Every CEO now explains AI in tasks. We named the discipline: task intelligence.
Last week I demoed our platform to an industry analyst. Near the end, they told me something honest: the research report they are planning will probably not use the words task intelligence, because buyers do not know the category yet. Fair. Their job is labels buyers recognize. Then, in the same week, I watched the CEOs of Nvidia, Anthropic, and Microsoft explain what AI does to work. All three of them explained it in tasks.
Three CEOs who disagree on everything, except the unit
Nvidia says jobs survive. Anthropic says half of entry-level white-collar jobs are at risk. Microsoft says work moves one level up. Listen closely and they are all counting the same thing.
Jobs are safe. Tasks change.
"A job is a bundle of tasks. Certain tasks get automated. The job transforms, it does not go away."
His standing rebuttal whenever someone says AI is wiping out jobs.
Half of entry-level white-collar jobs at risk in 1 to 5 years.
"You automate 90% of the job, people are 10 times more productive in the other 10% because they are 10 times more leveraged. But eventually it gets close to a hundred percent."
When challenged on Bloomberg that he was conflating tasks with jobs, his defense was that his essay spends five pages on exactly that difference.
Work gets reconceptualized one level up.
"Compressing of workflows, completing of tasks. That is where a lot of the value gets created."
His proof story: the Azure networking team told him they do not need head count, they need tokens.
On Bloomberg's The Circuit, the host put Jensen Huang's criticism to Dario Amodei directly: that he was conflating tasks with jobs. Think about what that exchange actually is. The sharpest public disagreement in AI, between the company building the chips and the company building the models, is an argument about whether someone got the relationship between jobs and tasks right.
Nobody argues about whether the task is the right unit anymore. They argue about what happens to the bundle.
And while I was finishing this piece, a fourth moment arrived. Mustafa Suleyman, the CEO of Microsoft AI, had told the Financial Times in February that most white-collar tasks would be fully automated within 12 to 18 months. The quote blew up as an AI-takes-your-jobs headline. This week, on The Verge's Decoder, he defused it like this: "I said 'tasks' in the quote that you've just said. So that does not mean jobs. Jobs and roles are the broader category, and tasks are the components of that." Read that again. When a CEO needs to walk back a jobs controversy, the safe ground he retreats to is the task. Even the corrections now happen in task vocabulary.
The economists were there twenty years before the keynotes
None of the three CEOs invented this framing. They borrowed it from labor economics.
Economists stopped studying technology and jobs as a whole-job question in the 2000s. The standard framework since then decomposes every job into tasks and asks which tasks the machine takes, which it amplifies, and which stay with the person. That framework is how economists explained computers hollowing out routine office work two decades ago. It is in the footnotes of Dario Amodei's essay. It is behind Jensen Huang's rebuttal. It is the grammar of the whole debate.
Here is the part I find funny. When farming mechanized, nobody needed a framework. The threshing machine took threshing, and everyone could see it, because threshing was a visible task with a name. Office work is different. The tasks inside a financial analyst's job or a support engineer's job are invisible from the org chart. The job title is one line. The work underneath is twenty.
So when AI started landing on knowledge work, the only honest way to talk about it was to make the invisible tasks visible again. That is what every one of those CEO explanations is doing. It is also, word for word, the problem we started working on.
Why we put those two words together
Skills intelligence existed. Process mining existed. Neither answers the question the CEOs are now asking.
When we built our platform, we needed a name for the thing enterprises were missing. The skills platforms could tell you who knows Python. The process tools could tell you what your systems record. Neither could tell you what your people actually do between nine and six, task by task, and which of those tasks AI changes.
So we named the discipline after its unit: task intelligence. The frame is simple enough to draw on a whiteboard. An organization has people and processes. People hold roles, and every role is one or more tasks. Processes are workflows, and every workflow is one or more tasks. The task is the atomic unit where value is created and delivered. AI does not arrive at your company at the job level or the workflow level. It arrives at the task.
I will not pretend the name was an instant hit. On sales calls, GSI buyers called what we do virtual IT labs, because that is the shelf they knew us from. Analysts filed us next to task mining. For two years the words needed a paragraph of explanation every time I said them. This month, the CEOs of Nvidia, Anthropic, and Microsoft did the explaining for me.
What the CEO consensus leaves out
Knowing that tasks matter is not the same as knowing your tasks.
Listen to where each CEO stops. Dario Amodei, pressed on where displaced people go, said the economy will probably create places for them, and that "it's just a matter of finding them fast enough." Finding them fast enough is the entire problem. His essay proposes taxes and measurement. It never tells an enterprise how to find which tasks, in which roles, change first.
Satya Nadella's Azure story is the most useful of the three, because it accidentally shows the prerequisite. Before that networking team could say "we do not need head count, we need tokens," someone had to know what the team's work actually was: the fiber cuts, the operator emails, the repair dispatches. They could only rebuild their job as supervision of an agentic system because they could list the tasks the system would take. The reconceptualization everyone quotes was downstream of an inventory nobody talks about.
That inventory is the missing artifact in almost every enterprise AI conversation I am in. A CHRO can pull a list of job titles in one query. Almost nobody can pull the list of tasks inside those titles. So AI plans get made at the title level, pilots get pointed at whole jobs, and a year later the bot is doing the easy slice while the people absorb the rest.
Here is a small example of why titles mislead. In the O*NET standard, a software developer scores around 8 out of 10 for AI exposure. Then we pull a real automation engineer posting from a real manufacturer, run the actual job description through task classification, and it comes out 5 out of 10, because the specific tasks that company needs are different from the generic bundle. Same family of title, different work. The difference only shows up at the task level.
So we built the map
If the task is the unit, the task inventory is the asset. Here is ours.
Every task in that set is classified three ways: automate, augment, or human-only. Across industries the split keeps landing near 30/40/30, and the interesting part is never the ratio. It is which tasks sit in which bucket for your organization, because that is what decides where agents go first, where people need AI alongside them, and where the role concentrates afterward.
Anthropic publishes something adjacent from the other side: their Economic Index reports how their models are actually used, task by task, split between automation and collaboration. The lab measures tasks flowing through the model. We map the tasks inside your organization before the model arrives. Both sides of the market have settled on the same unit.
If you run an organization, the question got simpler
The debate about jobs will run for years. Your move does not depend on who wins it.
Notice that Jensen Huang's optimism and Dario Amodei's warning prescribe the same first step. If jobs transform task by task, you need to know your tasks to manage the transformation. If half of entry-level work is at risk, you need to know your tasks to see which half. The disagreement is about the destination. The map is required either way.
We run this on ourselves. We have mapped our own delivery workflows into their tasks, found the ones no job description owned, and rebuilt around them. We do it live with customers: upload a job description, watch it decompose into classified tasks, argue with the classification, and leave with a learning plan tied to the tasks that change. That argument, by the way, is the best meeting in the whole process. People defend their work task by task, and that is exactly the conversation that never happens at the job-title level.
The words finally have famous company. The work of applying them, role by role, task by task, is the part the keynotes leave to you. That part is what we do.
Frequently asked questions
Tasks, jobs, and what the CEO consensus means for your organization.