Build the framework before choosing the model: A durable AI decision workflow for strategy consultants

Build the framework before choosing the model because your client’s business problem cannot wait for a model roadmap to cooperate. A release may move. Access may change. A model that performs well during discovery may behave differently after an update, under a new prompt structure, or when the evidence set expands.

The strategic question remains.

For strategy consultants, this is more than a technical inconvenience. Model-first work can make the engagement dependent on a temporary tool configuration. When that configuration changes, teams often repeat the research, reconstruct the prompt chain, and debate whether a different answer reflects better reasoning or merely different behavior.

A framework-first workflow changes the order. It defines the question, evidence, criteria, assumptions, trade-offs, dependencies, and decision owner before assigning any model a role. The models become replaceable contributors inside a visible method. The method remains the professional asset.

Research on language-model evaluation reinforces this distinction. Repeated-run studies have found meaningful output variation even when prompts and deterministic settings remain fixed. Evaluation research also warns that results can shift with the test setup, comparison method, benchmark quality, and domain criteria. The sensible response is not to avoid models. It is to make the decision system stronger than any single model run.

The fragility of model-first consulting workflows

A model-first workflow usually begins with a familiar question: Which model should we use?

That sounds practical. Often, it is premature.

The question quietly assumes that the assignment is already structured. It usually is not. The client may have supplied a briefing document, several spreadsheets, workshop notes, a partially agreed objective, and conflicting stakeholder expectations. Selecting a model at this stage optimizes the reasoning engine before defining the reasoning task.

The result is predictable. One model is asked to summarize. Another is asked to recommend. A third is used because it is newer or appears stronger on a general benchmark. The outputs differ, but the team has no stable criteria for deciding whether the disagreement is useful, irrelevant, or caused by inconsistent instructions.

Model-first workflows tend to fail in five ways:

1. The question drifts. Each prompt reframes the assignment slightly, so the models are no longer solving the same problem.
2. The evidence changes between runs. One output uses a document, another uses notes, and a third uses current web context.
3. The criteria remain implicit. The team compares writing quality instead of decision quality.
4. Disagreement becomes noise. Contradictions are noticed, but not classified by assumption, evidence, or trade-off.
5. The work is hard to preserve. When the selected model changes, the prompt history becomes the system of record.

That last problem is the expensive one. A prompt chain can produce a useful answer, but it does not automatically preserve the logic that made the answer useful. Strategy consultants need a visible reasoning trail that survives handover, challenge, revision, and later model upgrades.

The durable asset is the decision framework

A decision framework is not simply a template. It is the operating logic for the engagement.

At minimum, it should make nine elements visible:

• The precise decision to be made
• The decision boundary and time horizon
• The available options
• The evaluation criteria
• The evidence supporting each claim
• The assumptions that cannot yet be verified
• The trade-offs, dependencies, and risks
• The confidence level and unresolved questions
• The human owner and review trigger

This structure matters because model quality is task-dependent. A model that is strong at extracting evidence may be less useful at challenging assumptions. Another may generate broader alternatives but compress important distinctions. A third may produce the clearest synthesis. Without a framework, these differences feel inconsistent. Within a framework, they become roles.

For 250 years, consequential ideas have depended on people who could structure complexity, challenge assumptions and make the path forward visible.

That principle is still practical. A strategy consultant adds value by defining what must be compared, what counts as evidence, where judgment enters, and how disagreement changes the recommendation. Models can widen the analysis. They should not own the method.

Start with the strategic question, not the model menu

A useful strategic question is specific enough to organize evidence but open enough to permit real alternatives.

Weak question:

What should the client do?

Stronger question:

Which of three expansion paths should the client prioritize over the next two planning cycles, given customer fit, operational readiness, implementation dependencies, strategic differentiation, and execution risk?

The stronger version already contains the skeleton of the framework. It identifies the decision, the alternatives, the time boundary, and the criteria. It also prevents a model from quietly redefining success around the dimension it handles most fluently.

Before running any analysis, add three controls:

Define what is out of scope

Scope boundaries reduce attractive but irrelevant answers. For example, the engagement may exclude a full operating-model redesign or a new product category. Stating that boundary protects the recommendation from expanding into a different project.

Separate evidence from assumptions

A client statement, an uploaded report, a dataset, and a current web source do not carry identical weight. Label them. This makes it easier to see when a model has converted an assumption into a confident conclusion.

Define what would change the decision

A reusable framework records review triggers. New evidence, a dependency failure, a changed client constraint, or a material disagreement between model perspectives may justify revisiting the recommendation. A routine model release should not force a full restart.

Gather the evidence before asking for synthesis

The evidence pack should be assembled against the decision criteria, not collected because it happens to be available.

For a strategy consulting engagement, a practical evidence map can contain:

Evidence category	Purpose in the framework	Typical status
Client documents	Establish goals, constraints, and prior decisions	Source evidence
Structured data	Test scale, patterns, or operational readiness	Source evidence
Workshop notes	Capture stakeholder interpretation and disagreement	Contextual evidence
Sticky-note clusters	Group themes, concerns, and hypotheses	Working evidence
Current web research	Check time-sensitive external conditions	External evidence
Consultant judgment	Interpret trade-offs and implications	Professional judgment

Jeda.ai can support this sequence inside one visual workspace. Its AI Workspace overview describes visual outputs such as matrices, mind maps, flowcharts, and infographics, while its AI Whiteboard capabilities include document and data analysis, editable visual structures, Web Search, Multi-LLM comparison, Vision Transform, and AI+ extension.

The point is not to throw every file at every model. The point is to make the evidence trail visible. Document Insight can convert relevant documents into structured visual analysis. Data Insight can surface patterns from spreadsheets. Sticky notes can retain workshop observations. Web Search can add current context during an appropriate generation step. The framework then records where each input influenced the analysis.

Choose the framework before assigning model roles

The framework should fit the decision mechanism. A generic grid is better than an unstructured answer, but it is not automatically the right method.

Use a comparison matrix when the client must choose among defined alternatives. Use a risk matrix when uncertainty and consequence drive the decision. Use a scenario matrix when the recommendation depends on external conditions that may develop in different directions. Use a flowchart when the main issue is sequencing, approval logic, or conditional action. Use a mind map when the engagement is still defining the problem space.

In many engagements, the strongest workflow uses more than one view:

1. A mind map to clarify the question and evidence categories
2. A matrix to compare alternatives against common criteria
3. A diagram to show dependencies and relationships
4. A flowchart to communicate implementation or review logic
5. An infographic to summarize the final recommendation for a wider audience

Jeda.ai’s visual canvas allows these outputs to remain connected and editable. Vision Transform can convert an existing visual into another format when the reasoning is sound but the communication format needs to change. AI+ can extend or deepen a selected section while preserving its current context. Any added material still requires consultant review.

How-To 1: Build the framework through the AI Menu

This method suits strategy consultants who want guided structure before selecting model roles.

1. Open the AI Menu. Use the menu at the top-left of the Jeda.ai workspace.
2. Choose the Matrix category. Select an existing framework that matches the decision, or use the AI Recipe Maker when the engagement requires a custom structure.
3. Write the decision in one sentence. Include the options, time boundary, decision criteria, and intended professional outcome.
4. Complete the context fields. Add the client situation, audience, goals, constraints, and important exclusions.
5. Add the relevant evidence source. In the advanced options, use the appropriate document or data analysis route for the file type involved. Keep the source tied to the decision criteria.
6. Use Web Search when recency matters. Add current external context without replacing source review or consultant verification. Jeda.ai’s Web Search and AI+ release notes describe recipe-led workflows that combine structured inputs, current context, and visual output.
7. Choose the reasoning setup after the framework is fixed. Use one model for a focused task or Multi-LLM when the engagement benefits from multiple perspectives.
8. Generate and edit the matrix. Check every cell for evidence, assumptions, confidence, and missing information.
9. Extend only where needed. Use AI+ to extend or deepen existing sections, then review the added material manually.
10. Transform the format when useful. Use Vision Transform if the same reasoning needs to become a flowchart, diagram, mind map, or infographic.

The professional outcome is not merely a populated matrix. It is a reviewable reasoning system. Another consultant can inspect it, challenge it, and understand how the evidence moved toward the recommendation.

Assign roles to models instead of asking which one is best

Once the framework is stable, model selection becomes a narrower and more useful exercise.

Do not ask which model is best in general. Ask which model is fit for a defined role under defined criteria.

A practical three-role pattern is:

The evidence examiner

This role extracts claims from the evidence pack, checks whether each criterion has support, and identifies missing or contradictory inputs. Its output should be traceable to the source structure.

The assumption challenger

This role tests the logic. It looks for hidden premises, weak causal links, neglected dependencies, and conditions under which the option would fail. It should not rewrite the decision question.

The synthesis reviewer

This role compares the structured outputs, identifies genuine agreement and disagreement, and produces a concise synthesis without erasing minority concerns. The consultant then decides what enters the recommendation.

These roles can be reassigned later. If a stronger model becomes available, it can be tested against one role and the same framework. That is much safer than replacing the entire workflow because a new model appears more capable on broad benchmarks.

Compare disagreement by criterion

Multi-model analysis is useful when it reveals different reasoning paths. It is wasteful when it produces three polished answers that cannot be compared.

Keep the first outputs separate. Then classify disagreement:

Disagreement type	What it usually means	Consultant response
Evidence disagreement	Models used or interpreted sources differently	Recheck source mapping
Assumption disagreement	Models inferred different premises	Make assumptions explicit
Criteria disagreement	Models optimized for different definitions of success	Reconfirm the evaluation criteria
Risk disagreement	Models differ on likelihood or consequence	Add confidence and review triggers
Recommendation disagreement	Models weigh trade-offs differently	Preserve rationale before synthesis
Wording disagreement	Meaning is similar but phrasing differs	Do not treat as strategic conflict

This matters because surface variation can look larger than semantic variation. Research on model evaluation repeatedly distinguishes between reproducibility, reliability, and meaningful task performance. Newer evaluation work also favors decomposed, inspectable criteria over a single opaque score.

For client work, the implication is straightforward: compare outputs at the level of evidence, assumptions, criteria, and trade-offs. Do not choose the answer that merely sounds most confident.

How-To 2: Run the framework through the Prompt Bar

This method suits a custom engagement where the consultant already knows the framework structure.

1. Prepare the evidence map. Identify which documents, data files, notes, and current sources support each criterion.
2. Open the Prompt Bar. Select the Matrix command for a decision comparison, or another command that matches the framework.
3. Enter the full framework prompt. Specify the decision, options, criteria, evidence rules, assumptions, risks, dependencies, confidence scale, and required human-owned conclusion.
4. Add current context deliberately. Enable Web Search when the decision depends on changing external information. Keep current findings separate from client-provided evidence.
5. Choose the model configuration. Run a single model for a focused pass or use Multi-LLM to compare perspectives under the same prompt and evidence structure.
6. Keep initial model outputs distinguishable. Review where each perspective agrees, disagrees, omits evidence, or changes an assumption.
7. Use aggregation only after comparison. Let synthesis organize the strongest reasoning, but retain the visible disagreements and unresolved questions.
8. Edit the framework on the canvas. Correct weak claims, add missing evidence, adjust criteria, and record the consultant’s rationale.
9. Use AI+ only for extension or deeper treatment. Review the added material against the same evidence and criteria.
10. Convert and preserve the result. Use Vision Transform for another visual format when needed, then share or export the decision-ready visual work.

This sequence protects professional agency. Jeda.ai helps structure and compare the work, but it does not guarantee the decision. Evidence, judgment, verification, and client accountability remain human responsibilities.

Example prompt: Framework first, models second

The following prompt is designed for a strategy consultant advising a fictional mid-market software client. It uses one shared structure before any model role is assigned.

Create an editable decision matrix for a strategy consulting engagement.

Decision: Which of three expansion paths should a mid-market software client prioritize over the next two planning cycles?

Options:
1. Deepen the current customer segment
2. Enter an adjacent customer segment
3. Add a complementary service line

Evaluation criteria:
- Customer problem fit
- Operational readiness
- Implementation dependencies
- Strategic differentiation
- Adoption friction
- Execution risk
- Time to validated learning

For every option and criterion, include:
- Source evidence
- Assumption
- Trade-off
- Dependency
- Risk
- Confidence level
- Unresolved question

Use the same evidence set and decision criteria for every analytical perspective.

Assign three generic reasoning roles:
- Evidence Examiner: identify support, gaps, and contradictions
- Assumption Challenger: test premises and failure conditions
- Synthesis Reviewer: compare agreement and disagreement without removing minority concerns

Keep the role outputs separate before synthesis. End with a section titled Human Decision Record containing the recommendation, rationale, rejected alternatives, unresolved questions, decision owner, and review trigger. Do not present the AI output as a guaranteed decision.

A strong prompt does not ask a model to be “smart.” It defines what good analysis must contain. That makes the output easier to inspect, compare, and rerun with future models.

Preserve the work for future model upgrades

The final board should function as a decision record, not a frozen screenshot of the preferred answer.

Preserve these elements:

• The original strategic question and scope
• The evidence inventory and source mapping
• The framework structure and criteria definitions
• The model roles and reasoning instructions
• The separate model outputs
• The disagreement classification
• The final human rationale
• The rejected alternatives and reasons
• The confidence level and unresolved questions
• The decision owner, date, version, and review trigger

This record makes future upgrades controlled. A consultant can test a new model in the evidence-examiner role without changing the decision criteria. If the new output identifies additional evidence, the team can assess that addition. If it simply rewrites existing content more elegantly, the decision does not need to move.

That is the real advantage of framework-first work: it separates improvement from disruption.

The same visual record also improves handover. A colleague does not need to reconstruct the argument from chat history. The client can see where evidence ends and judgment begins. Stakeholders can challenge a criterion or assumption without reopening every part of the analysis. And when the engagement continues, the next review starts from an editable system rather than a document archaeology exercise.

What strategy consultants should own

AI can accelerate extraction, comparison, challenge, and synthesis. The consultant still owns the parts that make the work professionally consequential:

• Framing the decision correctly
• Choosing the right framework
• Setting evidence standards
• Defining and weighting criteria
• Distinguishing fact from assumption
• Interpreting disagreement
• Testing implications with stakeholders
• Recording why the recommendation was selected
• Deciding when new evidence justifies revision

This is why Jeda.ai should be treated as a visual intelligence workspace rather than a generic answer interface. It provides a place to assemble evidence, generate structured visuals, compare multiple perspectives, edit the reasoning, collaborate around the same board, and communicate the path from source material to recommendation. The workspace strengthens the method. It does not replace it.

Frequently asked questions

Why build the framework before choosing an AI model?

Because the framework defines the task that the model must perform. It fixes the decision question, evidence, criteria, assumptions, risks, and expected output. Model selection then becomes a test of task fit instead of a broad judgment based on novelty, reputation, or general benchmark performance.

Does framework-first mean model choice is unimportant?

No. Model choice still affects extraction quality, reasoning style, consistency, speed, and synthesis. Framework-first simply delays that choice until the role is clear. This creates a fairer comparison because each model receives the same question, evidence, criteria, and output requirements.

How should a strategy consultant compare multiple model outputs?

Compare them by evidence use, assumptions, criteria interpretation, trade-offs, risks, and unresolved questions. Keep first-pass outputs separate before synthesis. Similar wording does not prove agreement, and different wording does not prove strategic disagreement. The framework should expose the difference.

When is Multi-LLM analysis useful?

It is useful when the decision benefits from contrasting reasoning perspectives or when a single output may hide blind spots. It is less useful when the task is simple, the evidence is weak, or the evaluation criteria are undefined. Multiple answers cannot repair an unclear decision method.

What should remain human-owned?

The decision frame, evidence standards, criteria weighting, stakeholder interpretation, final recommendation, and review trigger should remain human-owned. AI can assist with analysis and synthesis, but it should not convert uncertain evidence into unqualified certainty or replace professional accountability.

How can AI+ be used without weakening the framework?

Use AI+ after a structured visual already exists and only to extend or deepen a selected section. Review the added content against the same evidence and criteria. AI+ should expand the working material, not silently change the decision question, scope, or recommendation standard.

What happens when a better model becomes available?

Test it inside one defined role using the existing framework and evidence pack. Compare the result with the preserved decision record. Update the analysis only when the new model adds material evidence, reveals a valid assumption problem, or changes a decision-relevant trade-off.

Which Jeda.ai visual should come first?

For a defined choice among alternatives, begin with a matrix. For an unclear problem space, begin with a mind map. Use diagrams for dependencies and flowcharts for conditional execution. The first visual should match the reasoning task; later formats can communicate the same logic differently.