The majority of the time I've been using AI tools at work, the exchange has followed the same pattern: I have a question, AI gives me an answer. But that pattern has started to change. The newest models can take an objective, work out the steps required to achieve it, carry out many of those steps, check their own work along the way, and hand back something pretty darn close to a finished deliverable.
In many ways, that’s an incremental improvement but I think it's larger than incremental because of what delegation has always required. A conventional software tool waits for instructions: do A, then B, then C. Handing over an objective instead (achieve X within these constraints) has only worked when the one receiving it could break the objective into tasks on their own. For most of our working lives, that meant the delegate had to be another person.
Case in point: Anthropic released Claude Fable 5.1 on September 1, 2026, and its product page describes the model in exactly these terms: "It plans the work, uses the tools it needs, recovers when a step fails, and keeps you posted without being asked." The company says it "handles complex, multi-stage knowledge work with minimal oversight" and that "teams can hand off large projects and review completed work rather than supervising every step" (Anthropic: Claude Fable).
That's the vendor's description, so I paid closer attention to what outside testers found. A same-day review by Katie Parrott and Dan Shipper at Every called the model a first glimpse for most knowledge workers, of "truly delegated AI work," the kind where you can assign something like a slide deck and move on to other things while it works. Their reviewers used it to build a simulation, a training pack, a survey dashboard, and an eight-stage market analysis (Every: Vibe Check: Fable 5.1). That review reflects a few practitioners' hands-on testing over a short period, and the models will keep changing. The direction is what I'd pay attention to.
My argument is that as AI learns to take an objective and return finished work, the skill that matters shifts to defining what should be built and judging whether the result is good. That shift has consequences for individual roles, and larger ones for how organizations are designed.
“Everyone's a Builder” Is Becoming Less of a Metaphor
When people can describe an application, a dashboard, or a training program and get a working version back, the hard part of building moves. More of the production work is something the machine can do. What stays with the person is the work upstream of production: deciding what should be built, why it matters, what good looks like, and whether the result is actually good.
For years, one question that separated people at work was whether they could personally build the thing. The question I expect to matter more is whether they can define what should be built and judge whether the result is good. Those are different abilities, and plenty of skilled producers will have to develop the second one deliberately.
The capabilities behind that second ability are familiar: judgment, systems thinking, curiosity, clear communication, and strategic clarity. They're often described as the things people keep because AI can't do them yet, which makes them sound defensive. I think they're becoming the controls people use to direct and check what AI produces. That's a more useful way to think about them, and a more demanding one.
The Assignment Matters More than the Prompt
Every's slide-deck example stayed with me. If you can hand a model an assignment and go do something else for an hour, you're delegating, and delegation is something executives have practiced for their entire careers.
A good AI assignment starts to resemble what a capable executive gives a capable employee:
Objective: the outcome, stated plainly
Context: what the work is for and what's already known
Constraints: limits on time, cost, scope, and sources
Resources: the data and tools it may use
Decision rights: what it can decide without asking
Quality standard: what good looks like
Verification: how the result will be checked, and by whom
Escalation conditions: when it should stop and come back to you
I suspect today's attention to “prompt engineering” will look dated before long (if not already). The skill that lasts is work design: deciding how a piece of work should be divided, assigned, bounded, and checked. Work design doesn't stay inside one person's job for long, either. Once a team can hand substantial work to AI, the organization has to answer questions it has rarely had to ask this explicitly:
Which work should people do, which should AI do, and which should they do together?
Who owns the outcome?
Where must human judgment stay in the process?
How will the work be verified?
How do roles and structures change once one person can direct a substantial amount of machine work?
That last question deserves candor. When one person can direct work that used to take a team, some roles will shrink, and a commitment to people doesn't justify pretending otherwise. It does mean being deliberate about which roles change, how people are prepared for that, and who gets to grow into the work of directing and judging.
These questions sit across organizational design, workforce strategy, leadership, governance, technology, and culture at the same time. That's the practical reason treating AI as a technology implementation falls short.
More Autonomy Makes Human Judgment Matter More
The same Every review offers a useful counterweight. A model capable of hours-long projects still made mistakes a careful manager would catch. In one test, five of 27 quotes it produced weren't in the source material. Given a 1,000-word cap, it came back at 1,288 words. One reviewer found it sometimes kept working when he interrupted to ask what it was doing; another found it could perform worse on heavy multi-tool tasks. The reviewers also noted that it will take as many steps as you let it unless you set a budget in advance (Every: Vibe Check: Fable 5.1).
None of that makes the model less useful, but it does show where the human work goes. A fabricated quote in a customer proposal or a board memo can cost far more than the hours the model saved. The more work AI can do on its own, the more depends on the decisions people make about objectives, boundaries, verification, and when to step in. When AI was limited, people had to do the work themselves. As it becomes more capable, people have to govern the work, which is a different capability and potentially a source of much greater leverage.
Why This Is a Transformation Question
Consider a hypothetical company two years from now. An executive asks an AI system to analyze customer data, employee feedback, operating metrics, market intelligence, and the strategic plan; identify the three most important transformation opportunities; model alternative operating designs; draft implementation plans; build the dashboards; and prepare the executive presentation. Suppose the system can do most of that. This is an illustration of where the technology appears to be heading.
If that happens, the technical question is largely answered, and the harder questions belong to the leadership team:
What should we ask it to optimize?
Who decides?
Can our leaders recognize a bad answer when it arrives looking polished?
Which decisions can be delegated, and what data can agents reach?
How should jobs and workflows be redesigned?
How do we keep people from automating a bad process?
How does accountability work when an agent does the work and a person owns the outcome?
Those aren't AI implementation questions. They're enterprise transformation questions, and they're the kind of questions Repeatable Transformation™ is designed to work through, because they don't get answered once. Each meaningful jump in what AI can do reopens them.
Here's what stumps me. If AI does more of the underlying work, how do people keep building the expertise they need to judge it? Many of the leaders I've worked with learned to recognize a weak analysis by producing a lot of mediocre ones early in their careers. If the next generation skips that stage, the judgment we're counting on to govern AI could get thinner over time. I'd be wary of anyone who claims to have solved that yet.
Where to Start
Start with one recurring piece of work your team produces, such as a monthly operating review or a customer proposal. Before anyone hands it to AI, write the assignment the way you'd brief a capable new hire: the objective, the context, the constraints, the resources, what they can decide on their own, what good looks like, how you'll check it, and when they should stop and ask. My expectation is that the gaps will show up quickly, often as a quality standard nobody has written down or a verification step that amounts to whoever happens to read it last.
A few questions to bring to your next leadership conversation:
Which of our recurring deliverables could AI produce from an objective today, and who would judge whether they are good?
Where do we need a person to approve before AI acts?
How will our people keep building the judgment that AI-assisted work depends on?
If everyone can be a builder, the advantage goes to organizations that are clear about what's worth building and honest about whether they built it well.
