Stanford's SALT Lab has published a large-scale study of how people and AI agents can work together across the U.S. workforce. Its central finding is not that every AI-actionable task should be automated. Instead, the research shows a clear worker preference: automate low-value, repetitive work, while preserving greater human agency when work is complex, consequential, or difficult to reverse. For businesses adopting AI in everyday processes, that distinction offers a practical alternative to treating automation as an all-or-nothing decision.
The study, Future of Work with AI Agents: Auditing Automation and Augmentation Potential across the U.S. Workforce, introduces WORKBank, a task- and worker-centered database built from 1,500 domain workers in 104 occupations. It covers 2,131 AI-actionable tasks and includes assessments from 52 AI researchers. The result is a framework for asking a more useful implementation question: where should an AI agent take the lead, and where should it assist a person who remains responsible for the decision?
What the SALT Lab research adds to AI adoption decisions
SALT Lab frames AI adoption through both automation and augmentation. Automation concerns tasks an AI system might perform with limited human involvement. Augmentation concerns work where AI helps a person perform better while that person retains meaningful control. The distinction matters because a workflow can be technically automatable without being a good candidate for AI-led execution.
The researchers developed the Human Agency Scale (HAS), a five-point spectrum that ranges from AI-led automation to equal human-AI partnership. It is designed to capture how much influence people want to retain over a task, rather than measuring only whether AI can perform it.
The paper also groups AI-actionable work into four decision zones: Automation Green Light, Automation Red Light, Low Priority, and R&D Opportunity. Together, these concepts help separate tasks that are suitable for automation from those where workers prefer agency, where AI support has less immediate value, or where further technical development is needed.
| Work context | Worker preference identified by the study | Practical workflow implication |
|---|---|---|
| Low-value, repetitive tasks | Greater openness to automation | Consider AI-led execution with clear operating boundaries. |
| Complex or high-stakes tasks | Greater desire for human agency | Use AI to prepare, analyze, or recommend, while a person reviews and decides. |
| Tasks needing more capable AI | Potential R&D opportunity | Do not assume current tools can safely deliver the intended outcome. |
This is particularly relevant to teams evaluating agents that can act across business systems. An agent may be able to draft a customer response, summarize a document, update a record, or route a request. But the appropriate level of autonomy changes when its output affects a customer, colleague, transaction, or decision that is hard to undo.
The research also documents collaboration patterns, moments in which AI teaches users, and productive friction in high-stakes contexts. Friction is not necessarily a defect. A pause for review, a request for clarification, or a visible recommendation can be valuable when it helps a person catch an error before an action has consequences.
How to apply human agency to AI-assisted workflows
The study is not a deployment playbook, and it does not prescribe a single operating model for every organization. It does, however, support a disciplined approach to deciding where AI agents belong. The most useful starting point is to map work at the task level, not by applying a blanket label such as “automate customer service” or “automate operations.”
A practical first pass can classify tasks according to their value, reversibility, and effect on other people. In practice, businesses can use four guardrails:
- Automate repetitive, low-value steps first, especially where the task is clearly defined and the output can be checked.
- Keep a named human decision-maker for work that has material customer, financial, legal, or employee consequences.
- Design review points before irreversible actions, rather than relying on a person to discover a problem afterward.
- Maintain an audit trail of the AI input, output, human edits, approval, and final action for workflows where accountability matters.
These are implementation choices, not claims that an AI system will be reliable in every setting. Their value is that they make the required degree of human involvement explicit. A workflow that drafts a response for an employee to approve is different from one that sends it automatically. Likewise, an agent that prepares a payment-related recommendation is different from one that executes a payment.
Human-in-the-loop is a workflow design choice
“Human in the loop” is often used as a broad assurance phrase, but SALT Lab's agency framing makes it more concrete. A person can be involved in several different ways: supplying context, choosing among options, editing an AI-generated output, approving an action, or taking over when the system cannot proceed. Those roles are not interchangeable.
For consequential work, businesses should decide in advance what the human must see and what authority that person has. If the reviewer only receives a final notification after an action is complete, the process has monitoring but not meaningful control. If the reviewer can inspect the recommendation and change or reject it before execution, the process better preserves agency.
Auditing should follow the task, not just the tool
AI audits are often discussed as broad assessments of a model or vendor. That can be useful, but operational risk usually appears in a specific workflow. The same AI assistant may be low risk when it summarizes internal notes and much more sensitive when its output is sent to customers or used to influence a consequential decision.
For each AI-assisted process, teams can document the task, the intended AI role, the required human role, the point at which an action becomes difficult to undo, and the records needed for later review. This creates a usable operating record without requiring a company to treat every AI experiment as a major transformation program.
A related scale-wide analysis summarized by alphaXiv found that a substantial share of conversations containing actionable tasks involve consequential or high-stakes work, and that dialogue tends to be longer when stakes rise. That context reinforces SALT Lab's core point: as the consequences of work increase, simple automation metrics become less informative than the quality of collaboration and the level of retained human agency.
For businesses, the immediate opportunity is not to delay all AI adoption. It is to direct autonomy toward the work employees find repetitive and low value, while building clear review and escalation patterns for the work that requires judgment. That can reduce manual effort without making an AI agent the unexamined final authority.
If your team is moving from AI experiments to repeatable workflows, Scalevise can help identify where automation can save time without removing the controls your process needs. We translate task-level opportunities into practical AI designs, including approval points, exception handling, and traceable handoffs between people and software. Explore a practical AI workflow automation project with Scalevise and request a consultation to map your highest-value opportunities.
Frequently Asked Questions
What is Stanford SALT Lab's Human Agency Scale?
The Human Agency Scale is a five-point spectrum introduced by the SALT Lab study. It ranges from AI-led automation to equal human-AI partnership and is intended to capture how much control people want in a task.
What is WORKBank?
WORKBank is the task- and worker-centered database created for the study. It draws on 1,500 domain workers across 104 occupations and covers 2,131 AI-actionable tasks.
Does the SALT Lab study recommend automating all AI-actionable tasks?
No. The study distinguishes automation from augmentation and finds that workers generally prefer automation for low-value, repetitive work while seeking more human agency for complex or high-stakes tasks.
What are the four AI work zones in the study?
The study uses four zones: Automation Green Light, Automation Red Light, Low Priority, and R&D Opportunity. They are part of its dual framing for evaluating automation and augmentation potential.
How can a business use the study in an AI workflow?
Map individual tasks, identify which actions affect other people or are hard to reverse, and set the AI role and required human review before deploying the workflow.
Conclusion
SALT Lab's research shifts the AI agent conversation from whether work can be automated to how people should collaborate with AI on specific tasks. Its Human Agency Scale and automation-augmentation framework give businesses a practical basis for prioritizing repetitive work, preserving judgment where consequences are higher, and making accountability visible in AI-assisted workflows.
