Mandate 02

Stop the Slop

The causes and cures of AI work slop.

ContentsOpening
  1. Opening
  2. The phenomenon has a name, a frequency, and a price
  3. You are already inside it
  4. Correct content is not meaningful content
  5. Why this keeps happening
    1. Cause one
    2. Cause two
    3. Cause three
  6. Not your fault. Still your responsibility.
  7. Why reading it over more carefully cannot work
  8. The Screening Protocol
    1. Train the machine
    2. Build the skill
    3. Raise the capacity
    4. Govern the effort
  9. The protocol applied to Thursday
  10. What this actually asks of you
  11. Sources

The report is due to the leadership team before Friday morning. An analyst blocked two hours on Thursday afternoon to write it, and an urgent meeting nobody planned took all two.

That leaves under an hour. The analyst opens an AI tool and asks for the report.

What comes back is good. The numbers are right, the tables are clean, and the graphs are formatted the way leadership likes them. At the bottom, under the last table, the system adds four sentences of conclusion. They read like conclusions. They recommend a direction nobody in the organization has agreed to and nobody has validated.

It is 4:45. The analyst is reading fast, thinking about the drive home, checking the figures against the source data because figures are what usually go wrong. They check out. The report looks right because the part that got examined is right. Those last four sentences go by the way the fine print on a renewal notice goes by. Send.

Friday morning the Vice President reads the recommendation aloud and asks where it came from.

There is a pause in that room that the analyst will remember longer than anything else about the report. The honest answer is that the AI wrote it. Saying so in front of the leadership team means admitting that a document carrying your name went out carrying a position you never actually took. The analyst had spent the last hour of Thursday confirming that the numbers were accurate. The Vice President wanted to know what the numbers meant and what to do about them.

This time it costs nothing except the pause and a drive home spent replaying it. The second time is different. The second time a word gets attached to the person instead of to the document, and that word is careless. The report after that goes to somebody else.

What went out was a polished, accurate-looking document quietly carrying something the sender never actually read. Intelligence had nothing to do with it.

The phenomenon has a name, a frequency, and a price

What happened to that analyst turns out to be common enough that researchers went looking for it.

Teams at Stanford and BetterUp studied the pattern and gave it a name. They call it workslop, and they define it as AI-generated content that looks like finished work while lacking the substance to move the task forward. The name has already outlived the study. Workslop now has an entry in the Collins English Dictionary, which is a fair sign that a term has stopped belonging to the people who coined it.

Their research found that 40 percent of workers had received work slop in the previous month. Roughly one piece of content in seven landing in a typical inbox qualified. The person on the receiving end spends close to two hours interpreting it, correcting it, or doing the work over.

How common it is, and what it costs the person who receives it

40%

of workers received work slop in the previous month

1 in 7

pieces of content in a typical inbox qualified

About 2 hours

to interpret, correct, or redo each one

Source. Stanford Social Media Lab and BetterUp Labs, “AI-Generated Workslop Is Destroying Productivity,” Harvard Business Review, 2025.

Then there is the finding that should get your attention. Recipients rate the sender as less creative, less capable, and less reliable than they did before. That judgment attaches to the person who hit send. It does not attach to the model that wrote the words, and it does not come off quickly.

That is worth setting down carefully, because it is the only place in this article that points at what you have to lose. The analyst on Thursday was not a careless person. Most coverage of work slop stops at the conclusion that people are being careless with a powerful tool, and if that were the whole story the remedy would be a reminder to pay closer attention. Everyone reading this has already tried that remedy. There is a reason it has not worked.

You are already inside it

Before I go further, I should tell you where I am standing while I write this.

My first encounter with work slop was in 2021, years before anyone had a word for it. I was using a writing tool called WordHero. I gave it a direction and let it run, and what came back had the shape of a blog post with nothing inside it I could put my name on. Cutting it down to something usable took longer than writing the post by hand would have taken. Nothing went out the door, and the lesson stuck anyway. Never trust the voice of AI completely, no matter how finished it looks.

I built a working habit out of that. Outline first, write in sections, review each section as it comes rather than generating a whole draft and hoping it holds together at the end.

The habit still slips.

I have sent short posts after a fast read, telling myself the stakes were low enough to absorb a small miss. Most of the time that was probably true. It was also the same decision the analyst made on Thursday, at a smaller size and a lower price.

Twice the stakes were not low. I found AI-drafted testimonials that had quietly inflated a client result, once on a webpage and once inside a book manuscript. I caught both, and not because I was screening carefully. I caught them because the sentence did not sound like anything that client would have said.

Once I did not catch it. I published a post called My Thanksgiving Epiphany that described an experience attributed to me that I had not had. It went out under my name to a community of colleagues who read my work, and I corrected it after they had already read it. As far as I can tell, nothing came of it. No damage, no correction demanded, no relationship changed. That is the honest accounting, and I would rather give you the story without a consequence bolted onto it.

I felt the same pull writing this article, and I felt most of it in the second half, as I got tired. That is the part worth your attention more than the confession is. Most of us read the first half of a long document closely and let the standard drift in the second half. If this were only a matter of discipline, being tired would not change the odds. It plainly does.

None of this is an accusation aimed at you. If you use these tools seriously, you have your own version of the Thanksgiving post somewhere behind you. Having built real discipline around AI does not exempt you from this. It changes how often, not whether.

Correct content is not meaningful content

There is a second kind of slop that an accuracy check will never catch, and I produced it for years before I understood it.

I once put together a briefing for a colleague, and later a version of the same material for a customer. Every fact in both was correct. I had checked. What I had not checked was how much of it there was, and the honest answer is that there was far more than either of them needed for what they were about to do. I had handed two people a smaller copy of the overload I complain about receiving.

Just because content is correct does not mean the audience gets the meaning.

The model underneath that is old and worth stating plainly. Communication has a sender, a message, and a receiver, and the exchange only succeeds when the receiver walks away holding the meaning the sender intended to send. Accuracy is a property of the message. Meaning is an outcome in the receiver. Those are two different things, and only one of them can be verified by inspecting the document.

A piece of work can pass every accuracy check and still fail the person reading it. It runs long. It is pitched above or below where that reader actually is. It carries three sections of background they already had.

Think about the last benefits enrollment packet your employer sent around. It can be accurate to the last footnote, compliant with every rule that applies to it, and forty pages long, and still fail the employee who needed three sentences and a deadline. Nothing in that packet is wrong. The person still ends up picking a plan by guessing.

Most writing has always tilted toward the sender, and there is a decent reason for it. Working out what a specific reader knows, needs, and can absorb takes real time from a person who does not have any, and the deadline does not negotiate. We have all made that trade quietly and repeatedly.

Here is what changed. Holding that discipline costs a machine almost nothing once you have specified it. Tell the system who the reader is, what decision they are facing, what they already know, and what they need to walk away with, and it will do the tailoring you have been skipping since your first job.

A form letter and a personal letter have always cost the reader the same effort to read. They have never cost the writer the same effort to write. That is genuinely new, and it may be the most underused upside these tools have.

Why this keeps happening

Once you stop reading each instance as a personal failure, a pattern comes up through it.

Three separate conditions produce work slop. They act at different moments, on different actors, and they have nothing in common except that all three are present in most modern work at the same time. Only one of them lives inside the AI.

Cause oneThe machine defaults to more, and to prose, and to nobody in particular

It defaults to prose, and to more of it. Ask a general-purpose AI system for almost anything and paragraphs come back. Sometimes paragraphs are right. Often the task would have been better served by a table, a checklist, a diagram, or three bullets, and prose arrives anyway because prose is the native shape of the thing.

It also arrives in quantity, and the quantity is measurably growing. Epoch AI tracks how long models take to answer standard benchmark questions, and that length is climbing exponentially, roughly 2.2 times a year for standard models and closer to 5 times a year for reasoning models. Reasoning models already return around eight times the output of their non-reasoning counterparts. These systems are getting more verbose as they get more capable, and they are doing it considerably faster than anyone’s reading habits are adjusting.

Answers keep getting longer

Epoch AI measured answer length to standard benchmark questions growing about 2.2 times a year for standard models and about 5 times a year for reasoning models. This chart compounds those two rates over three years. It shows what the rates mean, not a forecast. The scale is logarithmic.

Show the numbers
YearsStandard modelsReasoning models
Start1×1×
12.2×5×
24.8×25×
310.6×125×

Source. Epoch AI, “LLM Responses to Benchmark Questions Are Getting Longer Over Time.” Growth rates as reported in this mandate.

It does not know who is reading. Left alone, the system writes to an average reader who does not exist anywhere in your organization. It does not know whether this particular person wants the detail or the two-line version. It does not know whether they take information in through a chart or through a narrative. It does not know how much of the background they already hold, which means it cannot work out what to leave out.

Every one of those gaps has to be closed by somebody. A gap left open does not get filled with better-aimed information. It gets filled with more words.

It does not know you. There is no default sense of your standards, your judgment, or how you actually sound on a page. Without training, the output could have come from anyone who asked the same question, which is exactly what makes it feel slightly foreign when you read it back. The entire burden of catching what falls below your bar sits with you, every time, at the end, when you have the least left.

Cause twoNobody taught us the skills these tools assume

Nobody taught us where human judgment belongs. Someone has to decide where a person steps into the work at all. At the front, framing the task. In the middle, checking the sources. At the end, validating the conclusion before it goes anywhere.

Most of us default to the last one and review only the finished product, which happens to be the most expensive place to catch anything and the least effective. By then the document already looks complete, the time is already gone, and the pressure to accept what is in front of you is at its highest.

Nobody taught us to specify. Deciding, before you generate anything, what the work is actually for, what context the system needs in order to do it, what you want back, and what done and good look like on this particular task.

This used to be called prompting. That word was always too small for the job it names. What the work actually requires is specification, a front-end discipline, and the clarity of what you ask for now matters considerably more than the structure you ask for it in.

Nobody taught us to evaluate at this volume. Judging work against a standard you are holding in your head is expensive thinking. Bloom’s revised taxonomy places evaluating near the top of the hierarchy of thinking skills, well above remembering and understanding, which is roughly where most professional training stops.

That skill became a daily requirement for millions of people at the exact moment the volume requiring it multiplied. Nobody scheduled the training.

Cause threeDemand went up, capacity did not, and resilience decides the rest

Cognitive demand. The task sets the price, and AI work priced itself at the top. Evaluating sits near the top of Bloom’s ladder, and AI work parks a person up there for most of the day rather than visiting occasionally. Reading is the plainest case of a task setting its own price. It is a complex, multi-step process running partly in sequence and partly in parallel across several parts of the brain, all of it spending energy to pull meaning out of marks on a page. A children’s book with pictures and plain words asks very little. A dense technical report in industry jargon, with no illustration and no visual break, asks a great deal and asks it continuously. Most of what these systems hand you sits nearer the second, and that last claim is the softer ground of the three causes, resting on reading theory more than on current empirical research.

Cognitive capacity. The resource is real, it is bounded today, and it moves slowly. Cognitive load theory, first proposed by John Sweller, holds that working memory carries only a handful of items at once, somewhere in the range of four to seven. Push past that and comprehension degrades before the reader notices it happening. Attention and processing speed behave the same way, finite on any given afternoon and shifting with sleep, stress, and what has already been spent. Underneath sits an older constraint, since human attention evolved to sweep a field and act rather than to process bulk text, which makes sustained evaluation a trained override rather than a default. That last part reasons from how the system was built rather than from a study.

Capacity behaves like physical conditioning. At 4:45 on a Thursday it is whatever it is, and no amount of wanting raises it in the next twenty minutes. Across months of deliberate training it genuinely moves. Both halves of that matter, because a ceiling you can raise is a different problem than a ceiling you cannot.

Cognitive resilience. What holds when demand runs past capacity, and how fast you come back. Psychology uses the term for the ability to adapt, withstand, and keep performing under challenge and acute stress. Borrowing it from a clinical literature and pointing it at knowledge work is a transfer rather than a finding, which is worth saying out loud. Two variables live inside it, how much load a person absorbs before performance drops and how fast they return to standard afterward, and most of us track neither.

What makes resilience invisible from the inside is that the standard drifts rather than collapsing. Nobody experiences a moment of deciding to stop caring. The second half of a long document simply gets read faster than the first half did, which is what happened while this article was being written, and what happened to an analyst whose resilience had been spent in a meeting before the report was ever opened. Push that far enough and overload turns into overwhelm, a two-stage distinction that is my own extension rather than a settled finding. Overload is the immediate ceiling, where the mind goes numb because what is in front of it exceeds what it can process. Overwhelm is what repeated overload becomes, an ongoing exhaustion with a physical component.

None of these three describes a lazy person. They describe a fast machine, a skill set nobody built, and a brain that was doing its job long before any of this arrived.

Go back to Thursday. That analyst was not careless. They were tired, they were rushed, and they were evaluating at a speed nobody has ever trained them for. Work slop is what comes out when all three conditions arrive at the same desk on the same afternoon.

Not your fault. Still your responsibility.

Read that heading as two separate sentences, because both are true and neither one cancels the other.

You did not choose to make these systems default to volume. Nobody sat you down and taught you work process analysis, specification, and evaluation before handing you a tool that demands all three a dozen times a day. You did not design the brain you are reading this with, and you cannot trade it in for one with more working memory.

All of that is true, and it changes nothing about what happens next.

Work slop does not sit still. When you forward something you have not actually read, the person receiving it inherits the exact decision you just made, and if they are as stretched as you were, they make it the same way and pass it along. A backlog sits in a folder and waits for you. This moves through an organization on its own, one unread handoff at a time, which is why it is the Slop Contagion rather than a pile.

Somebody pays for every skipped evaluation, and it is never the person who skipped it. The cost gets paid downstream, once per recipient, by each person performing the read that did not happen upstream. That is the Second-Read Tax. The Vice President reading a recommendation aloud on Friday morning was paying it. Forwarding unread work is an act with a price attached rather than an omission.

Here is the uncomfortable part. You feel fine while you do it. Nothing looks wrong from where you are standing, the document was polished, the tables were clean, and you are a Silent Carrier of the thing. Both of those inflated testimonials passed through my own hands while I felt completely fine. So did the Thanksgiving post.

A contagion does not stop at its source. It stops at whoever decides it does. That person is the Last Carrier, and it is a position rather than a personality.

Where does it stop?

Unread work moves down a chain. Select the person who screens it before passing it on. Everyone after them receives work somebody actually read.

4second reads paid downstream
About 8hours, at about two each
Nobodyis the Last Carrier

Illustration. Hours use the Stanford and BetterUp finding of close to two hours spent on each piece received.

The position is open to anyone, and it asks for attention rather than for perfection. In every chain there is somebody standing at the point where the thing can be interrupted, and on Thursday at 4:45 that was an analyst with under an hour and a report to send.

Why reading it over more carefully cannot work

The natural response to everything above is a resolution to read more carefully next time, and that resolution has to come off the table before anything can take its place.

It is the right impulse pointed at the wrong target.

Work slop is what the problem looks like from the outside. Three untreated conditions produced that document, and the document is merely where they became visible to somebody else. Fixing the document treats the visible part.

A fever is a symptom. Bring the temperature down, send the patient home, and you have relieved the thing you could see while leaving whatever caused it fully intact. The next fever is already scheduled.

Editing is symptomatic relief. It clears the artifact in front of you and leaves every condition that produced it running. That is why genuinely careful people still ship slop, and why it keeps arriving in your inbox no matter how many times somebody downstream cleans one up.

There is a quieter problem inside the resolution too. Editing and evaluating feel like one activity and behave like two. Editing checks work you already understand because you produced it, against a standard you are already holding while you read. Evaluating asks you to judge work you did not produce, at volume, frequently with no benchmark to check it against, dozens of times in a day. One of those is familiar and cheap. The other is unfamiliar, expensive, and untrained, which is the Bloom problem from the last section showing up in your actual afternoon.

Two of the three causes are not in the document at all. One of them lives in a machine that was never trained on your standards. The other lives in you, and in how much capacity you have left at 4:45. Re-reading the page in front of you cannot reach either one, however carefully you do it.

The Screening Protocol

Here is what goes in place of a resolution to be more careful.

The remedy is a standing practice called the Screening Protocol. It has four layers with twelve tactics inside them, and only one of those four layers has anything to do with the AI itself.

Screening means here what it means in medicine. It is a standing program that tests before symptoms appear, run as a matter of course rather than in response to a problem. A mammogram is not a reaction to a lump. That distinction carries more weight than it sounds like it does, because if you hear screening as one last look before you hit send, you will collapse all four layers into the fourth one and land right back at proofreading harder.

The four layers divide the work between them. Two control the source, which is where the slop gets generated. One raises your own resistance so the thing cannot pass through you. The fourth sets the triage and the standard of care, and it is the layer that makes all of this survivable inside a real working week.

The Screening Protocol

Four layers, twelve tactics. Select a layer to see what it answers and what it asks of you.

Train the machine

Answers cause oneActs on the AI, once, and every task inherits it

  • Memory. Context the system would otherwise rebuild from nothing at the start of every session.
  • Voice. Enough samples of your real writing that what comes back stops sounding like it came from anyone.
  • Taste and judgment. The standards you apply without ever articulating them.

Build the skill

Answers cause two, and cause one per taskActs on you, while the work is happening

  • Work process analysis. Decide where your judgment enters before you start.
  • Specification. What the work is for, what context the system needs, what you want back, and what done and good look like.
  • Evaluation. Judge what came back against the standard you hold rather than against how finished it looks.

Raise the capacity

Answers cause threeActs on you, underneath the job

  • Reading skills. Reading is trainable, and most professionals stopped training it somewhere back in school.
  • Energy management. The high-effort thinking AI leaves you costs more per hour than most people expect, and it runs out on a schedule.
  • Environment. Space and time protected for demanding thinking.

Govern the effort

Decides whether the other nine get usedActs on the decision to apply them

  • Quality mindset. A standard you hold whether or not anyone is going to check.
  • Risk-matched effort. Screen hardest where being wrong costs the most. This is the place to begin.
  • Time management. A deadline has to include the time to validate the work.
One layer is about the machineThree are about you

Layer oneTrain the machine

This layer answers the first cause, and it is the only layer that does. What you build here persists. You do it once and every task afterward inherits it.

Memory. Your projects, your organization, the client history, the working background the system currently rebuilds from nothing at the start of every session. A system already holding that context spends less of its output telling you what you already know and more of it on the part you actually needed. This does not require putting sensitive material into a tool your employer has not approved. It requires thinking deliberately about which context can be held safely and persistently inside the tools you are permitted to use.

Voice. Enough samples of your real writing that what comes back stops sounding like it came from anyone. Without this, every draft carries a rewrite pass at the end that exists purely to make the words sound like yours, and that pass lands at exactly the moment you have the least left to spend on it.

Taste and judgment. The standards you apply without ever articulating them. What counts as sufficient evidence, which claims need hedging, what belongs in a recommendation and what belongs in an open question. This is the hardest of the twelve tactics and the least solved, and it is better described as an open area than as a finished technique. Even a partial attempt reduces how often the system has to guess.

Layer twoBuild the skill

This layer answers the second cause, and it also handles whatever is task-specific about the first one. It acts on you, while the work is happening.

Work process analysis. Decide where your judgment enters before you start, rather than discovering at the end that you should have been in it from the beginning. A short internal update and a recommendation going to a Vice President do not get the same answer.

Specification. What the work is for, what context the system needs, what you want back, and what done and good look like on this task. For Thursday’s report, specification would have said to separate validated findings from open questions, and to flag anything reading as a recommendation for a human to confirm before it went anywhere.

Evaluation. Judge what came back against the standard you hold rather than against how finished it looks. Polish is the easiest thing these systems produce and the least informative signal about whether the work is right. Evaluation also means comparison, against the source material, against what your organization has already decided, and against the judgment of whoever owns the outcome. AI can assemble information. It cannot take responsibility for whether a conclusion belongs in the report.

Training and specification are different jobs, and keeping them apart is worth the effort. Training handles what is stable about you and carries across every task you will ever run. Specification handles what changes with the task sitting in front of you today. Confuse the two and you will either explain yourself forever or expect a well-trained system to guess at a brief it never received. You do not re-teach a colleague your standards every time you hand them an assignment. You do brief them every time.

Layer threeRaise the capacity

You cannot screen what you cannot get through. This layer answers the third cause, and it works underneath the job rather than inside it.

Reading skills. Reading is trainable, and most professionals stopped training it somewhere back in school. Both speed and comprehension respond to instruction, and neither improves much from simply reading more, which is the assumption most of us have been running on for decades. The point is not becoming a speed reader. The point is treating reading and evaluating as a skill with room left in it rather than as an ability that finished developing at twelve.

Energy management. AI took the low-effort thinking off your desk and left you the high-effort thinking. Deciding what matters, checking whether a conclusion is supported, weighing one tradeoff against another. That work costs more per hour than most people expect, and it runs out on a schedule. It ran out on the analyst somewhere around 4:45.

Environment. Space and time both. Cal Newport made the general case years ago for protecting the conditions that demanding thinking requires, and evaluation is exactly the kind of work he had in mind. Comparing claims, checking a source, and noticing a quiet overstatement are hard to do while switching between messages and meetings. Evaluation done in fragments, between interruptions, is evaluation in name only.

Layer fourGovern the effort

This layer decides whether the other nine ever get used.

Holding a capability and applying it under pressure are two different things. You can have all nine practices above and still send unread work at 4:45 on a Thursday, because something has to decide whether you reach for them when the clock is against you. Training fixes competence. Design is the only thing that fixes the other one.

Quality mindset. A standard you hold whether or not anyone is going to check, carried as a habit rather than as a decision you remake under every deadline. This does not mean an hour of review on every email. It means the person sending the work has a conscious standard for what this particular piece of work requires before it leaves.

Risk-matched effort. Screen hardest where being wrong costs the most, and lightest where it costs the least. This is the rule that makes the other eleven affordable, and it is the place to begin for anyone starting from nothing.

Time management. Work slop escapes under deadline pressure specifically. When there is no room left to evaluate, the evaluation does not happen, and no amount of skill in the first three layers changes that arithmetic. A deadline has to include the time to validate the work, not only the time to generate something that looks finished.

Notice the shape of what you just read. One layer is about the machine. Three of them are about you.

The protocol applied to Thursday

Take those four layers back to the analyst with under an hour and a report to send.

Layer one would probably have prevented the sentences from existing. A system trained on that organization’s standards would already know that unvalidated conclusions do not belong in a report going to leadership, and it likely never writes that closing paragraph at all.

Layer two would have caught what layer one missed. A specification that separates settled findings from open questions produces a different document. An evaluation step, even a sixty-second one aimed at the ending rather than at the numbers, asks the question the analyst never got around to asking.

Layer three is the layer that actually explains Thursday. Two hours of preparation became under one. The mental energy that evaluation requires had already been spent in a meeting nobody scheduled. The environment was a tired person reading fast with the drive home pulling at them, which is close to the opposite of what judging unfamiliar work requires.

Layer four would have caught it anyway, and not by asking a depleted analyst for more effort. A standing rule that any report reaching a Vice President gets a second set of eyes on its final page would have caught it regardless of how the day went. That is triage. More effort placed exactly where the consequence is highest, and none of it demanded at 4:45.

Any one of those four probably prevents Thursday on its own. All four running together are what make Thursday unlikely to happen a second time.

What this actually asks of you

One more thing, because twelve tactics can read like a second job without it.

You do not have to stop all of it. You have to stop the work slop that matters, which means matching the screening to what actually breaks if the work is wrong. A status update three people will skim needs one honest pass. A recommendation a director is going to act on gets the whole protocol. Everything between those two gets something in between. That proportion is what makes twelve tactics survivable in a real week instead of one more standard you quietly fail.

Laziness was never the subject here. What all of this describes is something a lot of people have been feeling without having language for it. Work is coming out of these tools that is somehow both more than what was needed and less than what was needed at the same time. More volume arriving, and less meaning arriving with it.

Three things put us here. A machine that is fast and getting faster, a set of skills nobody scheduled the training for, and a brain with limits that were fixed long before any of this showed up. Naming them buys you something a resolution never could, because each one has a practice attached to it. You can train the machine. You can build the skill. You can raise your own capacity, and you can govern how much of any of it a given piece of work is worth.

The outcome keeps everything these tools give you. It protects the quality of the work and the professional standing attached to your name at the same time, which is what the Vice President was really asking about on Friday morning. Using AI less was never the goal here, and neither was working slower.

Screen before you send. Not because the tools, the training gap, and the limits of your own attention were your fault. Because you are standing where it stops, and the decision this time is yours.


Sources

  • Stanford Social Media Lab and BetterUp Labs. “AI-Generated Workslop Is Destroying Productivity.” Harvard Business Review, 2025.
  • Collins Dictionary. “Workslop.”
  • Epoch AI. “LLM Responses to Benchmark Questions Are Getting Longer Over Time.”
  • Anderson, Lorin W., and David R. Krathwohl, editors. A Taxonomy for Learning, Teaching, and Assessing. 2001.
  • Sweller, John. “Cognitive Load During Problem Solving.” Cognitive Science, 1988.
  • Newport, Cal. Deep Work. 2016.

Find your starting point

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

Take the assessment
Mandate 01The Promotion Nobody ExplainedThe 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.