Separate AI work and human decisions in the same workflow

A&A perspective

This article is for an owner deciding which work can move to AI while retaining clear human responsibility. Broad labels such as sales support, advertising operations or information sharing hide different actions. Our proposed method lists collecting, comparing, drafting, approving conditions and sending, then identifies the recipient and consequence beside each action.

A&A perspective

Consider how the output will be used, alongside time spent. Sorting material may be easy to correct, while automatically presenting customer terms from that list changes the commitment involved. A correct upstream result does not by itself authorize the downstream action. The scope decision should therefore include the connection between the two steps.

A&A perspective

The original Hagurumi discussion asks about frequency, judgment, consequences, verification and exceptions. Here we do not turn those questions into an automatic approval score. A frequent task may still need approval if an incorrect payment is difficult to undo. Our method does not let high frequency cancel out a serious consequence; decisive constraints are checked first.

A&A perspective

Ask who is affected by an error, who can correct it, what an approver must inspect and where exceptions go. A person pressing an approval button without access to evidence is not the intended control. Define human responsibility together with the material required for the decision and the ability to defer it.

A&A perspective

Before-action judgment approves templates, audience and prohibited conditions, with a route back when those conditions change. In-process judgment approves the specific output before it leaves; define a substitute approver and the pending state when that person is absent. After-action review samples outcomes only within a scope where errors can appropriately be corrected.

A&A perspective

These are choices per action rather than one policy for the whole company. A workflow can combine upfront policy, approval of important commitments and sampling of routine processing. Every step still needs an accountable owner. When removing approval, record why it is no longer required under the stated conditions, rather than treating staff inconvenience as sufficient reason.

A&A perspective

Our proposed design keeps conditions fixed where discretionary judgment is unnecessary. For the initial comparison, use the same type of inputs and record active work for a human process and for human review of AI drafts. Include checking, correction and exception decisions alongside drafting. An exception-free short trial does not establish that exceptions will never occur.

A&A perspective

Where approval is required, check whether the person could make the intended decision with evidence. Narrow the scope if every draft is rewritten; reconsider inputs if missing evidence increases deferrals. The owner should also decide how released time would be used. Faster processing alone is not an automatic claim of lower staffing or higher profit.

A&A perspective

Start with columns for action, AI responsibility, human responsibility, stopping condition and escalation owner. Replace a vague statement that AI answers with a specific allocation: AI drafts, the responsible person decides any discount, and an authorized operator sends. Later changes can then be reviewed for where decision authority moved, not simply for how many screens disappeared.

A&A perspective

The Hagurumi questions are a starting point for discussion, not a validated scoring model. For an uncertain step, one option is to keep judgment with a person and trial only the preparation of evidence. If review burden does not improve, reconsider the reason to automate that step. The intended outcome is a sustainable way to perform necessary business work, not a higher percentage assigned to AI.