The problem
AI has made polished analysis easier to produce. The scarce work is establishing what the analysis supports before someone relies on it.
In 60 seconds. The main risk is not limited to fabricated facts. A document can be fluent, cited, and numerically neat while its conclusion depends on a weak inference, a hidden assumption, or the wrong denominator. Decision readiness begins by testing the document against a declared use and keeping every material limit attached to the result.
AI is already part of ordinary professional work. It drafts, summarizes, restructures, researches, calculates, and prepares briefings. That can raise useful output. It can also raise the amount of polished material reaching the same finite group of reviewers.
The result is a review problem: production accelerates, while the work of rebuilding claims, tracing evidence, checking arithmetic, and testing reasoning remains bounded by time and attention.
Polish is not support
A document can pass a surface review and still fail the decision it was written to support.
Surface review can establish
- The prose is clear
- The citations exist
- The arithmetic shown is executable
- The document is internally coherent at a glance
- A human approved the final file
The decision still requires
- The source supports the claim made from it
- The calculation uses the right inputs and denominator
- The conclusion survives its important assumptions
- Contradictions and missing conditions are resolved
- The remaining uncertainty is acceptable for this use
The distinction is simple: evidence can be present without being sufficient.
Citations strengthen the basis, not the conclusion
A citation points to evidence. It does not prove that the evidence supports the sentence beside it.
The writer may have generalized from two hospitals to all hospitals, treated timing as cause, carried a pilot result into a national forecast, or divided by a narrow cost base while discussing total program return. Every source may be real. The conclusion can still outrun them.
That is why decision readiness is closer to an article with inspectable citations than to a certificate of correctness. The record gives a responsible reader more to work with. The reader still has to judge.
Three plain examples
Partly supported is different from supported
Source: A pilot reduced waiting times at two hospitals.
Claim: The program reduces waiting times in hospitals.
Finding: Partly supported. The pilot result is real for two hospitals. The wider population claim has not been established.
Correct division, wrong base
Source: A campaign returned $616,000 on the first $46,000 spent. The full campaign cost $250,000.
Claim: The campaign returned 13.4 times its cost.
Finding: The division is correct for the first $46,000. For the decision about the full campaign, the relevant return is about 2.5 times total spend.
Interest is not committed demand
Source: Customer inquiries rose 31 percent.
Claim: Demand will support a six-month expansion.
Finding: The source establishes more inquiries. The expansion case still needs conversion, orders, capacity, margin, and timing evidence.
Epistemic distance
Epistemic distance is how far a statement has moved from what the source directly establishes.
| Stage | What is added | Main discipline |
|---|---|---|
| Source and direct claim | A bounded statement about the material | Bind the exact passage, number, table, or visual |
| Synthesis | A view across several claims or sources | Preserve conflicts, scope, and missing evidence |
| Implication | A judgment about what the synthesis may mean | State the inference and alternatives |
| Scenario | Conditional future branches | Name assumptions, triggers, horizon, and branch conditions |
| Backcasting | A desired future state and proposed path backward | Separate aspiration, necessity, feasibility, and sequence |
| Recommendation | A proposed action under uncertainty | Name authority, trade-offs, reversibility, and what would change the advice |
The prose may become more specific as work moves down this table. Warranted specificity and reliability should be expected to decline unless each new step is supported and challenged. A precise recommendation may still rest on a long chain of contingent judgments.
Future-facing work can still be valuable. It should be presented as a conditional decision aid, not as a verified forecast.
Verification and the work that follows
A bounded integrity review checks one document for one declared use. Decision Grade applies only to the material and use that earned it.
Work after that review performs a different job. It may synthesize several checked documents, add research, build scenarios, test implications, or recommend action. The source material can be stronger because it was checked. The new reasoning still has to earn its own support.
The Decision Grade determination does not transfer to newly authored work. That work should become more decision-ready than ordinary prompting by preserving the evidence, assumptions, challenge, limits, and distance traveled.
Whether it is substantially better must be shown through comparative tests. It should not be asserted from architecture alone.
The operational question
Before an important document is used, ask:
What did we check, for which decision, and what remains unresolved?
A serious answer names the source version, declared use, checks performed, findings, challenges, incomplete work, outside-research needs, and the resulting use decision. A vague answer—“reviewed by AI and a human”—does not tell the decision-maker what they can safely rely on.
The goal is not more ceremony. It is a stronger basis for the next responsible act.