How New Work Works

May 14, 2026

The conversation about AI at work is stuck between two unhelpful camps.

One sells productivity. Buy the tool, save the time. The other sells dystopia. The robots are coming for everyone. Both are right enough to be exhausting and wrong enough to be useless.

The real shift is quieter, faster, and bigger than either camp admits. In the last eighteen months, AI moved from chatbot to co-worker. Systems that used to autocomplete sentences now plan, research, write, code, and execute multi-step work for hours without supervision. Ethan Mollick at Wharton has been calling this the move from copilots to agents — and the speed has caught even the optimists off guard. Frontier models from six months ago are not in the same category as the ones available today.

This is not a productivity upgrade. It is a recomposition of what work is. The shape of a knowledge worker’s day. The kind of judgment they exercise. The meetings that disappear, the artifacts that get made, the kinds of jobs that even exist a year from now — all of it is changing. Not in five years. Now.

We call this don’t work. Don’t work isn’t faster work or AI-assisted work. It’s a different relationship between people and the systems that do their thinking. Most companies don’t see this yet. The ones that do are pulling away.

The agentic shift, and why it changes everything

Until recently, the conversation about AI at work was about augmentation. A person writes a draft; the AI polishes it. A person debugs code; the AI suggests fixes. The human stays in the loop on every keystroke. Useful, but small.

The shift well underway now is agentic: AI systems that take a goal, plan a sequence of steps, and execute them autonomously. They take meetings off your calendar. They run research projects overnight. They draft proposals end-to-end and surface them for review, not for help. They monitor systems and intervene when things break. The unit of work being handed off has gone from “a sentence” to “a job slice.”

Anthropic’s Economic Index reports in early 2026 show this in real usage data, not surveys. Coding tasks have visibly migrated from augmentative use (the AI helping a developer) to automated workflows (the AI completing the work). Across the broader economy, about half of all jobs now have at least a quarter of their tasks regularly being done with AI. This isn’t a forecast. It is the current state.

This is the part most companies haven’t internalized. They are still treating AI as a productivity layer to bolt onto existing jobs. It isn’t. It is a recomposition of what a job is.

Why most companies are getting it wrong

In July 2025, MIT’s NANDA initiative published one of the most uncomfortable studies of the AI era: a comprehensive review of enterprise GenAI deployments finding that 95% of enterprise AI pilots produce no measurable financial impact. Across roughly $30–40 billion of enterprise AI investment, only about 5% of pilots delivered real ROI.

The cause, according to the researchers, was not model quality, not regulation, not talent. It was a learning gap between static tools and the workflows they were supposed to transform. Generic tools improved individual productivity but never integrated with how the company actually worked.

Three failure modes show up in nearly every company we’ve watched try to do this.

The pilot trap. A team finds a tool. Runs a small pilot. Generates a report. Files the report. Nothing changes. The pilot was always going to be the deliverable, not the change.

The IT bottleneck. AI gets routed through procurement and IT. By the time tools are approved, they’re three model generations behind. The people who could use them have moved on, and the people who got them no longer want them.

The training theater. HR rolls out an “AI fluency” workshop. People sit through it. They go back to their desks. Nothing transfers. The workshop wasn’t built for adults doing real work; it was built so the company could say a workshop happened.

The common thread: every one of these treats AI as a thing to adopt rather than a way of operating to evolve into. You can adopt a tool in a month. You cannot adopt a new operating model in a workshop.

The MIT study also turned up a finding that contradicts a lot of conventional wisdom: companies that partnered with outside builders succeeded twice as often as companies that built internally. The reason is uncomfortable. Most companies don’t yet have the internal experience to integrate AI into their own workflows. The ones that do are the ones who built that experience by partnering with someone who already had it.

There’s a related finding worth sitting with. In nearly every company surveyed, the most successful AI use wasn’t happening in official programs at all. It was happening in what MIT calls the shadow AI economy — workers using ChatGPT, Claude, or other tools without approval, because the official tools didn’t work and the unofficial ones did. Ninety percent of workers reported using personal AI tools for work; only 40% of their companies had any official AI subscription. The work is finding the AI. The companies aren’t.

What new work actually looks like

If we strip away the hype and look at the companies actually working AI-first, four shifts show up consistently.

1. The default shifts from doing to delegating.

In old work, the question was “should I do this myself or assign it to someone?” In new work, the first question is “should I do this myself or hand it to an agent?” Email triage. First drafts. Research synthesis. Calendar logistics. Meeting summaries. Lead enrichment. Status updates. These aren’t shared work anymore. They’re delegated by default — and not to a teammate. To a system.

This sounds small. It isn’t. When the default shifts, the rest of someone’s day reshapes around it. The hours that used to go to inbox triage now go to the work that actually requires judgment. The skill that matters most stops being doing-the-work and starts being directing-the-work.

2. Tools are picked at the task level, not the company level.

Old work picked tools at the company level. One CRM. One project tool. One writing tool. New work picks per task. The right model for synthesizing research is not the right model for drafting copy. The right tool for brainstorming is not the right tool for executing. The team’s job is to know which is which, and to switch fluidly.

This is why the “which AI should we buy” question is the wrong question. You don’t buy an AI; you build fluency across a stack of them.

3. Communication becomes the company's nervous system.

When agents handle the busywork, what’s left is the human work — and most of that work happens in conversation. Slack threads. Email exchanges. Decision documents. Brief calls. In new work, these aren’t byproducts of work; they are the work. The signal of where a company’s attention is going lives in its communication, not in its reports.

This is why leaders in AI-first companies are starting to treat communication as data. Not to surveil people, but to see what the company is actually working on. Reports lag reality by weeks. Communication is in real time.

4. The strategic question changes from "what should we automate" to "what should we become."

The companies pulling ahead aren’t asking “how do we automate task X?” They’re asking “what kind of company do we become if AI is the default for X, Y, and everything else?” The answer reshapes their roadmap, their hiring, their product strategy. Automation is a tactic. AI-first is a strategy.

Mollick has been blunt about where this leads: organizations that evaluate AI on free or outdated models consistently underestimate its capabilities and anchor their ambitions too low. The competitive question isn’t “what can AI do for our current processes.” It’s “what would our processes look like if we designed them around what AI can do today.”

What it takes to actually transform

Knowing these shifts is the easy part. Acting on them inside a real company is the hard part. From the engagements we’ve run, and from what the research now shows, four things have to be true at once.

Strategy comes before tools. Most companies start with “let’s buy Copilot for everyone” or “let’s stand up a custom GPT.” That’s tools-first thinking. It produces tools. It does not produce transformation. The MIT research is unambiguous: the single biggest predictor of pilot failure was lack of workflow integration. The tool wasn’t bad. The question of what it was for was never properly answered. Strategy-first means asking what this company needs to become better at, structurally, over the next 24 months. Once that’s answered, the right tools surface obviously. They become consequences of the strategy, not substitutes for it.

Education has to come with implementation. Workshops don’t transfer. Workshops with no implementation context never transfer. Companies that actually move have education fused with real work — people learning AI fluency on their actual tasks, with someone watching and coaching, not in a separate room with a slide deck. A 2025 follow-on study from Harvard Business School (the “GenAI Wall” paper) sharpened this further: AI doesn’t turn novices into experts when the knowledge distance between worker and task is too great. Productivity uplift requires enough underlying domain understanding for the human to direct and verify the AI’s work. Education isn’t optional. It is the moat between a stalled pilot and a working system. This is the piece almost no AI consultancy delivers. They build, then leave. The team learns nothing. Within six months, the build is being used wrong or not at all.

Leadership needs new instruments. If a CEO is making decisions on the same dashboards they used in 2022, they’re flying blind. The most important signal in an AI-first company — what teams are actually doing, where energy is going, what’s getting unblocked — lives in the communication layer. Leadership needs a way to see that aggregate without surveilling individuals. Without it, you can’t steer.

Culture decides whether any of this lands. The single biggest predictor of whether an AI rollout succeeds or quietly dies is whether middle managers and individual contributors believe AI is for them or against them. Anthropic’s 2026 economic survey of 81,000 Claude users found that workers experiencing the largest productivity gains from AI were also the ones expressing the most concern about job displacement. The two feelings live together. If the messaging from the top is “we’re rolling out AI to make the company more efficient,” the receivers hear “we’re rolling out AI to replace you.” The rollout dies in month four. Companies that handle this well do it explicitly. An augmentation promise to employees. Role-evolution maps for each team. Manager coaching on how to talk about AI without triggering replacement fear. None of this is soft. It’s the difference between a successful transformation and an expensive consulting bill.

How we work on this

We built Don’t Work because we kept watching companies hit one of these failure modes and assume the problem was the tool. It isn’t. The problem is that nobody is offering the integrated thing — strategy, build, education, leadership intelligence, cultural embedding — as a single coherent program.

We deliver it in four composable modules. Build is the technical work: agents, integrations, workflow redesigns. Native is the curriculum that makes every employee AI-fluent on their actual work. Compass is the leadership intelligence layer that turns communication into strategic signal. Embed is the cultural change management that prevents the whole thing from quietly failing.

Most clients need some mix. Some need all four. The math decides.

If this resonates and you want to see what your company’s version of new work would actually look like, we built a free agent that will map it for you. Five minutes per teammate. You get back a ranked roadmap of what’s worth building, learning, and changing — with timelines and costs.