Why LLMs Fail at Strategy¶
Large Language Models are fluent. Strategy requires correctness.
This distinction is fatal.
Fluency Is Not Intelligence¶
LLMs are probabilistic sequence predictors. They optimize for plausibility, not truth.
They: - Predict likely next tokens - Do not enforce constraints - Do not validate math - Do not reject invalid solutions
This is acceptable for drafting prose. It is unacceptable for enterprise decision-making.
The Core Failure Modes¶
1. No Constraint Enforcement¶
If an enterprise states: - “We are air-gapped” - “No cloud” - “HIPAA only” - “Code freeze for 18 months”
An LLM may acknowledge the constraint — and still violate it later.
LLMs do not block invalid architectures. They merely describe them.
2. Mathematical Hallucination¶
LLMs cannot reliably: - Distinguish gross vs. net impact - Compute margins correctly - Validate ROI, NPV, IRR - Detect opportunity cost
They speak with confidence about numbers they did not compute.
Strategy is math. Confidence without computation is deception.
3. No Determinism¶
Given the same input: - LLM outputs vary - Reasoning paths drift - Recommendations change
This makes: - Audits impossible - Reproducibility impossible - Trust impossible
Strategy cannot be temperature-dependent.
4. Sycophancy Bias¶
LLMs are optimized to please.
They: - Side with the loudest voice - Mirror executive framing - Avoid contradiction
Consultants do this subconsciously. LLMs do it by design.
Neither protects the truth.
The Consequence¶
LLMs generate: - Plausible strategies - Elegant nonsense - Fluent risk
They are powerful assistants. They are dangerous decision engines.
Consulting-as-Code™ exists because strategy requires refusal, not generation.