided what to call it.

Waiting is not neutral. Waiting is the only choice with no upside.

This is the siren before the storm. The threat is real and shelter can be built. The institution may never sound the siren in the form you expected, because the institution is learning at the same time you are.

There is another side to this change. The same shift that threatens a professional who waits can elevate a professional who moves. When routine coordination shrinks, judgment becomes more concentrated. When production accelerates, choosing what deserves to be produced becomes more valuable. When a machine can create ten plausible answers, the person who can recognize the right one matters more.

Responsibility that once sat two levels above you is moving closer. Direction. Standards. Accountability. It arrived without the title and without the announcement.

This is not a demotion. It is a promotion nobody explained.


Would you hire a professional today who could not use a spreadsheet? Of course not. That stopped being a skill and became a floor. The question is no longer whether AI belongs in professional work. It is whether you will integrate it into your own work before the new standard is used to judge you.

That is why this is a mandate.

Premise 1The Mandate Has Already Arrived

Learning AI is not optional. It’s a mandate. Organizations expect it and training may never come.

Competence and tenure used to create protection because the world of work changed slowly. A person could master the systems, understand the unwritten rules, and build value through accumulated knowledge. Those things still matter. The pace of change accelerated and it is still accelerating.

The mandate comes from two forces no individual controls. The first is a capability gap between people who can use AI inside their work and people who cannot. The second is an unannounced expansion of the work itself. It is not too late to start. Learn deliberately and begin before the institution finishes deciding what its program should be.

Employees without AI skills won’t be needed. AI workers will.

AI arrived with product demonstrations, quick-start guides, and lists of features. It did not arrive with instruction in the craft of using it inside a specific profession. Organizations often provide access, general guidance, and rules about what cannot be shared. They rarely teach an analyst, program manager, clinician, engineer, or attorney how to redesign the work they actually own.

That distinction matters. An apprentice blacksmith does not become a smith because someone placed tongs and a hammer nearby. Access to the equipment was never the missing piece. Instruction in the craft was. The same is true here. Knowing where to type a request is not the same as knowing how to use AI to research a decision, challenge an assumption, create a reliable work product, or improve a process.

I saw that gap in the team from the opening. The scientists did not need another feature tour. They needed a different mental model. I explained that generative AI produces probable responses rather than deterministic answers. Then we broke their own work into steps and identified where AI could help.

One conversation changed the problem. They had read their weak results as evidence that AI could not handle scientific work, or that they were not good at using it. The real problem was instructional. Nobody had shown them how to translate expert work into a sequence the tool could support.

Capable people often read an instructional gap as a personal failure. It is not. The gap is real. Closing it is still our responsibility.

The training gap is measured. Section’s AI Proficiency Report found that even in organizations deploying agents, only about one in three employees received agent-specific training. Senior leaders were far more likely to receive both access and support. The people closest to the daily work were often handed the tool and left to invent the craft themselves.

About one in three employees received agent-specific training

Employees in organizations already deploying AI agents. Each dot stands for one in thirty people.

Source. Section, AI Proficiency Report. The report gives about one in three, drawn here as 10 of 30.

The labor market has begun to reflect that gap. Stanford and Harvard studies, one of employment patterns and one of roughly 200 million job postings, show entry-level hiring contracting more quickly in occupations exposed to AI. That does not mean work disappeared. It means the work that once converted academic knowledge into professional judgment is changing. Employers can automate parts of the old apprenticeship while still expecting the judgment it used to produce.

This is not a reason to blame people who are not already fluent. The path is new, uneven, and poorly marked. Few organizations know how to teach domain-specific AI capability because few have built it deeply enough to understand what the curriculum would require. The absence of instruction is real. It cannot become a reason to stand still.

A professional who waits for perfectly matched training stays broad and shallow. A colleague who takes one part of the work and learns to do it with AI begins compounding. A year later the difference looks like talent. It was practice. Several years later the difference becomes expensive to close because the colleague is no longer learning a tool. They are building judgment about when and how to use it.

The hiring calculation changes quietly around that difference. One candidate can describe AI in general terms. Another can show how they used it inside the work, what they verified, and what improved. The second person has evidence. The first has exposure. Organizations may struggle to define AI fluency, but they can recognize demonstrated capability when it appears in front of them.

You were handed the tool. You were never taught the craft. The craft is yours to build.

The work expanded. The job description did not.

The job description at issue is not the document in an HR system. It is the working agreement about what a professional is expected to produce and how that work is judged. That agreement moved. The title, the pay band, and the formal review criteria often did not move with it.

Organizations already expect the same work to be completed with AI. They expect it faster, at higher quality, and at greater volume. That expectation arrived without a memo or a title change. A 2026 BCG survey of more than 11,000 workers found that regular frontline use had risen rapidly while most workers received little guidance about what to do with the time AI freed. Many also reported greater cognitive load. The old task got faster. The space around it filled.

The next expansion is visible, although it is not fully here. Managers in the same survey expected agents to absorb a large share of their own tasks within several years. That is their belief, not a settled forecast. The direction is still useful. Tools assist a task. Agents carry a task. When that change reaches ordinary work, coordination is the first layer to compress.

Coordination collapses. Control does not.

Status collection, workflow routing, report packaging, and routine follow-up can be absorbed. Judgment, specification, exception handling, and accountability remain. They become concentrated in the professional who directs the system and answers for what it does. The work does not simply shrink. The center of gravity moves upward.

IBM offers a more complete picture than the usual headline. AI absorbed work associated with roughly 200 human resources roles. Total employment later rose because investment moved into software engineering, marketing, and sales. Arvind Krishna described the growing work as roles where employees face up or against other humans rather than doing rote process work. The climb was not automatic, but it was real.

Radiology carries the historical lesson. In 2016, Geoffrey Hinton suggested that training new radiologists no longer made sense. A decade later the field offered a record number of residency positions, compensation had risen roughly 48 percent since 2015, thousands of posts sat open for months at a time, and nearly half of radiologists were using AI. The field did not escape the technology. It absorbed it. The work expanded around the tool.

I watched the same expansion happen on a smaller scale. I built a process that helped my team produce scientific article summaries faster. The quality of each summary did not suddenly improve. Expert people were already producing expert work. What changed was the time required and how often the summaries could be produced.

The gain became the new baseline. The team could provide summaries regularly without letting them consume the job. The time that came back went into relationships, support, and work requiring human judgment. Nobody revised the formal job description. Nobody called a meeting to announce a higher expectation. The capability existed, everyone could see it, and the standard moved.

Nobody behaved unfairly. The managers’ reasoning was sound. That is what makes the pattern hard to see. Productivity gains do not always return as spare time. They often return as a larger definition of what the role can carry.

Not every organization reinvests as IBM did. Some use technology to handle more volume with fewer people. That fear belongs in the argument. It does not define the whole argument. Organizations eventually rebuild around new capability, and rebuilding creates work that the old structure could not imagine. The near-term pressure is real. The long-term shape is not simple subtraction.

A professional can miss the first expansion because nobody announced it. The second one is visible. That asymmetry matters. What felt like a failure to keep up was often a promotion nobody explained. The rote half of the job is being absorbed. The judgment half is being enlarged.

Defending the disappearing half is understandable. It is also dangerous. A person can become excellent at collecting updates, packaging reports, and moving information through layers just as those layers lose value. The better response is to identify the decisions, exceptions, relationships, and standards underneath that coordination. Those are the parts of the role moving closer to the professional who remains accountable.

The half of the job that survives is the half you were never given enough time to do.

Do not wait for employer training. Start now.

The tools arrived. The training did not. No memo is coming soon enough to protect the time that waiting costs.

Microsoft’s Work Trend Index documented how adoption actually spread. Most knowledge workers were already using generative AI, and a large majority brought their own tools rather than waiting for an official rollout. The behavior came first. Policy and training followed. That pattern is uncomfortable for a responsible professional, but it is not new.

Spreadsheets entered many offices the same way. Accountants and analysts bought early software with their own budgets, brought it to work, and used it before information technology departments had a rollout plan. The people who adopted quietly did not merely save time. They learned how the work changed when calculation became cheap. Formal training arrived after the early adopters had already made adoption inevitable.

My own first useful AI lesson came from a video. I learned that a broad request such as “summarize this” gives the system almost no guidance about what matters. I applied that lesson to scientific literature reviews my job already required.

The first result was generic. The next attempt became a chain. I defined the purpose of the review, the methods, the key endpoints, the secondary endpoints, what the results meant, and the implications for colleagues. Building and testing that chain took two to three hours on my own time and with tools I paid for myself.

Using AI was not in my job description. The description named the output and said nothing about the method. That silence is where a professional either builds a capability or does not.

The process saved time and made frequent summaries practical. It did not remove the need to read the article or judge which findings mattered. I shared the method with my team. Some people got good results. Others did not adopt it and continued doing the work manually. A working artifact handed to someone for free is still not adoption. Practice is the part nobody can hand over.

The modern organization may be further behind than the individual. McKinsey found only one company in a hundred reporting that AI had reached maturity in its operations, while executives underestimated how much their own employees were already using it. Waiting for the institution to become certain means waiting for a group that is trying to learn the same lesson at a larger and slower scale.

Start does not mean break policy. It does not mean place protected information into an unapproved system. It means use what is approved, work with public or synthetic material when necessary, and begin building the mental models that transfer. The path can respect the rules without surrendering the initiative.

The cost of delay is compounding. A peer completes one real project, learns from it, and attempts another. The next project begins from experience rather than zero. Six months of small attempts becomes a body of judgment. The person who waits begins the first attempt when the peer is already refining a method.

By the time formal training arrives, it is often designed for people at the starting line. The early practitioners may be the people asked to help design it. The difference was not permission. It was time in motion.

Waiting also changes the emotional cost of beginning. How ordinary will your first attempt look a year from now? Today it looks like everyone else’s, because everyone is still learning. The same attempt feels exposing after peers have accumulated a year of practice. Delay does not protect confidence. It raises the stakes attached to the first move. Starting responsibly now keeps the learning private, practical, and proportionate to the work.

Nobody is coming. That is the bad news, and it is also the opening.

What you can do now

Choose one approved tool and one task that already belongs to your role. Use public, synthetic, or properly approved information. Define the output you need, the audience who will use it, and the standard it must meet. Then complete the task with AI and compare the process with the way you did it before.

Do not wait for the organization to design the perfect course. Ask what support already exists, follow the rules that protect the organization, and build inside those boundaries. Initiative and responsibility belong together.

The first goal is not a dramatic productivity claim. It is evidence. You need to see where AI helped, where it created work, what required judgment, and what you would change next time. That evidence becomes the start of domain-specific capability.

Share the lesson, not merely the artifact. A completed output may save a colleague time once. Showing the choices behind it can help them build a capability of their own. Explain the goal, the sequence, the point where human expertise mattered, and the check that prevented a weak result from moving forward. That is how individual initiative begins becoming organizational capacity.

Premise 2Expertise Is Never Finished

Learning AI is not a course you take. It is expertise you never stop developing.

Professionals often treat AI as a body of knowledge that can be completed. Take the course. Earn the certificate. Learn the current interface. Return to work. That model fails because the technology changes, the work changes, and expertise only develops through repeated use.

The deeper obstacle is often misdiagnosed. Resistance to AI is rarely just a knowledge gap. It is a threat response. It touches competence, autonomy, status, and belonging. More information does not resolve a threat. Experience does.

Albert Bandura identified four sources that shape self-efficacy, the belief that we can succeed at a task. Three of them point directly to a practical way forward. Complete a real win. Make experimentation emotionally safe. Watch peers succeed and learn beside them.

Practice on small, real projects.

The fastest path to useful capability begins with one small task from your real work, carried from beginning to end. Small is not a concession. It is the design. The first attempt is sized so it can be finished, examined, and used as the foundation for another attempt.

Real matters as much as small. Adults learn differently when the problem belongs to them. Malcolm Knowles built adult learning theory around this distinction. Adults resist abstract instruction imposed without context. They respond to problems connected to immediate goals and responsibilities. A generic exercise teaches a feature. A real project teaches a professional how the feature behaves inside their world.

Bandura explains why finishing matters. Mastery experience is the strongest source of self-efficacy. It is not encouragement and it is not positive thinking. It is evidence gathered by the learner that a difficult action can be completed. One finished project changes the belief carried into the second project.

New pilots do not log their first hours in a jet during a storm. They fly a small, forgiving aircraft on a calm day. The conditions are chosen so the learner can complete the cycle and understand what happened. Nobody mistakes the smaller plane for the limit of the pilot’s ambition. It is the correct first environment for building the capability the larger plane demands.

My first practical win was an ordinary email. I chose it because I already knew how to do the work and the output was specific. I knew who needed the message, what I wanted to say, and what the right tone sounded like.

The task usually took twenty to thirty minutes of typing, editing, and wordsmithing. With AI supporting the draft, it took about ten. That was typical for me, not a promise to anyone else. The content held because I supplied the substance and could judge the result.

The important decision came before the first instruction. I did not choose a complex workflow. I chose work I understood well enough to evaluate. That small finished result gave me evidence that I could do the next thing.

Choose a task with a visible beginning and end. Summarize a public report. Compare three options against criteria you already use. Turn meeting notes into a first draft of a project update. Create questions that expose gaps in a plan. Keep sensitive information out until the method and the rules support it. The task should matter enough to be real and remain small enough to finish.

Starting with an ambitious system feels serious. It often produces the wrong lesson. The project stalls, the professional interprets the stall as evidence of poor aptitude, and the next attempt becomes harder. A series of courses can create the same trap. Knowledge accumulates while the belief required for action does not.

One finished thing beats ten started things. The reason is not discipline. Finishing changes what you believe you can do next.

A finished bounded project creates something more useful than an artifact. It creates a reference point. The next time a similar task appears, the professional begins with a method, a known risk, and a standard for comparison.

The first win is small on purpose. It is the only one that has to be.

Make time for play, not just practice.

Real learning moves in a loop. Try something. Observe what happened. Extract the lesson. Try again. The loop only keeps running when a failed attempt is cheap enough to study.

There are two ways to run that loop, and professionals reliably run only one. Practice is experimentation with a result already named. You know the output, you know the standard, and you are working toward both. Play is experimentation with no result in mind. You are finding out what the tool does rather than making it do something you already decided on.

Practice takes care of itself. Work arrives with a deadline and a defined output, and a professional will always find the hours for it. Play is the one nobody schedules. It has no deliverable, so it loses every calendar contest it enters.

Running only practice changes the conditions. When every attempt is aimed at something that counts, a weak result becomes a verdict. The learner protects competence by narrowing the experiment, hiding the failure, or stopping. None of those responses indicate a lack of ability. They indicate that every available hour was spent on work being judged.

Play is a deliberate technique for defusing that threat response. It is not permission to be careless with live work. Software engineers test code in a sandbox because errors there are cheap, visible, and reversible. The sandbox is not a lack of seriousness. It is the discipline that makes serious deployment possible.

I learned the cost of getting this wrong with Jasper, an early AI writing tool that charged credits for every generation. Each attempt visibly spent a limited resource. What if the instruction was poor? What if the answer was useless? What if the credits disappeared before I learned anything valuable?

So I rationed. I tried small prompts, guarded the credits, and never went far enough to discover what the tool could do. The caution intended to prevent waste became the waste. Weeks passed with almost no meaningful practice.

More information did not change the pattern. A real project did. A class required me to finish something that counted twice, as work my job actually needed and as the condition for completing the course. The stakes went up, not down. That is what got me past the meter. I had to spend the credits, so I stopped counting them.

What came afterward mattered more. Having finished one real thing, I could afford to play. The confidence did not arrive by lowering the stakes. It arrived by clearing a bar I cared about.

That is the part most professionals have backwards. Practice is not the enemy of play. Practice is what buys the confidence that makes play affordable. But practice will always take the hour, and play never asks for one.

David Kolb described experiential learning as a cycle that begins with concrete experience, moves through reflection and explanation, and returns to active experimentation. A single pass is activity. Repeated passes become learning. Jack Mezirow added that genuine perspective change requires a safe place to examine assumptions after an experience disrupts them.

AI provides an unusually good sandbox when the stakes are controlled. Ask for three approaches to a problem and inspect where they differ. Change one part of the instruction and observe what moves. Give the same task different context. Deliberately ask for a weak version, then define what would make it better. The goal is not the artifact. The goal is seeing cause and effect.

This kind of experimentation should be separated from the work being judged by a manager or client. Use public information, old material, or a synthetic example. Save the useful methods. Discard the rest. The professional can be rigorous without making every experiment visible.

You are not bad at this. You have been grading yourself on a first draft.

When every hour goes to practice, people narrow their use to tasks where success is almost guaranteed. That protects the appearance of competence and prevents competence from growing. Play reverses the incentive. A weak result becomes useful information. The professional can ask what changed, adjust one variable, and run the loop again without turning the result into a judgment about identity.

Practice will always take the hour. Play never asks for one.

Learn faster with peers than alone.

Capability transfers through people. Curricula can organize knowledge, but much of the knowledge that makes AI useful never appears in a course. It lives in the choices a practitioner makes while doing the work. What context they include. What they verify. What they ignore. When they stop the tool and take control.

Bandura called one part of this vicarious experience. Watching someone like us succeed raises our belief that we can succeed. The effect is especially important when the barrier is fear. A peer does not only transfer a technique. They transfer permission.

Technology adoption research reaches the same conclusion from another direction. The Unified Theory of Acceptance and Use of Technology names social influence as a central driver of whether people intend to use a technology. Peer influence is not a soft extra. It is part of the mechanism that moves knowledge into behavior.

Medicine has learned this lesson in high-stakes settings. New surgical techniques spread through proctoring. A surgeon who has performed the procedure stands beside one who has not. Journal articles carry the evidence, but observation and guided practice carry the technique. The profession learned that written knowledge alone could not transfer everything the hands and judgment needed to know.

After struggling with several tools, my obvious next move was to try another tool. Instead, I joined a program that combined a course with a community. I became one of the least advanced people in the room and treated that as the benefit.

One peer showed me how to build a presentation through a sequence of instructions with human review between the steps. I applied the method to real work. A presentation that took roughly two hours came down to roughly one. The result was useful, but the transferable value was the method. I had watched someone slightly ahead make the process visible.

The community did more than answer questions. It made capability observable. Watching peers work through the same uncertainty transferred both technique and permission.

Find one person slightly ahead, not a distant celebrity presenting a polished system. Ask to see one real workflow. Ask what failed before it worked. Ask what they verify every time. Share one of your own attempts and invite them to point to the part you cannot yet see.

A small peer group can make this routine. One person brings a task. One demonstrates a method. Everyone names what transferred to their own work. The meeting does not need a curriculum. It needs real artifacts and enough trust for people to show the weak versions.

You will not read your way into expertise. At some point you will watch someone do the work and realize the capability was available to you too.

Isolation hides the part that matters most. Published examples show finished output. Peers can show the failed instruction, the missing context, the verification step, and the moment they took control back from the tool. That is where transferable judgment lives. Without access to that process, a learner can imitate the artifact while missing the capability that produced it.

The fastest route runs through one person who is slightly ahead of you.

What you can do now

Pick a task you can finish in one sitting and already know how to judge. Keep the stakes low enough that an imperfect result teaches you something instead of creating harm. Complete the whole cycle. Try, review, adjust, and try again.

Then show the work to one person who is slightly ahead of you. Ask them to identify one choice they would make differently. Do not ask for a complete education. Ask to see one real method applied to one real piece of work.

Repeat that pattern until experimentation feels normal. Courses can provide useful structure, but expertise grows through completed work, thoughtful reflection, and contact with people who make invisible choices visible. The goal is not to finish learning AI. The goal is to build a learning practice that does not finish.

Protect that practice from evaluation before it is ready. Keep early experiments away from sensitive information and high-consequence decisions. Use material that lets you examine weak results honestly. Once the method is reliable, move it into real work with the review and approval the work deserves. Confidence grows from evidence, not from pretending the risk is lower than it is.

Premise 3Build What Outlasts the Tools

Tools are temporary. Capabilities last indefinitely.

Regular use is not the same as capability. Many professionals use AI as an answer engine. They ask, receive, and move on. That can be useful. It is not fluency, and it does not become durable because it happens every day.

Tool-centered learning forces a professional onto a treadmill. Every release creates another interface to master and another feature list to chase. Capability-centered learning builds something that survives the release cycle. It has structure, depth, and stronger thinking skills underneath both of them.

Build AI skills in stages.

Learning AI is not one skill. Research, communication, analysis, operations, workflow design, and agent direction are different capabilities. Each has its own learning curve. Each also depends on more basic capabilities being reliable.

Bloom’s revised taxonomy gives the sequence a useful spine. Remembering supports understanding. Understanding supports application. Application creates material for analysis. Analysis makes evaluation possible. Evaluation supports creation. The levels can move quickly when the learner has strong domain knowledge. They cannot be skipped without moving the missing work downstream.

The stages are compressible. They are not skippable.

Bloom’s revised taxonomy, from the lowest level on the left. Select a level, then try skipping it.

Select a level to see what it makes possible.

Source. Anderson and Krathwohl, Bloom’s revised taxonomy.

A professional may try to build an agent before learning to specify a task. The build may appear technical, but the failure began earlier. The agent cannot reliably execute work whose inputs, boundaries, and definition of good were never made explicit. The missing stage returns as review burden, rework, or risk.

Medical education proceeds in stages for the same reason. Anatomy is not the destination, yet a learner cannot skip from interest to residency. Each stage creates the raw material needed by the next. Experienced professionals already possess deep domain knowledge, which lets them compress parts of the path. Compression is an advantage. Skipping is not.

When I committed to my first substantial AI project, the tool made me slower for several weeks. I spent more time shaping instructions than the work would have taken by hand. My first result that looked successful was a wall of text. It seemed impressive until I tried to edit it. Rewriting from scratch would have been faster.

I reached a point where I wondered whether I had misunderstood how any of this worked. Then I recognized two things. This was the worst it would ever be. My skills would improve through practice, and the models would continue improving too.

By the fifth week, my instructions became more precise, the edits shrank, and hours began turning into minutes. The slow weeks were not wasted time. They were the stage everyone wants to skip and nobody can.

Begin with direct use on bounded tasks. Learn how context changes an output and how criteria improve it. Move into repeatable workflows where outputs become inputs to another step. Then move into delegation, where an agent carries work across time and makes bounded choices. At every level, retain the ability to judge what comes back.

The sequence should not become a reason to move slowly. Urgency and stages are compatible. A person can move through a well-designed progression faster than someone who repeatedly leaps to advanced work, hits an invisible gap, and starts again.

Without stages, breadth of exposure gets mistaken for depth of capability. The missing foundation eventually appears as rework. An agent built before task specification creates more supervision. A workflow built before reliable direct use hides weak instructions inside a larger system. An advanced demonstration can conceal those gaps for a while because it looks impressive. Real work exposes them. Stages prevent the professional from discovering a basic weakness only after it has become an operational problem.

The stages are compressible. They are not skippable.

Learn principles over tools.

The tool mastered this year may be renamed, rebuilt, or replaced within two. The principles underneath survive. How to define the goal. How to supply context. How to specify what done and good mean. How to verify a result. How to keep control when the work crosses a boundary.

Learning principles does not mean avoiding tools or sampling them casually. It means going deep enough on one capable tool to see the principles beneath the interface. Play describes the mode of practice. Depth describes the commitment. They support each other.

The Anthropic Economic Index found that experienced users of one system achieved higher success and wrote increasingly sophisticated instructions over time. Tenure with the tool became something more than familiarity. It became a richer model of how to structure the interaction. The specific buttons were temporary. The accumulated judgment transferred.

A woodworker who has mastered five hand tools can outperform someone who owns fifty and has used each twice. The tool count was never the variable. Depth teaches pressure, sequence, limits, and recovery. Those are the principles that make the next tool easier to learn.

My prompt library began as one long document. As the prompts became more useful and more detailed, each moved into its own file and a spreadsheet tracked where everything lived. Later, the most-used instructions moved into project folders and standing system instructions. Copying and pasting largely disappeared.

That library survived three changes in AI tools. It began in WordHero, moved to Jasper, moved again when ChatGPT arrived at the end of 2022, and settled into Claude by the middle of 2025. Every move meant a new interface, a different feature set, and a different way of holding instructions. The software changed. The structure did not. Goals, context, criteria, verification, and reusable knowledge moved from one system to the next.

The asset was allowed to start ugly. Its value came from accumulated use, not the software that happened to hold it.

Choose one capable system and use it across a meaningful range of work. Keep a record of what produced a reliable result and why. Separate principles from product behavior. A principle sounds like “state the decision criteria before asking for options.” Product behavior sounds like “click this menu.” One belongs in your professional method. The other belongs in temporary notes.

Depth also reduces exhaustion. A person who understands the principles does not need to chase every launch. They can ask whether a new tool changes a capability that matters. Most releases do not. When one does, the underlying method provides a place to put it.

Without that depth, every product announcement feels like an emergency. The professional restarts, learns another interface, and loses the record of what worked before. The cycle creates constant motion without accumulation. Principles provide continuity. They let someone evaluate a new feature against a stable question. Does this improve work that matters, or does it merely give me something new to try?

Depth is the only investment in this space that no release can depreciate.

Grow your thinking. Not just its application.

Where do you think judgment comes from? Most people treat it as a fixed trait a professional either has or lacks. It is downstream of thinking skill, and thinking skill can grow. AI makes that growth more important, because production is no longer the main constraint.

Thinking operates in three plain places. It selects the project worth doing. It defines what done and good mean before the work starts. It evaluates what comes back and decides how to redirect it.

AI can support all three, but it cannot remove the need for them. If the wrong project is selected, speed magnifies the mistake. If quality is undefined, a fluent answer can pass without being useful. If review is weak, a plausible output moves downstream carrying errors that become harder to see.

Dell’Acqua and colleagues, studying 776 professionals, found that individuals using AI could match the output of two-person teams working without it. The machine changed the amount one person could produce. It did not make every person equally effective. Once the tool becomes broadly available, the variance moves to the human side.

Freestyle chess demonstrated the same pattern after computers surpassed human players. The strongest results did not always come from the strongest grandmaster or the strongest machine. Teams with a better process could beat both. The engine was not the advantage. The way the human structured the partnership was.

I tested that distinction while preparing a substantial piece of writing. I did the research, built the outline, defined my writing style, and documented the language I avoid. AI produced draft chapters that I could easily have released.

I did not release them.

The ideas and frameworks were mine, but the language did not consistently sound like me. I read the work out loud. That exposed phrases and rhythms that looked acceptable on a screen but felt wrong when spoken. The manuscript needed roughly three rounds of revision before it represented the standard attached to my name.

The win was not skipping review. It was doing three focused rounds of review instead of writing every sentence from scratch. AI accelerated production. Judgment decided what could leave the room.

Bloom’s revised taxonomy adds a level that professionals often overlook. Metacognitive knowledge is awareness of our own thinking. It notices how a decision was framed, where confidence came from, and what standard was used. A professional grows thinking by making that process visible, inspecting it, and changing it on purpose.

Keep a decision record for important AI-assisted work. Write the project chosen, the standard for good, the output accepted or rejected, and the reason. Over time the record shows recurring weaknesses in specification and review. It turns experience into a curriculum built from your own work.

Plausible output makes this weakness difficult to notice. The language is polished, the structure looks complete, and the answer arrives quickly. A professional without a defined standard can accept it because nothing appears obviously wrong. Better thinking creates the standard before the output arrives. It lets speed serve judgment instead of replacing the pause where judgment should have happened.

The professional becomes better at applying thinking skills they never strengthen. Those skills become the ceiling on everything built above them.

The tool was never the advantage. Your thinking was.

What you can do now

Stop counting the tools you have tried. Choose one capable system and use it deeply enough to recognize the principles beneath it. Keep a record of the goals, context, criteria, and verification steps that produced dependable work. Those lessons belong to you. The interface does not.

Name the capability you are developing. Research, analysis, communication, workflow design, and agent coordination are different forms of work. Identify your current stage in one of them and practice the next stage without pretending the foundation is complete.

For important work, write down why you accepted or rejected the result. That one habit turns review into learning. It strengthens the part of the partnership every professional still owns. Selecting what matters, defining what good means, and deciding what can move forward.

Review that record after several projects. Look for repeated failures in how tasks were selected, specified, or evaluated. Those patterns reveal the capability that needs development next. A new tool may be interesting, but a recurring weakness in judgment is more important. Work on the weakness first. The improvement will transfer to every tool that follows.

Premise 4Integrate AI Into the Work

Using AI is not adding it to work. It is integrating it into work.

The mandate turns here. The first three premises establish the requirement, the method, and the capabilities. This premise names what to build.

The sequence has four stages. Add AI to work you already do. Redesign the work around AI. Direct agents that carry the work. Build safeguards at the same speed as access. These are not four options. Each stage creates the conditions for the next.

Four stages, each one the condition for the next

Select a stage to see what it asks and the test that tells you it is done.

Add AI to the workflow you already have.

Write down the actual steps of one recurring process. Then decide where AI can research, draft, compare, classify, or challenge.

You cannot redesign work you have never written down.

Redesign workflows around AI.

Ask which steps disappear, which can run together, which should move earlier, and where human judgment adds enough value to remain.

Would the workflow still run without AI? If yes, you adopted. You did not integrate.

Develop agent creation and coordination skills.

An agent needs more than a task. It needs a workspace, boundaries, approval gates, and a destination for reviewed work.

If this agent stopped running tonight and nobody told you, what would go undone tomorrow?

Build safeguards as fast as capabilities.

Give the narrowest access needed. Separate drafting from sending. Require approval for irreversible actions.

Build the safeguard at the same moment you grant the access.

From Premise 4 of this mandate.

Add AI to the workflow you already have.

Most daily users still work in answer mode. They ask a question, receive a response, copy part of it into the real work, and move on. That is AI beside the work. The first step toward integration is to place it inside a workflow the professional can describe.

Write down the actual steps of one recurring process. Begin with the trigger. End with the person or system that receives the result. Include the unofficial checks, the handoffs, the waiting, and the corrections that everyone remembers but nobody documented. Then decide where AI can research, draft, compare, classify, or challenge.

Process mapping before automation has a long history in manufacturing. Firms that automated an undocumented process often automated their own dysfunction and made it run faster. The map was never the boring preliminary task. It was the point where hidden work became visible.

I applied this idea to an unglamorous process many professionals recognize. Annual development and self-review forms arrived as a set of online questions. The thinking mattered, but typing and wordsmithing inside small boxes created friction around work that already carried a deadline.

I copied the questions into an AI-guided interview and supplied existing documents that contained some of the answers. Then I spoke through the remaining questions and pasted the reviewed result back into the form. The process took about half the time.

The larger gain was reduced stress. Administrative work can feel secondary while remaining non-negotiable. Removing the friction returned attention to the parts of the job that needed it. This was still a bolt-on. It did not transform the organization. It demonstrated what becomes possible once a workflow is made visible.

A six-month randomized study of 7,137 knowledge workers across 66 firms illustrates the ceiling of adoption without redesign. Regular users of an AI assistant spent 3.6 fewer hours a week on email, a 31 percent reduction. Meeting time did not significantly change and total document output did not move. The tool improved one task while the surrounding process remained intact. The gain was real and modest.

Email got faster. Meetings and document output did not move.

Change for regular users of an AI assistant over six months. 7,137 knowledge workers at 66 firms.

Time on email−31%
Time in meetingsNo significant change
Documents producedNo change

Source. Dillon, Jaffe, Peng, and Cambon, Early Impacts of M365 Copilot, a six-month randomized field experiment.

This stage is supposed to be modest. The workflow is still initiated and directed by a person throughout. The goal is to learn where AI helps, where it creates review burden, and where a step exists only because the old process required it. The map creates the evidence needed for redesign.

Most professionals have never written down a workflow they own. They know how to perform it and cannot yet articulate all of it. That gap matters because a process cannot be redesigned while it remains tacit. AI exposes the gap by asking for the instructions the professional never had to state.

Stopping here creates a readiness illusion. Frequency rises. Confidence rises. The work remains the same. Repeated tasks feel like fluency because the interaction becomes comfortable. Yet the professional still begins every task, transfers every result, and remembers every next step. The tool reduced friction without carrying work. Redesign, agents, and governance all depend on a workflow somebody can see, so none of them can start while the workflow lives only in your head.

You cannot redesign work you have never written down.

Redesign workflows around AI.

Using AI inside an existing process and rebuilding the process around AI are different acts. The test is binary. If the workflow still runs without AI, it was not redesigned around AI.

Factory electrification provides the exact historical parallel. Early plants replaced a steam engine with an electric motor and connected it to the same central drive shaft. Productivity barely moved. Decades later, plants were rebuilt around distributed electric power. Machines could be placed in the order the work required rather than near a shaft. The productivity gain came from redesign, not from the new motor alone.

Here is the honest result when I apply the test to my own work. In the corporate environment, nothing I have built passes. Every AI use there remains an aid to an existing human process. The reason is not a lack of capability. Employed workflows include other people, required formats, policies, and established definitions of the job. A bolt-on can be built by one professional. Redesign usually requires permission and participation across a system.

Outside that environment, two processes pass. One is a standing intelligence digest that finds material, organizes daily and weekly notes, extracts themes, and identifies useful distinctions. Doing the theme extraction alone would take about a working week. The final weekly review would take another ten to fifteen hours.

The second is a structured multi-perspective review. Several AI roles argue competing positions while I judge the result. There is no manual version. One person cannot hold genuinely competing views and independently adjudicate among them.

Those processes do work that would otherwise be absent. That is integration. The corporate answer remains incomplete, and naming that limitation matters. Redesign is real, reachable, and harder inside an organization than most AI advice admits.

An AI assistant bolted onto an unchanged process is the electric motor on the old shaft. It may make one step faster. It does not change the architecture. Redesign begins by asking which steps disappear, which can run together, which should move earlier, and where human judgment adds enough value to remain.

The redesigned workflow keeps a human in the loop at specific points rather than everywhere. The first point is specification. A person defines the goal, context, boundaries, and standard before AI begins. The second is review. A person judges what comes back before it moves downstream. Other checkpoints are added where the consequence of error demands them.

This structure is not a compliance ritual. It produces better work. Research on metacognitive laziness, including work by Fan and by Shen and Tamkin, finds that AI can improve immediate output while the user’s underlying capability erodes. The erosion is difficult to detect because the artifact looks better. Other interaction patterns prevent that loss. The difference is design.

BCG finds that organizations producing stronger returns from AI invest more heavily in restructuring work than average adopters. That is the corporate version of the same lesson. Technology acquisition does not create integration. Work must be reorganized around the new capability.

Most of the value from AI comes from changing the work

How BCG splits the value organizations get from AI.

10% Algorithms 20% Technology 70% Workforce change

Source. BCG, AI Transformation Is a Workforce Transformation, 2026.

A redesigned process may not function without AI, but it must still function under human control. The professional can explain why the task exists, what the system is allowed to decide, what evidence supports the result, and when a person must intervene. Dependence on capability is not surrender of accountability.

Without redesign, a professional gets a frustrating outcome. Adoption is real, satisfaction may be real, and the volume or quality of final work barely changes. They conclude the technology was overhyped. The evidence is accurate. It describes the old process, not the limit of the tool.

Redesign also protects human capability. If AI handles every step while a person watches vaguely from the side, immediate output may rise while the ability to judge it fades. Specification and review prevent that drift. They make the human contribution explicit at the beginning and the end, with additional checkpoints wherever the consequence of error requires them.

Would the workflow still run without AI? If yes, you adopted. You did not integrate.

Develop agent creation and coordination skills.

An agent carries work rather than answering a request. Building and directing one is a management skill before it is a technical skill. The essential acts are delegation, specification, review, exception handling, and accountability. Managers already know the shape of this work.

The transition resembles the move from individual contributor to first-time manager. Skill at doing the work earns the opportunity. It does not automatically produce skill at directing the work. The new role asks for clarity about outcomes, boundaries, resources, and quality. Prompting alone cannot carry that responsibility.

I learned this while using an AI assistant to prepare a vendor document from a template. I asked the assistant to review the template and interview me for the missing context. I expected the conversation to produce an approved plan before anything touched the deliverable.

Instead, the assistant began writing directly into the document. The deliverable became the workspace, and the machine took control of the browser window to make changes. The output looked productive because it appeared in the final artifact. The process took about two hours on work I could have completed in about twenty minutes.

The request had been reasonable. The coordination design had not. Nobody specified where drafting would happen or where approval had to occur. The review was not ignored. It had no defined place in the process.

An agent needs more than a task. It needs a workspace, boundaries, approval gates, and a destination for reviewed work.

Begin with a job, not an impressive demonstration. Define the result the agent owns, what starts the work, what information it can use, what it may decide, and when it must stop. Define where the output goes next. If the result has no downstream use, the agent is producing activity rather than capacity.

That distinction is the honest caveat. Many agent systems create artifacts that a human rewrites, rechecks, or quietly ignores. They appear productive while adding supervision. This is not a reason to avoid agents. It is the reason to build them around work that can genuinely leave a person’s queue.

Use one diagnostic. If this agent stopped running tonight and nobody told you, what would go undone tomorrow?

If the answer is nothing, the agent is a demonstration. If a report fails to reach a decision-maker, a record fails to update, or a review queue stops moving, the agent holds real work. The professional can then measure reliability against an outcome rather than against the amount of content produced.

Coordination that once required meetings, reminders, and status collection can move into a system. Control stays with the person who defines exceptions and watches the right indicators. The dashboard does not remove management. It changes management from chasing information to acting on what needs judgment.

The distinction becomes visible downstream. An activity-producing agent fills a folder with drafts that someone must rewrite or quietly ignore. A capacity-producing agent completes a bounded job that another process can use. Its value is not how busy it appears. Its value is that reliable work happens and a human no longer carries every mechanical step.

Stop asking whether the agent is working. Ask what stops when it does.

Build safeguards as fast as capabilities.

A bad email may cost an apology. A deleted database can cost a company. Exposed financial or employee information can do lasting harm. As AI gains access and agents gain authority, risk grows faster than visible capability.

The answer to more capability was never less capability. It was a discipline that made the capability survivable.

In 1935, the prototype that became the B-17 crashed during a demonstration flight. The aircraft was more capable and more complex than what pilots had flown before. The response was not to abandon the aircraft or simplify it until the advantage disappeared. Pilots developed a preflight checklist that made the complexity manageable. The B-17 went on to fly an extraordinary record without a similar accident.

The most instructive AI error I have seen was not a fabricated number. Every figure was real and every one appeared in the source. The system reported figures for one endpoint under the heading of another.

That error is quiet. A fabricated number is loud and gets caught. A correct value under the wrong label can survive an obvious check and travel with a professional’s name attached. Catching it required reading the source and understanding which endpoint mattered.

The failure produced an unwritten rule. AI could extract information and raise questions, but a qualified professional still had to read, verify, and judge the source. The tool removed much of the mechanical work and returned attention to critical thinking. It never received authority over the conclusion.

When I was asked what happens if someone skips that verification, I did not have an example. We do not skip it. That is not even a question.

Safeguards should rise with the level of access. An assistant working from public information needs verification and clear quality criteria. A workflow touching internal data needs approved systems, limited permissions, and traceability. An agent that can change records, send messages, or trigger another system needs stop conditions, monitoring, and a person who owns the outcome.

An agent can hold the work. It cannot hold the blame. Every delegation increases the accountability carried by the person who authorized it. That accountability should shape the design before access is granted, not after the first incident reveals the blast radius.

Build controls at the moment capability expands. Give the narrowest access needed. Separate drafting from sending. Require approval for irreversible actions. Preserve a record of inputs, decisions, and outputs. Test failure conditions with synthetic data. Decide who receives an alert and what they can do when one appears.

This is governance as professional discipline, not governance as paperwork. In regulated work, ambiguity is not a creative feature. The standard is knowing where information came from, what changed, who reviewed it, and what evidence supports the decision. Less regulated work benefits from the same clarity before the consequences force it.

Without safeguards, access expands through a series of reasonable small steps until the total exposure is something nobody would have approved as one decision. The first serious incident becomes the story leaders use to restrict AI for everyone. Capability stalls because trust was assigned before it was earned.

Safeguards also preserve credibility. One untraceable result can make leaders distrust a useful process they do not yet understand. A visible standard changes that conversation. The professional can show what the system received, what it produced, who reviewed it, and why it moved forward. Trust then rests on evidence and disciplined practice rather than enthusiasm.

Build the safeguard at the same moment you grant the access.

What you can do now

Map one workflow you own from trigger to final handoff. Mark each place where information enters, a decision is made, work changes hands, or quality is checked. That map will show where AI can help and where human judgment cannot disappear.

Begin with assistance. Learn what the system does reliably. Then test whether the workflow can be redesigned around that capability. If an agent will carry a task, define its workspace, permissions, approval gates, stop conditions, and downstream destination before it begins.

Build the safeguard during the same design session. Decide what must be verified, what action requires approval, and who owns the result. Integration does not mean removing people from the work. It means placing human attention where judgment, accountability, and consequence require it.

Measure the result by capacity rather than activity. Ask whether a meaningful task left someone’s queue, whether quality remained visible, and whether the people involved can explain the system when it fails. More drafts, messages, and dashboards do not prove integration. Reliable work moving through a governed process does.

The Closing ChallengeBecome the AI-Powered Professional

The AI-Powered Professional.

An AI-Powered Professional is an expert in their domain and an expert at using and managing AI for the work they do in that domain. The expertise is specific. An attorney does not need to build manufacturing agents. A program manager does not need to master legal analysis. Each needs to understand how AI changes the work they already own.

This identity was earned across thirteen principles before it was named. It is not a label placed on ordinary tool use. It describes a professional who can give AI a question, then a task, then a workflow, and then responsibility inside a governed system.

They redesign work with AI at the core while maintaining human control and judgment. They choose what matters, define what good means, direct the work, and answer for the outcome. They become the C-level executive of their own department even when the title does not change.

That is what the unannounced promotion was preparing you to become.

The manager in the opening had been given access and a mandate without a method. That interval between access and instruction belongs to whoever moves into it.

We do not need universal mastery. We need domain experts who can direct new capability without surrendering the standards their professions require. We need people who can move faster without confusing speed with value. We need people who can build more without losing sight of what deserves to be built.

This is not a call to become a technologist. It is a call to become more capable in the profession you already chose. Your domain knowledge gives you the context AI lacks. Your judgment gives the work a standard. Your accountability gives the output a responsible owner.

Those qualities do not become less important as AI improves. They become the difference between more output and more value.

The first move requires no permission, no new budget, and no title change.

Choose one workflow you own. Write down every step from the moment it begins to the moment another person or system receives the result. Mark where AI can research, draft, compare, classify, or challenge. Mark where human judgment must remain. Finish the map this week.

That map is not the destination. It is the first visible act of a different professional identity.

The mandate has arrived. Start now.

Bruce Fleck, PhD

The Promotion Nobody Explained


Source notes

The mandate keeps citations out of the reading flow. These notes identify the principal research, theories, and historical cases used in the text.

  1. BCG, AI at Work 2026. Survey of more than 11,000 workers across 14 markets.
  2. Microsoft, Work Trend Index 2024. Workplace use and bring-your-own-AI adoption.
  3. Section, AI Proficiency Report. Access and training gaps inside organizations deploying AI and agents.
  4. Stanford research on employment patterns among younger workers in AI-exposed occupations.
  5. Harvard research using roughly 200 million job postings to examine changes in junior and senior hiring.
  6. Arvind Krishna, Wall Street Journal interview, 2025. IBM’s automation of work associated with several hundred human resources roles, and the reinvestment into software engineering, marketing, and sales. Quoted through secondary coverage rather than a primary read.
  7. Radiology after the 2016 automation prediction. Reported by Deena Mousa, Understanding AI, and by Fortune, 4 May 2026, which carries the residency, compensation, and vacancy figures.
  8. McKinsey, Superagency in the Workplace, 2025. Roughly one company in a hundred reporting AI maturity, and executives underestimating how much employees already use the tools.
  9. Albert Bandura. Self-efficacy theory and social learning theory.
  10. Malcolm Knowles. Andragogy and problem-centered adult learning.
  11. David Kolb. Experiential learning and the cycle of experience, reflection, conceptualization, and experimentation.
  12. Jack Mezirow. Transformative learning and critical reflection following a disorienting experience.
  13. Venkatesh, Morris, Davis, and Davis, 2003. The Unified Theory of Acceptance and Use of Technology.
  14. Anderson and Krathwohl. Bloom’s revised taxonomy of cognitive processes and knowledge types.
  15. Anthropic Economic Index. Findings on tool tenure, success, and increasing sophistication of use.
  16. Dell’Acqua and collaborators, NBER Working Paper 33641, The Cybernetic Teammate. Study of 776 professionals working alone, with AI, and in teams.
  17. Dillon, Jaffe, Peng, and Cambon, Early Impacts of M365 Copilot. Six-month randomized field experiment across 7,137 knowledge workers at 66 firms.
  18. Research on AI use, metacognitive laziness, and skill transfer, including Fan and collaborators and Shen and Tamkin.
  19. BCG, AI Transformation Is a Workforce Transformation, 2026. Value from AI splitting roughly 10 percent algorithms, 20 percent technology, and 70 percent workforce change.
  20. Historical cases used as analogies include spreadsheet adoption, factory electrification, surgical proctoring, freestyle chess, and the B-17 checklist.

Find your starting point

See where you are with AI today and what to practice next.

Take the assessment
Mandate 02Stop the SlopThe causes and cures of AI work slop.