Claim Decomposition

A bold claim presented as a single statement benefits from a psychological effect: if the overall narrative sounds coherent, people tend to accept the whole package without examining each supporting premise. Decomposition disrupts this by forcing each supporting premise to stand on its own merit. A prediction like "AI will replace most knowledge work within five years" sounds plausible as a headline but decomposes into dozens of sub-claims about technology capability, adoption speed, regulatory response, labour market dynamics, task complexity, and organisational inertia — many of which are individually debatable.

AI models are effective decomposition tools because they can systematically unpack logical dependencies without being swayed by the rhetorical force of the original claim. The model does not find the narrative compelling or threatening; it simply identifies the chain of premises that must all be true for the conclusion to hold. This makes it a natural complement to assumption auditing. This mechanical impartiality is the technique's core strength.

The practical value is triage: once you have the sub-claims listed, you can identify the weakest link — the single premise most likely to be false and most damaging to the overall argument if it is. Testing that one premise is far more efficient than trying to evaluate the entire claim holistically. In many cases, the weakest sub-claim is also the one that was least examined by the original claim's proponents, precisely because it was buried inside the larger assertion.

The most effective prompt pattern, illustrated in this prompt teardown, asks the model to list the assumptions that must all be true for the claim to hold, then rank them by two dimensions: likelihood of being false and impact if false. This two-dimensional ranking surfaces the assumptions that matter most: the ones that are both fragile and load-bearing. A premise that is unlikely to fail and would not change the conclusion even if it did is not worth investigating; a premise that could easily be wrong and would invalidate everything above it is the priority.

For complex claims, hierarchical decomposition works better than flat lists. A top-level claim decomposes into three or four major premises, each of which decomposes further into supporting assumptions. This tree structure makes it clear which sub-claims are foundational (everything above them collapses if they fail) and which are peripheral (they affect confidence but do not undermine the core argument). The tree also reveals dependencies: sub-claims that share a common assumption are correlated risks.

Combining decomposition with <a href="/aisapedia/confidence-calibration">confidence calibration</a> adds another dimension: ask the model to rate its own certainty for each sub-claim. The combination of a fragile sub-claim rated with low confidence is a strong signal that the overall claim needs more evidence before being acted upon. Conversely, a claim where all sub-claims are rated as certain by the model still warrants spot-checking — model confidence is a heuristic, not a guarantee.

When an AI model produces a strategic recommendation or analytical conclusion, the output itself can be decomposed. Ask: "What facts does this conclusion depend on? Which of those facts did you verify versus assume? Where did you extrapolate beyond the data I provided?" This prompt forces the model to separate its grounded claims from its inferences, making it explicit where the analysis transitions from evidence to extrapolation.

This is particularly valuable for <a href="/aisapedia/cascading-error-analysis">cascading error analysis</a> in multi-step workflows. If step three's recommendation depends on step one's market sizing, decomposing step three reveals this dependency chain explicitly. You can then verify step one's outputs before trusting step three's conclusions — and if step one's market sizing turns out to be wrong, you know exactly which downstream conclusions need to be re-evaluated.

In team settings, decomposition serves as a shared evaluation framework. Instead of debating whether an AI's recommendation "feels right" — a conversation that tends to be dominated by the most senior or most confident person in the room — the team evaluates each sub-claim independently, reaching consensus on which components are solid and which need additional validation. This structured approach is faster, more inclusive, and more productive than holistic judgment calls.

When you present a recommendation backed by AI analysis, stakeholders often have a binary reaction: they either trust the AI output entirely or dismiss it entirely. Decomposition provides a middle ground by making the recommendation's structure visible, supporting AI transparency practices. A stakeholder who sees "this recommendation depends on these five premises, three of which are verified and two of which are assumptions" can engage with the specific assumptions rather than making a blanket trust-or-reject decision.

This transparency builds long-term trust in AI-assisted decision-making. When teams consistently surface the assumptions behind their AI-aided analysis, decision-makers learn to calibrate their confidence appropriately. They stop being either AI-credulous (accepting everything the model says) or AI-sceptical (rejecting everything). Instead, they develop the nuanced stance of evaluating AI outputs premise by premise — which is exactly the skill that produces the best outcomes in AI-augmented workflows.

Business plans and project proposals frequently contain compound claims that bundle optimistic assumptions together: 'We can launch in Q3 and capture the market before competitors respond.' Decomposition separates this into its component parts using task decomposition principles — development timeline feasibility, market readiness, competitive response speed, resource availability — each of which carries independent risk. A project that looks viable as a single narrative may reveal an unacceptable risk profile when its assumptions are examined individually.

The technique is especially powerful when evaluating vendor claims, partnership proposals, or technology adoption decisions. Vendor marketing bundles features, timelines, and capabilities into coherent narratives that sound compelling as wholes. Decomposing the narrative into individual verifiable claims — each feature exists today, the integration timeline is realistic, the support commitment is contractual rather than aspirational — reveals which parts of the vendor pitch are substantiated and which are forward-looking statements that may not materialise.