Choose the balance one task at a time—not by labeling an entire job “automatable” or “human-led.” Automate tasks when you can evaluate the result and the workflow can detect and recover from errors. Preserve human judgment where context, consequences, accountability, or the ability to oversee the system matter. Then measure how the arrangement works in practice and adjust it.
What does “AI automation vs. human-led” mean?
It is not an all-or-nothing choice. Human-AI configurations can range from fully autonomous to fully manual, with different divisions of work between them. NIST notes that oversight needs vary by system and context; some uses may not require human oversight, while others do. See NIST AI RMF 1.0, Appendix C.
A useful way to make the choice is to describe the work as tasks and outcomes. NIST’s 2024 AI Use Taxonomy: A Human-Centered Approach sets out 16 AI use activities and explains that tasks combine one or more activities. The count describes the taxonomy, not a performance measure or a recommended automation threshold. Separating tasks helps reveal where AI contributes, where people contribute, and where handoffs or review are needed.
How to decide which tasks to automate
Use these questions to assess each task. This is a practical decision framework informed by NIST and ILO guidance, not a formula or threshold issued by either organization.
Free tools Windows power users keep installed
One-click scans. No signup required.
#1 Best Overall
| Decision factor | Questions to ask | What it means for the workflow |
|---|---|---|
| Task and intended outcome | What activity is being done, and what result does the user or organization need? | Describe the work in terms of its purpose before choosing what AI should do. One workflow may contain tasks suited to different arrangements. |
| Context dependence | Does the decision rely on context that is difficult to represent as measurable inputs? | Preserve a human role when missing or hard-to-model context could change the decision. NIST cautions that mathematical representations of complex human phenomena can lose necessary context. |
| Consequences and recovery | What happens if the output is wrong? Can the workflow detect the error and recover? | Match controls and escalation paths to the consequences. Automation level alone does not manage the risk. |
| Oversight capability | Who can monitor, challenge, or intervene? Do they have the skills and authority? | Assign named roles and provide relevant proficiency and training. A person’s nominal presence does not by itself amount to effective oversight. |
| Work integration | How central is the task to the occupation, and how does automation fit into the rest of the work? | Consider whether AI removes a task, complements people, or changes what they do. The ILO identifies task centrality, integration into work, and management choices as factors in these outcomes. |
| Evidence from operation | What will you examine about quality, errors, time or effort, escalation, and overrides? | Evaluate the actual deployed arrangement and revise it based on results rather than assuming adoption or speed means success. |
How to set up a human-AI workflow
- Describe the task and desired result. Break a workflow into activities instead of deciding whether a whole job can be automated. State what a successful outcome means for the person or organization using it.
- Specify who does what. Record the AI’s contribution and the person’s contribution, including review, exception handling, communication, and accountability. NIST’s taxonomy can help establish common language for describing AI use and evaluation needs.
- Assess context and consequences. Identify edge cases, context the system may not capture, and the effects of incorrect output. NIST notes that human-AI interaction outcomes can vary, and that modeled representations of complex human phenomena can lose context.
- Assign operational roles and authority. Define who operates the system, uses its output, monitors performance, challenges a result, intervenes, and remains accountable. Make sure people have the proficiency and training needed to carry out those responsibilities. NIST’s AI RMF Playbook, Govern function recommends defining human roles, proficiency standards, training protocols, and oversight policies.
- Choose an initial configuration and evaluate it. Track outcomes that matter to the task, such as quality, errors, time or effort, escalations, and how well human review works. These are practical measures to consider, not a universal list required by NIST or the ILO.
- Revisit the arrangement. Change the division of work if the workflow is missing its intended result or if people cannot perform the oversight assigned to them. NIST’s Govern guidance recommends procedures for tracking configuration risks and outcomes.
Does automation mean jobs will disappear?
Not necessarily. The ILO explains that automating tasks can complement human labour rather than lead to redundancies. The effect depends partly on how central the task is to an occupation, how the technology is integrated into work, and whether management retains people for other tasks or oversight. Its artificial intelligence topic guidance describes those factors; they do not establish a universal prediction for a particular job or workplace.
That distinction matters when assessing a workflow: a system can take over one activity while people continue to handle exceptions, judgment, communication, or accountability. The relevant question is how the task mix and responsibilities change in the actual organization, not whether AI is present.
What to measure after deployment
Evaluate the configuration in the real workflow, not just the AI output in isolation. Choose measures that reflect the intended outcome and the cost of failure. For example, examine quality and error patterns alongside time or effort, escalations, and whether reviewers can identify and address problematic results. Track who reviews, overrides, or intervenes, and whether those actions resolve the issue. NIST’s taxonomy highlights evaluation needs, while its Govern playbook recommends tracking risks and outcomes for human-AI configurations. Neither establishes a universal performance threshold.
Keep the measures tied to the task. A faster process is not necessarily a better one if errors become harder to detect, people lack authority to intervene, or the result no longer meets the user’s needs. Use the findings to revise the task division, oversight, or escalation process.
Recommended Free Tools
Rank #3
What the guidance does—and does not—settle
NIST and ILO guidance support a context-dependent choice: define the task, consider consequences and work integration, assign capable human roles where needed, and evaluate results. They do not prescribe one correct automation percentage or a universal point at which human review should stop. The decision therefore needs to be made for the specific task and workflow.
ISO/IEC FDIS 42105 describes draft guidance on human control and monitoring across the AI system life cycle. On the cited ISO page, it is listed as under development at the final-draft approval stage, not as a published final standard: ISO/IEC FDIS 42105.
Quick Recap
Best Value
- Book - powershell for sysadmins: workflow automation made easy
- Language: english
- Binding: paperback
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
