The split happens at the task level. Tasks that follow clear steps and need to happen the same way every time go to AI. Tasks where the right answer depends on reading a situation, carrying a relationship, or owning the outcome if something goes wrong stay with your people. That distinction rarely maps to a job title.
Why the job title is the wrong unit of analysis
An account manager might spend 40 minutes on a client call where everything depends on reading the room, then another hour copying notes from that call into CRM fields. Those two tasks have almost nothing in common from an AI-fit standpoint. The call requires listening, adapting mid-sentence, and picking up what the client left unsaid. The field update follows the same template every time, for every call, all year. Treating them as a single unit produces a single answer, and that answer is almost always wrong in one direction or the other.
The more productive question is: within this role, which specific tasks does a person handle better than AI, and which ones does AI handle better than a person? That question has an answer for almost every role, and the answer is almost never all-or-nothing in either direction.
What AI handles well inside a role
- High-volume repeatable work: updating records, formatting reports, routing incoming requests, moving data between systems. Same inputs, predictable outputs.
- Availability-dependent tasks: alerting on deadlines, sending acknowledgments when a form is submitted, monitoring conditions you have defined and flagging when they change.
- Consistency-dependent processing: applying the same classification logic to hundreds of items, extracting the same fields from incoming documents, checking a set of items against a known pattern.
- First-draft generation: converting notes into a structured brief, generating a draft from a completed form, populating a template from data already in your system for the team member to review and send.
What stays with your people
Any task where a wrong answer has consequences that are hard to reverse stays with a person. Client negotiations, pricing decisions on unusual deals, complaint handling where the relationship history matters: these require reading the specific situation and being ready to own the outcome when something goes wrong. That combination of judgment, accountability, and relationship is the persistent core of what your people are there to do. Handing the mechanical, repeatable work to AI is what clears the path to it.
A four-question test for any task
- 01Does it follow clear, repeatable rules, or does the right answer depend on reading this specific situation? If clear rules, AI is a strong candidate.
- 02If the AI gets this wrong, how easy is it to catch before it reaches someone it should not? The harder the error is to spot, the more a person needs to stay close to it.
- 03Does the context this task needs live in your systems, or in someone's head and memory? AI works from what it is given. If the knowledge has never been written down, the task is not ready to hand over.
- 04Does someone need to own the outcome if this goes wrong? If accountability must land on a named person, that person needs to be in the loop, even if AI does most of the work.
How to structure the handoff
Start by writing down exactly what the task involves: the inputs, the steps, and what a finished output looks like. Documenting your processes at this level of detail is what most teams skip, and it is the step that determines whether any handoff works. AI cannot take over a task that has never been articulated: the ambiguity that makes it hard to describe also makes it hard to delegate to a new employee. Once the task is written, look for the moments where human judgment enters: the steps where the person does something other than follow the obvious next action. Those are the steps to keep or to build a human checkpoint around. The rest can move to AI.
When one task is running cleanly, extending the same approach to another workflow comes faster, because the team member has already learned to read the AI output, knows where it needs a second look, and can explain the setup to a colleague. The first handoff takes the most effort and teaches you the most about where your task definitions were still vague.
How you know the split is working
The augmentation is working when the team member's day contains less mechanical work and they are spending more time on the decisions that actually require them. The AI error rate on the handed-off tasks stabilizes as the definition gets tighter, and the person stops thinking of it as an experiment because it has become a reliable part of how the work gets done. The split has gone wrong when the team member spends more time correcting AI output than they used to spend doing the original task directly. That is almost always a signal that the task boundary was drawn in the wrong place, and the fix is to pull the handoff back one step.
Frequently asked questions
How granular should I get when breaking down a role?
Granular enough to name the inputs and outputs of each task separately. A useful unit is roughly an hour of work or less: small enough that you can say clearly whether a task produces the same output every time or requires reading the situation first. Roles that stay blurry at this level usually need a process documentation pass before any AI work starts.
What if a team member feels like they have become a reviewer for AI output?
That reaction usually signals the split is in the wrong place. Heavy review makes sense for high-stakes tasks where catching an error before it reaches someone matters more than saving time. For lower-stakes tasks, the right move is often to narrow the review scope or let verified AI output go directly to the output without full review. The goal is for the team member's time to move toward harder work, not toward reviewing output that does not actually need their attention.
Can an entire role eventually move to AI?
Some roles shift almost entirely: basic scheduling coordination, first-line document processing, routine status updates. Most roles have a persistent core that stays human because clients or colleagues expect a person to own it. The clearest test is whether accountability can genuinely transfer. If something goes wrong, a human needs to be on the hook for it, even if AI did most of the work.
How do I introduce this without making it feel like a restructuring?
Start with one task inside one person's role, and frame it as making one specific part of their job easier. Choose something where the time savings will be obvious to that person, not just to you as the owner. When it works, the team member usually becomes the one who asks what else can move over. The change spreads from inside the role rather than being imposed on it.
What is the most common mistake in this process?
Handing over a task before it has been fully defined. The version of the task that exists in someone's head is almost always more nuanced than what gets written down in the first pass. AI surfaces every ambiguity in the definition because it cannot fill gaps the way a trained person does. When the first AI output is incomplete or wrong, that is useful information about what the task description was still missing, not evidence that AI cannot do the work.
If you want to map which tasks in your team are ready to hand to AI and which should stay with your people, we can walk through it with you.