Everyone’s feeling the pressure to keep up with the AI frontier. But few teams have a sensible, shared rule for what’s worth building. Leaders are high on hype, low on details. Organizations say they want innovation but are slow to provide their teams with the right tools (so people use their own). The confusion is slowing us down…and stressing us out.
My feed is full of people very confident about a future nobody understands.
The thing about algorithms: they feed you fear. And when you click the bait, they feed you more. So my feed (and maybe yours too) is a solid wall of anxious think-pieces predicting half the workforce gone in eighteen months. Meanwhile, I’m at my desk correcting Claude on things a real expert would have caught the first time.
I talk to companies that are way ahead of the pack, furiously building agentic teams at a pace the humans can’t keep up with. I also talk to companies that are tragically behind, trapped behind an IT policy machine that locks down everything but The LLM Everyone Wants to Throw Against a Wall. (You know which one I’m talking about.)
What I hear from almost everyone is that they’re tired. Tired of being scared and uncertain. But also, tired of deciding, moment to moment, when and how to use AI in their work.
The teams with access to frontier tools (i.e., tools that sit at the leading edge of what AI can do) are usually the teams with the most free rein to innovate. They get excited about building. They build. They tweak. They build more. They lie awake because they didn’t finish the cool thing they were into during the day. Then they look up, and it’s Friday, and while they’re now the proud owners of their first honest-to-god agent, it’ll take four months of running just to break even on the time. (I find myself here a lot.)
On the other end of the spectrum are teams that have just been given access to these tools or are still sitting on the sidelines, wondering how AI is supposed to help them. They’re getting little guidance except “use this,” and they’re checking that box by having Copilot draft their emails. Some have the sense (vague or real) that the world is moving faster than they are, and the ambient stress is wearing on them.
Both teams are exhausted. And even though they look different, I think they’re exhausted for similar reasons.
Now, every person is the Business Analyst of their own job.
There’s a phrase I use for what’s happening to knowledge work: everyone is now the Business Analyst of their own job.
We’re all now responsible for articulating the shape and boundaries of our role, continuously and in real-time. The decisions we make about how to use AI are really decisions about what we want our job to be—where we see ourselves adding value, and where we don’t.
The trouble is that it’s difficult to think about job design and task delegation at the same time.
Tasks are reducible to rules, rubrics, prompts, and agentic loops. The job is the outcome you’re there to achieve. The job is the reason there’s a box with your name on it on the org chart. It’s the reason they interviewed multiple candidates and picked you.
Not every person in the role, doing the same tasks, would be equally effective.
The part that makes you effective is the job.
AI can do so many tasks. Just when you think we’ve hit the frontier, it gets smarter. “Can Claude do it?” is the new “There’s an app for that.”
Teams that are actively building, augmenting, and automating are overwhelming themselves because they’re thinking solely in terms of tasks. “What can AI do?” They’re struggling to ask, in specific terms, “What should AI do?” When we confuse tasks with the job, we miss something important: the job is the logic by which we should decide what tasks AI should do.
And when teams don’t have shared discussion and understanding about where jobs end and AI begins, you get parallel reinvention, process fragmentation, and fuzzy ROI on what is built. Not to mention, a lot of exhausted people.
On the other hand, companies that lock everything down suppress any norm-building at all, leaving people with FOMO and a strong temptation to work around policy. In that world, you have the same issues as in the first set, but no security oversight.
Different failure, same root. Leaders aren’t helping their teams deliberately build a shared layer: team guidance (not just policy) that preserves autonomy and experimentation, but dumps the confusion and burnout.
Automation announcements don’t count as a shared layer
Some companies skip the hard work of delineating jobs vs. tasks altogether. They simply announce what is going to be automated.
A memo goes out. AI agents will handle X, Y, and Z. Headcount reduction will follow. The press release calls it a transformation. The people who remain understand that this new world was built by executives (and tech partners) with a pretty good grasp of the tasks but little understanding of the job.
The remaining team absorbs both the jobs left behind and the complexity of integrating AI into workflows nobody on the ground designed. Same load. Smaller team. More complexity. And still no shared layer, because everyone’s waiting for the company to tell them what’s coming next.
(So far, that hasn’t been working out well, and not just for the people who were laid off. MIT famously found that 95% of businesses see no ROI from their AI investment, while 56% of publicly traded companies announcing AI-linked layoffs saw a drop in stock price, with the average decline roughly 25%).
Why teams shouldn’t wait for organizational guidance
The strongest teams right now are working together to support individual innovation. Formally or informally, they’re building norms, principles, and shared learning. They’re respecting organizational policy, but they aren’t waiting for their organizations to spoon-feed guidance.
In these teams, individuals are free to innovate and experiment, but the culture is unapologetically practical: don’t build without a reason and a plan.
This is where leadership has real leverage. A shared layer doesn’t require a policy rollout or a governance committee. It requires leaders who are actively building norms with their teams, helping people understand where AI is most likely to have an impact. That way, “I built an agent” stops being confused with “I used AI intelligently.”
The downstream effects are concrete. When teams are learning out loud, parallel reinvention turns into parallel experimentation. The best results from each person’s work blend into tools and patterns the whole team can use.
And when people are building with a clear reason, it gets easier to stop. To be two hours into something, recognize the technology isn’t there yet, and set it down. To notice when a long-running AI conversation has stopped making you sharper and started making you lazier.
Here’s a principle to start with
Here’s my rule—the one I use on my own work and the one I think teams could start with as a high-impact common language. It’s based on Cynefin, and essentially it sorts my role into three categories:
- Simple work. These are repeatable, low-risk tasks that I never considered to be part of my value proposition.
- Complicated work. These are tasks or functions that have multiple steps, real interdependencies, lots of branching logic. Complicated work has to be done right, or there are real consequences. And usually, I need to be able to explain the logic behind what the complicated work produced. This is where the boundary between the “job” and the “tasks” gets a bit blurry, and it’s why I’m slow to let AI orchestrate it.
- Complex work. Ambiguous, high-impact work that unfolds and evolves. It’s very difficult to reduce this work to steps and rules and pass/fail rubrics. Diagnosing where a team is stuck and recommending how to get unstuck is complex work I’ll never delegate to AI.
Where AI Fits
Simple work is augmented or automated wherever possible. Agents, scheduled tasks, skills. I make sure simple work has a stable context (the rules, files, connectors it needs to do the work well).
Complicated work is assisted, not automated. I know that AI is able to do increasingly complicated things. But when 1) there are real downstream consequences for getting something wrong, 2) not all of the complications are known yet, and/or 3) I need to be able to explain the decision-making that got AI from point A to point B—it makes no sense to “set it and forget it” with an agent or automated workflow.
Complex work is the job, not the tasks. I can’t delegate this. I can’t automate it. I wouldn’t want to. Here, AI may be a thinking partner, but the decision stays mine. Nothing is prompted until I understand my base thinking first. No AI input is integrated into my work unless I’ve deeply explored the places it might be wrong.
How to resist agents and why
Resist the urge to build agents without a reason and a plan.
The slogan I keep on a sticky note: resist the urge to build agents without a reason and a plan. The people doing this right are using the right technology for the right kind of work. For some jobs an agent is a no-brainer. For many, it’s like flying an airplane twenty-five miles (possible, impressive, and almost never worth it). Use AI in the way the work actually demands.
Look. You know and I know that I could bang out an issue of Big Dumb Questions 2-3x a week easy…if I used AI. I’ve got a rich idea database, a voice honed over forty-five issues, a format that works.
The part of me that feels FOMO around the people letting agents run their life can picture a world where I write my raw ideas and let an agent draft-and-grade in a loop until it clears my editorial bar.
But here’s what I think would happen. I can explain my voice all day long, but ultimately, my voice blends with the LLM itself. That means the LLM wants to continuously pull my voice closer to the average “Universal Voice” we’ve all come to associate with AI workslop. And a grader agent would make that worse, running the work through loops of pass/fail tests until it forgets to ask: Is this even worth saying?
It’s complex work and it’s mine to own.
One more example. Have you heard of people building an agentic “board of directors?” The idea is that subagents each play executives—your CEO, your CIO, your COO. They debate strategy and decide together on one for you to evaluate.
Two problems with this. First, you’re locked out of the logic. (How did those “people” resolve their fight, what got sacrificed, what did the loop give you that staying in a manual back-and-forth wouldn’t have?)” And secondly, if everyone’s running their strategy through the same model, you risk converging on the same answer as everyone else, generated by something that doesn’t feel the stakes of any of it.
Make the decision Valuable, Easy, and Safe, together
If you’ve followed my work you know I diagnose execution drag with VES: Valuable, Easy, Safe. The same test can be applied here. If you’re thinking of building something with AI, ask:
Is it valuable? (This is worth doing now, in light of your other priorities).
Is it easy? (The effort is proportional to the value).
Is it safe? (Are you thinking about the risk AI introduces?)
If you only do one thing…
Save and share this guidance with your team, and talk about whether it’s right for you. What work is complex, and should stay with the team? What’s complicated, and of those things, what’s worth solving together? What’s a solo project worth doing? And where is the low-hanging fruit—simple things that save a ton of time and free you up to do your job.