Welcome to PossibNow. This is a place to understand what AI makes possible and apply it to work that matters. This week, we ask which human bottlenecks protect quality and control, and which prevent agents from doing more useful work. We look at how professionals can embrace that opportunity, keep developing their expertise and try one small change in an assignment.
Anthropic’s workplace study found that interactive Claude sessions spent most of their time waiting for human input.
Around 80% of session time went on that wait. The October 16 research examines work inside Anthropic, where employees already use agents extensively. Even there, people determine how quickly much of the work can move.
What’s Next
Humans are the bottleneck. Is that a good thing?
There is a strong argument for keeping us in that position. We need to ensure quality, question weak reasoning and control decisions that affect other people. An agent completing an assignment quickly gives us little benefit if we cannot trust what it has done. Some pauses deserve to remain.
But there is also a cost to making ourselves the point through which every step must pass. We can keep adding AI to the way we already work, asking it to help with one task and then waiting to decide the next. That approach leaves our own time and attention determining how much the whole system can attempt.
The opportunity extends beyond producing the same report a little faster. We could investigate questions we currently leave unexplored, compare alternatives we would never have time to examine ourselves and keep useful work moving between the moments when we can attend to it. Removing unnecessary handoffs gives us a way to explore that possibility.
That means accepting some discomfort. I can understand the appeal of remaining involved in every step. Each approval gives you a sense of control. It can also stop you discovering what an agent could accomplish with a larger assignment and room to work.
The sooner we embrace agents, the sooner we can make a step change in the work we attempt.
The professional question is how we make that change while keeping our expertise and our livelihoods.
You should instead put your expertise to work in choosing the problems, setting the ambition and challenging the result. Your experience helps you recognize a customer need, see why an apparently sensible recommendation would fail and identify the evidence that could change a decision. I want professionals to use that knowledge to direct not just more work, but more ambitious work than they could personally execute.
You do that by continuing to develop your skills. Read the evidence that matters, question the agent’s assumptions and work through the choices you cannot yet explain. If we stop understanding the work, we weaken the expertise we need to direct it. Embracing agents should give us more ambitious problems on which to exercise our judgment.
This connects to the small group I discussed last week, drawing on Andrej Karpathy’s rough estimates of people using frontier models professionally in math and code. I think professionals in many more fields can begin exploring that level of ambition. We have to give agents substantial work and learn from what happens when we stop directing every move.
Preserving every familiar step feels like the way to protect our professional value. But that is a stopgap at best. I would rather learn how to apply our expertise to what agents make possible. That involves getting comfortable being uncomfortable, while staying close enough to the substance to recognize when the work needs us.
Thanks for reading What AI makes PossibNow! Subscribe for free to receive new posts and support my work.
Useful now
How the Claude Code team does code review
Anthropic’s team asks which checks fit the risk of a change. It combines automated review with different requirements for human approval, then watches reversions and incidents to see whether the approach holds up. Although the example concerns software, the reasoning travels. A routine formatting check and a consequential recommendation deserve different treatment. Reducing review only helps if the work holds up after approval.
McKinsey argues that faster analysis can overwhelm the people who must understand assumptions, weigh trade-offs and reach agreement. Its suggestion to surface the assumptions and uncertainties that require a decision fits this week’s theme. The challenge is to redesign how decisions reach you as agents take on more execution. More analysis earns its place when it improves a decision or opens a useful line of inquiry.
Jan Bosch argues that abundant production makes selection more valuable. He puts particular weight on explaining why something works, so that a preference becomes judgment someone else can understand. The professional contribution involves choosing what deserves attention and explaining why. That judgment can guide an agent toward a stronger result, even when the agent produces most of the work.
Microsoft’s AI transformation of HR
Microsoft makes the case for moving HR’s attention toward diagnosis and counsel as AI handles more information gathering and process navigation. It provides a perspective beyond software. Preparing workforce information and interpreting what it means for a particular team require different contributions.
What’s possible
Mollick’s essay considers what happens as agents organize more of their own work. His examples include Codex and personal agents, extending the discussion beyond a chatbot that waits for the next prompt. For professionals, the possibility is a larger assignment with fewer instructions about execution, allowing us to reconsider work we could never staff or carry out alone. His account also leaves an open question about how well self-organizing agents handle ordinary organizational work over time.
There and Back Again, The Act of Taste
Lew Ludwig describes overlooking an error in AI-generated teaching notes because he had not read them closely enough. He knew the subject, but his expertise could not catch a mistake on a page he skipped. His essay challenges the easy version of stepping back. Explicit criteria can help us check work consistently. We still need to apply our attention where an error could pass those checks.
Optro reports that 46% of respondents said employees were spending more time reviewing, validating or correcting AI outputs. Its survey covered 417 enterprise audit, risk, compliance and security decision-makers. The vendor-sponsored, self-reported findings offer a counterweight to claims about effortless productivity. More agent output can leave people with more checking. The question is what we need to change in the assignment and review process so that greater agent capability produces useful work rather than a larger queue for a person.
Shaping global governance of AI agents
Brookings distinguishes an agent’s capabilities from the authority people give it in a particular workflow. Reading a record, changing it and communicating a commitment carry different consequences. Its argument for explicit permissions and continuing monitoring gives control a practical meaning as agents take on longer assignments. A good output alone cannot tell you whether the agent stayed within the agreed boundaries while producing it.
The loop this week
Pick a workflow you already run with your favorite AI agent, especially one where you have taught it to follow your usual method.
Start a fresh conversation. Describe the result you need in words and give the agent the context, documents and data it needs. Leave out your usual step-by-step instructions, template and finished example.
The CRAFTS structure I teach is designed to help you, the human, communicate with the AI models. You can also use a short, ordinary paragraph and see the difference. Give the agent enough information to do the job and let it choose how to deliver.
Context: [Relevant background, documents, data and sources.]
Role: Take ownership of completing this assignment.
Action: Produce [specific output] so [audience] can [achieve a goal], by
[deadline, for example 5 pm Thursday]. Look at the job from multiple
angles. Complete as much work as possible without asking me for input,
approval or human review.
Format: Choose any format, tools and workflow that best achieve the
goal.
Tone: Choose the style that best serves the goal and audience.
Safeguards: Do not assume or invent missing facts. Source your work,
flag anything you cannot verify, and check the finished output yourself
before returning it.When the result comes back, compare it with what you normally create. Was it better or worse? What did you like? What would you change? Did it give you something useful that your usual output misses?
Use that comparison to improve the next assignment. Add any context the agent needed, explain what you want to keep and change, and consider what more ambitious job you could give it next.
Thanks for reading What AI makes PossibNow! Subscribe for free to receive new posts and support my work.
Get the next one on Friday.
Applied AI for professionals, by Sal Carranza. One workflow a week, free.