On Monday, a customer success manager asks AI to turn a messy export into a renewal-risk analysis. It works well enough that she repeats it before the next account review. Two weeks later, her director asks her to build the analysis for the whole team.

Nobody announced a role change. No responsibility moved on the org chart. A useful experiment simply became expected work.

This is one of the quieter ways AI is changing jobs. The first use feels like resourcefulness: a person crosses into analysis, writing, design, research, or technical work that once belonged elsewhere. The important moment may come later, when the same task returns. Repetition converts a clever workaround into a workflow, and a workflow can become a responsibility before anyone decides whether it should.

Teams need a small checkpoint for that transition. By the third meaningful repeat, stop treating the work as an experiment and review the role around it.

Repetition changes the question

New OpenAI Economic Research published on September 16, 2026, examined more than 1.5 million work-related ChatGPT messages from April through July. Among roughly 6,200 workers observed consistently, previously used tasks associated with another occupation grew from 13.1% of occupation-specific AI activity in April to 25.9% in July. In a separate matched analysis, workers returned the next month to a cross-occupation task 23.6% of the time, compared with 8.4% among comparable workers who had not used that task in the prior month. OpenAI's report

The study observes ChatGPT use, not complete job descriptions, and recurrence does not prove that a manager formally assigned new work. That limitation is precisely why the pattern matters. Jobs can expand through repeated behavior before the expansion appears in a title, staffing plan, or performance conversation.

The first attempt asks, "Can AI help me do this?" The recurring task raises harder questions:

Am I now accountable for the result?
Which existing work will make room for it?
Do I understand the domain well enough to judge the output?
Does another team still own the source, standard, or final decision?
Is this developing my role, or merely filling a gap nobody has staffed?

A productivity tool cannot answer those questions. They belong to the worker, manager, and domain owner.

Use the third repeat as a trigger

Three is not a scientific threshold. It is a practical signal that the task has survived novelty and is likely to return.

At the third meaningful repeat, schedule a short role review. Do it sooner when the work affects customers, money, security, legal obligations, hiring, or safety. Do it later for trivial, reversible help such as reformatting an internal note.

Bring a compact record:

New task: Produce a weekly renewal-risk analysis from support and billing data.
Why it appeared: The analytics queue could not support the account-review cadence.
AI's role: Clean the export, group accounts, and draft explanations.
Human judgment: Confirm customer status, challenge anomalies, and decide which risks to raise.
Time and displacement: About two hours each week; quarterly account planning has slipped.
Dependencies: Finance owns paid-status definitions; data owns the approved query logic.

This record makes invisible expansion discussable. It separates the minutes saved by AI from the responsibility added around those minutes. A task may be faster to produce while still creating review, coordination, maintenance, and reputational work.

Choose one of four outcomes

The review should end with an explicit treatment for the recurring task. There are four honest choices.

Adopt it

Keep the task in the role because it fits the team's purpose and the worker wants to own it. Define the expected result, decision authority, quality standard, and time allocation. Remove or reduce something else if the work is material.

Adoption is not "keep doing what you are doing." It is a decision that the new responsibility belongs, with support underneath it.

Develop it

Keep the task provisionally while building the missing capability. Pair the worker with a domain expert, identify an approved example, and choose a review point. Development is appropriate when the task is a valuable stretch but independent judgment has not caught up with AI-assisted production.

For the renewal analysis, that might mean two sessions with a data analyst, access to documented metric definitions, and review of the next four outputs. The aim is not to turn a customer success manager into a data scientist. It is to make the boundary between useful analysis and false confidence visible.

Reassign it

Move the work to the team that should own it, or split production from judgment. The customer success manager might contribute account context while analytics maintains the query and metric logic.

Reassignment is not a failure of initiative. An experiment can reveal demand that deserves a proper service, shared workflow, or new role. The person who found the workaround has still improved the organization by exposing the gap.

Retire it

Stop the task when its value does not justify the workload, risk, or maintenance. AI makes it cheap to create recurring outputs that nobody uses. A weekly report can survive because it is easy to generate, not because it changes a decision.

Before continuing, name the audience and ask what they did differently after the last two versions. If the answer is nothing, ending the workflow may be the most productive outcome.

Review accountability, not just capability

AI often makes the production step look smaller than the job around it. A draft appears in minutes, but someone must verify the inputs, interpret uncertainty, answer questions, correct the workflow when systems change, and carry the consequences of a bad result.

That distinction matters in performance conversations. If recurring AI-assisted work is valuable, record it as part of the person's contribution. If it requires new judgment, provide learning time and access to expertise. If it adds accountability, match authority to it. Do not let the work remain visible to the organization while staying invisible in the role.

Managers should also ask a blunt capacity question: what stopped when this started? "AI saved time" is not a sufficient answer if the saved minutes immediately filled with a new obligation. Sustainable job design requires subtraction as well as expansion.

Keep a role-change log in meeting memory

Recurring work often becomes official through ordinary meetings: a manager asks for the analysis again, a team begins relying on it, or a provisional owner accepts one more cycle. Those moments are easy to lose because the recap records the task but not the shift in responsibility.

When a repeated task is discussed, capture four things in the meeting record:

whether the work is still an experiment or now expected;
who owns the output and who owns the underlying standard;
what support, review, or authority accompanies it;
when the team will reconsider the arrangement.

This turns meeting memory into evidence of how the job is actually changing. It also gives the worker a fair reference when priorities, staffing, or career development are discussed later.

AI will keep making adjacent work newly possible. That can be energizing, especially when it gives people room to learn and solve problems without waiting in a queue. The risk begins when possibility quietly hardens into obligation.

The third repeat is a useful moment to notice. Ask whether the task belongs, what it displaces, and what judgment it requires. Then make a choice the person doing the work can actually see.

Caspi supports live recap, suggested questions, contextual chat, proactive flags from connected tools, post-call action items, and persistent meeting memory. Those capabilities can help teams preserve the moment a recurring experiment becomes a real responsibility, so role changes are discussed and designed instead of merely inherited.