Where does preparation become a business decision?
The boundary is the moment an output commits the firm, changes what a client can expect, or resolves an ambiguity that requires context and accountability. Identify that moment on a recent job. Assign the decision to a person with the authority, evidence, and time to make it.
Consider an illustrative interiors studio reviewing a change request. An assistant can assemble the contract, pull the latest selections, and draft a summary of the differences. The client asked to substitute a finish, and the draft calls the change equivalent. A designer notices the proposed finish has a different lead time and changes the way two surfaces meet. Treating the swap as equivalent could commit the studio to a deadline and level of finish it has yet to evaluate.
The designer needs the underlying selections, the supplier’s current lead time, and the original promise to the client. A clean summary helps only if it makes those facts easy to inspect.
How do you test whether assistance preserves judgment?
Give the team a small set of representative examples, including awkward exceptions. Compare assisted and unaided decisions on accuracy, review effort, and the ability to explain the choice. Include a case where the evidence is incomplete and the correct response is to pause.
Brynjolfsson, Li, and Raymond found that the effects of an AI assistant varied across workers in one customer support setting. That result gives executives a reason to test effects by role and experience. A senior reviewer’s workload may change differently from a junior colleague’s, and neither experience is captured by a single average time saved.
Ask the junior reviewer to explain which evidence would change the answer. Ask the senior reviewer what they had to recheck. If the tool helps people move faster through familiar work, use some of the regained time to discuss the exceptions where the team learns its craft.
What belongs in a decision matrix?
For each decision, record the action, authorized person, required evidence, conditions that trigger escalation, and how the team proceeds if the tool is unavailable. Keep authority visible at the point where work is approved. Revisit the matrix when the scope of a tool changes.
FAQ
Can a tool send routine messages automatically?
For a bounded message, define what facts may be used, what it may promise, and when it must wait for approval. Start with examples where the consequences of a mistake are limited and reversible.
What should we measure besides time saved?
Track missed exceptions, review effort, and whether staff can explain and reproduce the decision. Check the experience of the client and the person doing the review.
Who owns a bad automated decision?
Assign responsibility before deployment. The owner needs authority to pause the workflow, correct the outcome, and change the conditions under which it runs.
Sources
- Fritjof Capra, Principles of Life, for the distinction between embodied judgment and computational abstraction.
- Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, Generative AI at Work, for observed differences in the effects of assistance across workers in a specific support operation.