A decision framework for managers and founders: decompose the role into tasks, price the oversight honestly, and keep a named human accountable for every result.
Every open role now comes with a second question attached. Before you write the job description, you have to decide whether the work needs a person at all, or whether a model plus a smaller team would do it better. This is not the same as asking whether AI can do the job. Almost any job contains tasks a model can do. The real decision is narrower and harder: which parts of this role are genuinely substitutable, which parts are not, and what it actually costs to run the substituted version once you count the oversight.
The mistake most managers make is treating the role as a single unit. A role is a bundle of tasks, and the bundle is what the org chart records, not what the work requires. To make the call well, decompose first. Our hiring guides walk through this at length, but the core move is simple: list the tasks the role actually performs in a normal week, then sort them into two piles.
Separate substitutable tasks from durable ones
The first pile is production work. Drafting, processing, formatting, first-pass analysis, routine responses, moving data from one system into another. These are the tasks where a model produces a usable output in seconds and the marginal cost of another unit is close to zero. The second pile is durable work: judgment under ambiguity, direction of other people or systems, holding a client relationship, and accountability for an outcome when it goes wrong. Models are strong on the first pile and structurally weak on the second, because the second pile is defined by owning consequences, not by producing artifacts.
The split is rarely clean, and the same job title usually contains both piles. Consider a customer service representative: the task of answering a routine billing question is substitutable, but the task of defusing an angry enterprise client who is threatening to churn is not. If you deploy AI against the whole role you break the second task. If you deploy it against only the routine tier, you free the person to do the work that actually retains revenue. The unit of decision is the task, never the title.
Ask who owns the result when it is wrong
Substitutability is a necessary test, not a sufficient one. The decisive question is accountability. When the output is wrong, who answers for it? A model cannot be accountable. It cannot be fired, cannot sign off, cannot stand in front of a client or a regulator and own the error. So every task you hand to a model still needs a named human who owns its output. If that human does not exist, you have not removed the role, you have only removed the safety net.
This is why the highest-substitution tasks often still need a person nearby. A data analyst role shows it clearly. A model can produce a query, a chart, and a plausible narrative in seconds. But someone still has to know whether the query answered the right question, whether the data was clean, and whether the conclusion is safe to put in front of the board. The production work collapsed to seconds. The accountability did not move at all.
Count the total cost, not the headcount saved
The seductive number is the salary you avoid paying, and it is also the least reliable one. The real cost of a deployed AI workflow includes the oversight time of the person checking output, the error risk when a wrong output ships unchecked, and the rework when it does. A workflow that saves ten hours of drafting but adds four hours of review and carries a one-in-twenty chance of an expensive mistake is not the bargain the headcount line suggests. Oversight, error risk, and rework are real line items even when nobody budgets for them. Add the ramp and maintenance cost of prompt design, tool integration, and ongoing tuning, and count all of it before you count the salary you saved.
A tool that multiplies a team, or a genuine replacement
Once you have the true cost, the decision usually resolves into one of two shapes, and they are very different. In the first, AI is a tool that lets a smaller team do more. The people stay, and each of them now directs work that used to take a junior a week. Headcount holds or grows slowly while output multiplies. In the second, AI is a genuine replacement: the task leaves the human org entirely and the workflow runs with a person only at the accountability checkpoint. Most roles are the first shape. Treating a multiplier role as a replacement is how teams end up understaffed and over-automated, shipping fast and correcting slowly.
Do not starve the senior pipeline
There is one trap that does not show up in any single hiring decision but compounds across all of them. The tasks that are easiest to automate, routine drafting, first-pass analysis, ticket triage, are also the tasks juniors learn on. Automate the entire junior training ground and you save a salary this year and lose your senior bench in five. Seniors are not hired fully formed. They are juniors who spent years doing the substitutable work badly, then adequately, then with the judgment that makes them durable. Remove that ladder and you are buying present efficiency with future capability. This is a company-level risk, not a role-level one, which is why the workforce planning view has to sit above individual hiring calls.
A step by step way to make the call
For one open role, run this sequence before you post the job or buy the tool.
- List every task the role does in a normal week, then mark each one substitutable or durable using the accountability test: if no human has to own the output, it is not truly substitutable.
- For each substitutable task, estimate the oversight, error risk, and rework it carries, not just the production time it appears to save.
- Decide what the remaining human work adds up to: a full role, a smaller role, or a single checkpoint. A checkpoint means you are replacing; a smaller role means you are multiplying.
- Check the pipeline effect. If the automated tasks are how your juniors learn, keep a training path even where the economics say automate.
- Write the role, or the workflow, around the durable work that is left, and name the person accountable for every model output it produces.
The framework does not tell you never to deploy AI. It tells you to deploy it against tasks rather than titles, to price the oversight honestly, and to keep a human owning every consequence. Run it that way and most roles come back as a smaller team doing more, not an empty seat. The empty seat is the rarer answer, and it is worth being sure before you choose it.
Run your structural assessment
The AI Career Index is free to take. Get your structural snapshot in 10 minutes, including your durability score, archetype, and projection.
Take the Free Assessment