Draw the delegation boundary inside a task. “AI can summarize this” does not answer who may see the source, who verifies the summary or who may act on it. Conversely, requiring a person to perform every formatting step can add work without addressing the actual risk.
Separate the stages, name the authority for each and decide what evidence makes the result usable. NIST’s human-AI guidance discusses why responsibilities must be explicit; this worksheet is a practical aid, not a legal classification or a validated risk score.
Examine the stage and its consequence
Most workflows include gathering information, transforming it, generating options, choosing, and communicating or acting. Accountability remains with the responsible person or organization throughout, even when a bounded action is automated.
Ask four questions at each stage:
- What happens if it is wrong, incomplete or disclosed to the wrong person?
- Who has the evidence, skill, time and authority to detect and reject a bad result?
- Does the interaction require someone’s attention or participation, including accessibility support?
- Is this a skill the person or team needs to keep practicing?
Do not assume that preparation is harmless. Gathering interview notes can disclose personal information; a summary can remove a condition that changes a decision. Data minimisation and the approved tool/account boundary apply before that source enters the model. Provider controls do not grant permission.
Choose one of four operating modes
| Mode | Role for AI | What makes it usable |
|---|---|---|
| Automate | Perform a bounded, authorized step | Defined scope, tested controls, monitoring and a stop route |
| Assist | Prepare a draft or comparison | A capable reviewer checks the material result |
| Challenge | Critique an initial answer or plan | The owner checks objections and evidence |
| Keep with the responsible person/process | Leave the decisive step outside the model’s role | Direct judgment, required professional process or relational attention |
These are operating choices, not permanent labels for entire occupations. A message that merely confirms a validated booking differs from one that promises a disputed refund. Explicit authorization and tested limits matter more than whether both are called “customer email.”
A worked split: interview preparation
This fictional workflow is an authored teaching example, not hiring advice or a tested selection system.
A team wants a consistent interview structure. AI receives only an approved public role description and drafts neutral questions about its stated tasks. It receives no candidate CVs, identifiable notes or inferred traits. The responsible hiring team checks the questions against its lawful, job-related process before use.
| Stage | Proposed split | Failure to check |
|---|---|---|
| Prepare questions | AI draft; hiring team reviews | Invented job requirement or inappropriate question |
| Conduct interview | Interviewers and candidate participate directly | Script prevents listening or needed accommodation |
| Record evidence | Approved note process with source attribution | A paraphrase becomes a claim the candidate did not make |
| Assess and decide | Responsible hiring process | Model ranking or speculation replaces required judgment |
A generated question such as “Describe how you checked an error in a report” can be reviewed against a role requiring report review. “Prove you can work every weekend” would add a requirement absent from the supplied role. The check concerns the source and process, not whether the question sounds professional.
Make review capable of changing the outcome
An approval button is weak oversight if the reviewer cannot inspect evidence, reject the output or stop the next action. NIST’s AI RMF Core calls for documented roles and empowered, trained people. Its generative-AI profile adds risks such as confabulated content that a plausible draft can conceal.
For skills you want to retain, schedule an unaided attempt or transfer task instead of assuming review alone supplies practice. That is a deliberate practice choice, not a claim that this worksheet measures skill loss. For relational work, use assistance that supports the person’s actual participation; accessibility tools should not be excluded merely because they help produce speech or text.
Use the stage decision card for one recurring workflow. Review it after a material source, permission, consequence or reviewer change. For a broader personal starting point, see personal boundaries for AI use.



