Lately I've been questioning the assumption sitting underneath most automation conversations: that removing a task removes the work behind it. Every system I've watched scale tells me the opposite. You automate something, and the work doesn't disappear. It moves somewhere you weren't looking.
The Part That Doesn't Disappear
When a task gets automated, what actually gets removed is the visible, repeatable part of it, the part that fit inside a checklist. What stays behind is everything that checklist was quietly absorbing: judgment calls, exceptions, edge cases that never showed up often enough to write down.
Before automation, a person handled those without anyone noticing, because the decision and the execution were fused into one job. Automation splits them apart. Execution goes to the system. Judgment stays with a person, except now that person has to find it, name it, and handle it deliberately instead of by habit.
This isn't specific to AI. Any system that compresses a manual process into an automated one leaves the same residue behind it:
- A build pipeline replaces a person clicking deploy, but someone still has to own what happens when it fails at 2am.
- A recommendation engine replaces a merchandiser choosing products, but someone still has to decide what the engine optimizes for, and catch it when it optimizes for the wrong thing.
The task disappears. The decision behind the task does not. Zoom out far enough and this is just conservation of complexity: you can relocate it, but you can't delete it.
The Decision Surface
I've started thinking about this as the decision surface of a task: the set of judgment calls a task requires that never get written down, because a human absorbed them silently while doing the work. Automating a task doesn't shrink that surface. It exposes it.
Once exposed, someone has to own it explicitly, because the system won't own it for them. This matters more with agentic AI than with any earlier wave of automation, and the adoption curve backs that up. McKinsey's State of AI 2025 survey found that 62% of organizations are already experimenting with agents that plan and execute multi-step workflows on their own, not just generate text.
An agent that plans and executes doesn't make a single decision. It makes a chain of them, closer to a call stack than a single function call. Ask an agent to handle a customer refund, and it's making a sequence:
- Read the request and infer what the customer actually wants.
- Classify the situation against a policy that may not cover it cleanly.
- Decide whether the case needs to escalate to a human.
- Draft a response that matches tone, policy, and precedent at once.
Each frame in that stack is a place where the agent's judgment can diverge from what you actually wanted, and every unresolved frame gets pushed up to a human. A person supervising ten agents isn't supervising ten tasks. They're supervising a decision network with more nodes than the process ever had when a human ran it end to end. We've tracked a similar expansion in what happens to middle management and span of control when AI takes over coordination: the layer absorbing that extra network is rarely the layer companies planned for.
This is why deploying agentic systems at scale tends to increase headcount demand rather than shrink it, at least in the operational layer. The organizations that get surprised by this made a sizing error: they budgeted headcount against the old task instead of the new decision surface the agent revealed. It's the same mismatch showing up industry-wide: 99% of companies are running AI transformation, and 84% haven't restructured a single job to match it.
A team that used to need five people to write content now needs two people to write it and three people to audit five hundred variants a week for consistency, because output volume scaled faster than judgment did. The capital is already moving this direction: PwC found that 88% of senior executives plan to increase AI-related budgets over the next year, citing agentic adoption as the driver, not headcount reduction.
The deeper mistake sits one level up, in what gets optimized:
- Proxy: fewer people per task.
- Target: decision throughput, the rate at which an organization can resolve the judgment calls a system surfaces, reliably, without every case routing to the same three senior people who become the bottleneck.
You can hit the proxy and miss the target completely. Most companies that cut headcount right after an automation rollout do exactly that. They shrink the group meant to absorb the decision surface at the exact moment that surface got bigger. McKinsey estimates generative AI carries $2.6 to $4.4 trillion in annual value across 63 use cases, and none of that value gets captured without people configuring, auditing, and correcting the systems that produce it.
Beyond Automation
Once you see it this way, the pattern stops being about AI specifically. It's about any system that increases the volume or speed of output faster than it increases the organization's capacity for judgment.
Scaling infrastructure doesn't remove the need for people who understand failure modes, it multiplies the places failure can occur. Scaling a sales motion with better tooling doesn't remove the need for judgment about which deals matter, it increases the number of deals where that judgment has to get applied.
The technology is never the whole system. It's the part of the system that runs without asking permission. Everything it can't decide on its own becomes a job, and being AI-ready is really a staffing question in disguise: who owns the decision surface a deployment is about to expose, decided before launch instead of discovered during an incident.
The macro data lines up with this. The World Economic Forum projects that 22% of formal jobs will be structurally affected by 2030, but also forecasts net growth of 78 million jobs over the same period, mostly in roles that configure, audit, and interpret systems that didn't need configuring or auditing before. That's the decision surface showing up as a labor statistic.
There's a second-order effect worth naming too. Research published in Science found that access to ChatGPT raised job satisfaction by roughly 0.50 standard deviations among knowledge workers, a large effect for a single tool. Automation that surfaces the interesting judgment calls and absorbs the repetitive execution doesn't just create work. It tends to create better work.
The companies that get squeezed by automation aren't the ones that adopted too much of it. They're the ones that treated adoption as the finish line instead of the starting condition for a new layer of work. Every system that runs on its own eventually needs someone deciding what "on its own" is allowed to mean. That job doesn't show up on the org chart until the first time it's missing.
Originally published on the Linksoft Technologies blog.
Top comments (0)