Lead at Scale · White Paper · July 2026Why Generic LLMs Suck at Business
A Case Study: One Model, One Question, Two Knowledge Bases
We asked the same AI model to define “strategy” twice — once from a curated expert compendium, once from its general training — then applied both definitions to one company’s US-expansion decision. Both answers read well. Only one survived scrutiny.
The TL;DR (Short Version)
This case study documents a single working session in which the same AI model defined “strategy” twice — once grounded in a curated expert compendium (The New Strategy Bible), once from its general training — and applied both definitions to one company’s decision: a Canadian protein powder maker weighing US expansion.
Side by side, both definitions read as credible, and both produced plausible expansion approaches. The model framed them as co-equal alternatives with trade-offs. One round of blunt human challenge broke the generic answer, by the model’s own admission.
The diagnosis: the generic definition described a plan and mistook it for strategy — the exact confusion the compendium documents as the most common error in business. The paper closes with operating rules for using LLMs on business questions.
A Single Session Produced Both Answers
Same model, same moment, same question — two knowledge bases.
The prompt was deliberately clean: “Give me two definitions of strategy. One based on the context of this project. One based on your general knowledge, ENTIRELY ignoring your context.”
The “project” was a curated compendium: The New Strategy Bible, a reconciliation of field-tested sources — Beinhocker on value, Kim & Mauborgne on differentiation, Collins on the flywheel, Ries on learning, the Drucker–Grove–Doerr OKR lineage. The alternative was whatever the model absorbed from the open internet.
The Definitions Look Equally Credible on the Page
Nothing warns the reader which one will fail.
Both answers, condensed from the session:
“A coherent working theory of how a business builds competitive advantage over time by accumulating know-how and productive assets faster than competitors.”
“It is a process, not a plan — an overarching framework that keeps plans, projects, and tactics pointed in the same direction: a set of hypotheses about where and how you can differentiate, refined through iteration and learning.”
“The set of choices an organization makes about where to compete and how to win — allocating limited resources toward a defined goal in a way that creates a defensible advantage over rivals.”
“It links ends (objectives), ways (approaches), and means (resources), typically involving deliberate trade-offs about what to do and, equally, what not to do.”
Nothing here warns the reader. The generic version echoes respectable frameworks (ends-ways-means, where-to-play/how-to-win). A manager could quote either in a board meeting without embarrassment.
Applied to a Real Decision, They Prescribe Different Companies
A Canadian protein powder maker weighs US expansion.
The test: how should a Canadian protein powder company approach US expansion? Each definition was asked to answer.
Concretely: one path starts with a contract-manufactured test shipment into a single state. The other starts sizing a warehouse. Same company, same market, different first purchase order.
The Model Framed Them as Co-Equal Approaches
A balanced both-sides table — the kind that reads as rigor.
Asked for the practical differences, the model produced a balanced comparison:
| Dimension | Expert-grounded answer | Generic answer |
|---|---|---|
| What you produce first | A testable hypothesis plus open questions | A set of committed decisions |
| Being wrong | Expected; entry steps designed to surface errors cheaply | Costly; the aim is to choose correctly the first time |
| Speed | Slower to commit capital, faster to start learning | Slower to start, faster to scale once chosen |
| Resources | Staged against validated hypotheses | Allocated to the plan from the outset |
| "Done" | Never; a continuing learning loop | When the plan is executed and the position held |
It closed with two conclusions that sounded reasonable: that large irreversible commitments favor the generic approach, and that in practice most companies blend the two. Both were about to fail inspection.
One Round of Challenge Broke the Generic Answer
No red-teaming — a practitioner pushed back twice, in one message.
No elaborate red-teaming was involved. A practitioner read the output and pushed back twice, in one message.
“If entry requires large irreversible commitments up front — a US manufacturing line, exclusive retail contracts, regulatory investment — the general approach forces the hard choices before that capital is locked in.”
“This sounds like nonsense.”
The model agreed the claim “collapses under scrutiny” and reversed it: irreversibility is an argument for staged commitment. Facing a US manufacturing line, the grounded response is to contract-manufacture or import to test demand, and build only once volume is proven.
“In practice most companies blend them: use the project approach to find and validate the source of advantage, then use the general approach to commit resources and scale once the hypothesis holds.”
“This is just an extension of the Project approach.”
The model conceded: once you validate before committing, you are inside the hypothesis-driven framework — the “general” step is its scaling phase. Its summary: “Calling them alternatives, as the earlier answers did, overstated the contrast.”
The model’s closing self-diagnosis, verbatim: “These aren’t two co-equal approaches you pick between or mix. The general definition describes a plan — where to compete, how to win, resourced commitments. The project definition describes strategy as the process that generates and refines that plan.”
The Generic Definition Described a Plan and Called It Strategy
What survived the session.
What survived the session:
- The confusion is inherited: The internet routinely uses “strategy” for plans, budgets, and goals. A model trained on that corpus reproduces the blur fluently — exactly the strategy-vs-plan confusion the compendium documents as the most common error in business.
- Fluency hid the failure: Both answers were well written. The defect surfaced only because a human challenged the output; a reader taking the generic answer at face value would have started allocating capital against an untested bet.
- The stakes are concrete: For the sample company, the two answers diverge at the first purchase order: a contract-manufactured test in one state, or a warehouse lease and a committed marketing budget. One of these is recoverable if the hypothesis is wrong.
- What the genuine contrast is: The real choice is strategy-as-learning-process versus commitment-without-learning. The second is common practice; it just should not be called strategy.
- The operating rule: Ground the model in curated, field-tested expertise, require it to answer from that base, and challenge the output. The model in this session was capable of the right answer the whole time — it produced one. The knowledge base decided which answer it gave.
Takeaways
Operating rules for using LLMs on business questions.
The operating rules this session supports:
- Treat generic answers as drafts, never verdicts: A public LLM’s answer on a business concept is a fluent average of an incoherent corpus. It is a starting point for scrutiny, and unsafe as a basis for allocating capital.
- Ground before you ask: Load curated, field-tested expertise — your own written standards or a compendium you trust — and require the model to answer from it. The grounded answer in this session held; the generic one broke.
- Challenge once, hard: A single blunt push (“This sounds like nonsense”) was enough to expose the defect. Build that round into any AI-assisted decision process.
- Watch for the plan/strategy swap: If the “strategy” you receive is a resourced list of commitments — segment, channel, budget — you have been handed a plan. Ask what theory of advantage generated it.
- The model is the constant; the knowledge base is the variable: The same model produced the answer that survived and the answer that collapsed, minutes apart. What changed was what it was given to stand on.
Ready to agentize your standards?
Start free, or schedule a free consultation with Clayton to see how it works for your team.
or email clayton@sagely.ltd