Choosing the Right AI Role for Workflow Risk and Control

Welcome back. Last time, you mapped Maya’s meeting-follow-up workflow: she gathers scattered notes, decides which outcomes are confirmed, drafts an update, and verifies it before sending. That map showed where help might matter. Today you’ll decide what role AI should play at each point—assisting, recommending, or automating—based on what could go wrong and how much control Maya needs.

This is a product-design decision, not simply a choice of model capability. An AI system may be able to produce a polished follow-up, yet still be the wrong party to decide what attendees should be told. Allow about 40 minutes for the lesson and resources.

First ask whether the work calls for AI

A workflow pain point is not automatically an AI opportunity. Suppose Maya repeatedly has to find the current approval status of an action item. If her team records that status in a reliable field, a filter might solve the problem. If the status is buried in loosely written notes and chat, AI might help locate relevant passages—but it cannot make an unrecorded approval real.

People + AI Research’s User Needs + Defining Success frames the choice as solving a human problem first, then deciding whether AI should take over work or strengthen a person’s ability to do it.

Before choosing a role, name the unit of work. “Prepare the follow-up” contains several different judgments and actions. A single role for that entire workflow would conceal an important boundary: privately organizing notes is not the same as telling attendees that Mina has committed to sending a draft.

There is also a no-AI option. When an action follows clear, reliable rules, ordinary software may be more predictable. The following IBM Technology video illustrates why a product can combine rules, AI assistance, and human review rather than asking an AI agent to make every decision.

Assist, recommend, or automate: who makes the consequential choice?

The three roles describe how responsibility is divided, not three mutually exclusive technologies.

Assist means AI helps a person do the work while the person directs the task and owns the result. Maya might ask for a draft that separates confirmed decisions from pending items. She can inspect its wording, revise it, or discard it. Producing draft text is substantial help, but generating text is not the same as having authority to send it.

Recommend means AI proposes a choice for a person to evaluate. It might flag “Mina might send a draft Friday if legal approves it” as pending, show the passage that prompted that suggestion, and let Maya confirm or change the classification. The defining feature is the proposed judgment: the system is not merely writing; it is advising Maya how to interpret an item.

The running-app mockup below shows a lower-stakes recommendation pattern. It presents route categories under “Recommended,” with workout history named as the basis. The screen offers options; it does not show the app choosing a route and starting a run on the person’s behalf.

Automate means the system completes a bounded action without requiring approval for each instance. For example, with Maya’s permission, a tool might privately group copied meeting notes by date in a workspace where she can inspect and undo the grouping. Automation is not synonymous with “AI”: if dates are already in a structured field, a simple rule may do this better.

These labels describe the role at a particular decision point. A product can automate private housekeeping, recommend interpretations, and assist with writing while reserving the final send action for Maya. Nor is recommend inherently safer than assist: a persuasive but unsupported recommendation can steer a person toward an error. What matters is the decision the output influences and whether the person can evaluate it.

People + AI Research gives more detailed reasons people may welcome delegation in some tasks but want to retain responsibility in others.

Let the cost of a mistake set the control boundary

The question is not just “How often will AI be correct?” Ask what happens when it is wrong. In Maya’s workflow, misgrouping a private note may be easy to spot and undo. Incorrectly stating that Mina has committed to a Friday delivery could cause attendees to plan around a promise she never made. Both errors might begin with the same source sentence, but they have different consequences.

Four considerations help locate the appropriate boundary:

• Impact: Who could be affected, and how seriously? A private draft and a message to the whole team have different reach.
• Detectability: Will the error be obvious before anyone acts on it? Fluent writing can make a mistaken commitment harder to notice.
• Reversibility: Can the user fix the mistake, or has information already been shared and relied upon?
• Responsibility: Who is expected to stand behind the outcome? Maya remains accountable for what her follow-up tells attendees.

These considerations can change within one task. A suggested classification becomes more useful when the exact source passage is visible; a draft becomes harder to verify if it mixes supported claims with plausible additions. Likewise, “human review” is meaningful only if Maya has enough context to catch a consequential error. An Approve button on an unsupported, polished message does not by itself provide control.

The Feedback + Control chapter of the People + AI Guidebook adds a useful check: people may want to retain control not only because of risk, but because a task carries personal responsibility or preferences the system cannot easily know.

Control has a cost too. Requiring Maya to approve every harmless formatting change would make the tool laborious. The aim is proportionate control: less interruption for contained, reversible work; more opportunity to inspect and decide where an error could mislead other people. Users should also have a workable path when they decline the AI feature or its output is unusable.

Apply the roles to Maya’s workflow

Return to the illustrative note: “Mina might send a draft Friday if legal approves it.” Maya needs help finding and phrasing this item, but she also needs to preserve its uncertainty. A defensible role selection looks like this:

For the end-to-end experience, choose assist as the primary AI role: help Maya prepare an accurate update, but do not take responsibility for publishing it. Inside that experience, a recommendation can help with the difficult classification, and narrow automation can remove low-risk housekeeping. The role choice is therefore precise about both what AI does and where it stops.

This is a design hypothesis, not a finding about how real coordinators prefer to work. Research might show that Maya’s team already has trustworthy approval fields, making AI classification unnecessary. Or testing might reveal that source passages are too incomplete for the system to make useful recommendations. Role selection should be revisited when evidence changes.

Key takeaways

Choose an AI role per meaningful task or decision, not per product. Assist when AI should help a person produce work; recommend when it can propose a judgment the person must evaluate; automate a bounded action when the benefit of delegation outweighs the cost of errors and the user can intervene appropriately. Prefer non-AI rules when the logic is already clear.

For Maya, the critical boundary is between preparing a follow-up and asserting its contents to others. Risk, detectability, reversibility, and responsibility explain why those actions deserve different levels of user control. In the next lesson, you’ll turn this role choice into a focused problem statement for an AI-powered productivity experience.