AI for Work
AI Should Own the Draft, Not the Decision
Cihan's view: USE AI for bounded drafts and comparisons; SKIP handing it an undefined decision with no evidence, owner, or review step.
AI Should Own the Draft, Not the Decision
Thesis: AI is most useful at work when it prepares a bounded piece of work and makes its evidence visible. It should not quietly become the owner of a consequential decision.
That sounds cautious. It is actually a productivity rule.
A good work system separates three things that are often collapsed into one prompt:
- Preparation: gather, summarize, classify, compare, and draft.
- Judgment: decide what matters, what is missing, and which trade-off is acceptable.
- Action: send, approve, change a record, spend money, or affect another person.
AI can be valuable in the first category without being trusted with the second or third.
The strongest counterargument
The obvious objection is that this creates a review bottleneck. If a person must check every output, perhaps the organization has only moved the work from typing to proofreading. For low-risk, repeatable tasks, that criticism is fair. A human approval step should not mean rereading every comma in a routine draft.
The answer is to make the boundary narrower, not to remove it. Let AI produce a recommendation with the source material, assumptions, and exceptions attached. Let a person review the parts where context or consequences matter. Automate the handoff only when the task has a clear definition of done and a reversible failure mode.
That is close to the risk-management logic in NIST’s Generative AI Profile: manage risk in context, pay attention to human over-reliance, and evaluate the system across its lifecycle rather than treating a fluent answer as proof of reliability. Anthropic’s guidance on agentic use makes a related point from the product side: more capable computer-use systems introduce new risks and need safeguards.
Neither source says “never automate.” They support a more useful question: what exactly is the system allowed to decide or change?
A practical workflow
- Name the job. Replace “review these accounts” with “classify renewal risks using these fields and flag missing evidence.”
- Show the evidence. Ask for source links, quoted inputs, assumptions, and an explicit “insufficient evidence” label.
- Reserve the consequence. Keep pricing, hiring, legal, financial, customer, and reputation-sensitive actions behind a named owner until the workflow has earned more autonomy.
A prompt can make this concrete:
You are preparing a decision brief, not making the decision. Use only the provided sources. Separate facts, inferences, and open questions. For each recommendation, show the supporting evidence and the condition that would change it. End with “insufficient evidence” where the record does not support a conclusion. Do not send, approve, edit, or execute anything.
This is not an argument for timid AI. It is an argument for useful scope. The fastest way to learn whether a workflow deserves more automation is to give it a small job, inspect the failure modes, and expand only what remains dependable.
TRY / SKIP / USE
- TRY: AI for a bounded first draft, comparison, classification, or meeting follow-up where the inputs and reviewer are explicit.
- SKIP: A vague “run the process” instruction, especially when the output changes records, money, access, or another person’s options.
- USE: Evidence-linked briefs with a clear owner, a defined escalation rule, and a reversible next step.
My verdict: USE the preparation layer; SKIP the fantasy that a confident answer is an accountable decision.
Cihan’s view: “AI should reduce the cost of preparing a decision, not hide who is responsible for making it.”
For one practical AI-for-work task each week, with a clear TRY / SKIP / USE verdict, subscribe to The Weekly Verdict.