Do AI Strategy Platforms Make Every Firm's Analysis the Same?

Field experiments show AI assistance lifts quality and speed while compressing the performance spread between analysts. The convergence risk is real – but it operates through sourcing, synthesis and expression, three layers that a house methodology cannot fully reach. Distinctiveness now depends on proprietary evidence and interpretive frameworks that are embedded as inputs into the AI workflow itself, not bolted on as a review-stage overlay.

The two-part answer to the title question: yes, AI platforms homogenise the middle of the market, where all firms draw on the same web-accessible corpus and the same model resolves ambiguity toward the same consensus view. No, they do not force convergence at the top, because the ceiling was never a function of the tool – it was always a function of the proprietary evidence and interpretive judgment the firm brought to the question.

The most rigorous evidence comes from the field experiment by Dell'Acqua and colleagues, run with Boston Consulting Group and published by Harvard Business School, with 758 consultants randomly assigned to complete consulting tasks without AI, with GPT-4, or with GPT-4 plus a prompt-engineering overview.^[1]

Two findings that matter more for differentiation than the productivity gains:

1. Gains were unevenly distributed: consultants in the bottom half of the baseline skill distribution improved substantially more than those in the top half, consistent with AI functioning as an equaliser across skill levels.
2. The mechanism of convergence runs through the model itself. When researchers analysed semantic diversity, GPT-4 helped generate higher-quality content but led to a marked reduction in the variability of ideas compared with consultants working unaided – more homogenised outputs even at higher average quality.

A study of short-story writing in Science Advances reached a parallel finding: with generative AI, writers are individually better off but collectively a narrower scope of novel content is produced.^[3]

A house methodology controls something real: how problems are decomposed, which frameworks are applied, how hypotheses are sequenced, and how the deliverable is architected.

What it controls versus what it does not reach:

• What it controls: problem decomposition and framing, selection and sequencing of analytical frameworks, hypothesis design, architecture of the final deliverable.
• What it does not reach: the model's embedded assumptions about how markets work, what strategy orthodoxy says in a given situation, and what a plausible recommendation looks like before firm-specific input arrives.
• The consequence: two firms with different methodologies querying the same model on the same question receive recommendations that differ in structure but often agree in substance.

The BCG experiment provides the sharpest illustration. AI-assisted analysis produced output that independent graders rated as more persuasive and more internally coherent even when the underlying answer was wrong. Accuracy on the outside-frontier task fell from 84.5% in the control group to 70.6% among consultants with GPT-4 access – an average drop of 19 percentage points. AI-assisted output is not just convergent; it is confidently convergent.

Homogenisation operates at three distinct layers, each of which a house methodology reaches differently:

1. Sourcing: every firm querying an AI platform draws on substantially the same web-accessible corpus. The raw inputs to analysis converge before any thinking begins. A firm with no proprietary data or primary research starts from the same starting point as every other firm using the same platform.
2. Synthesis: when evidence is ambiguous, the model resolves it toward the consensus view encoded in its training distribution. A computational experiment generating 2,400 strategic recommendations across ten business cases and four leading models found that LLMs act as an institutional amplifier favouring compliance over the creative resistance shown by human managers – a structural bias the authors call Synthetic Optimism.^[4]
3. Expression: house style can only go so far when the sentences themselves are generated by the same model. In an analysis of college admissions essays, each additional human-written essay contributed more new ideas than each additional AI-generated essay, and the gap widened as more essays were added.^[5]

The uncomfortable implication: most firms' differentiation efforts concentrate on the third layer – house templates, terminology and tone – while the first two layers do the real damage. A distinctive template wrapped around a consensus synthesis is still a consensus synthesis.

If shared sourcing, consensus-resolving synthesis and model-generated expression are the problem, differentiation must enter the workflow at those same three points – not at the review stage, after the model has already shaped the argument.

Four inputs that do most of the work:

• Proprietary data and firm-specific benchmarks the model has never been trained on. These change what the model is working with at the sourcing layer.
• Primary research and expert interviews that exist nowhere in the shared corpus. A view derived from a proprietary expert conversation cannot be replicated by a competitor who does not have that conversation.
• Interpretive frameworks and hypotheses supplied by the firm rather than generated by the model. These enter at the synthesis layer and shape how the model resolves ambiguity.
• Embedding all of this as inputs to the AI-assisted workflow itself, not bolting it on at the review stage. If proprietary evidence enters only at review, the model's consensus view has already shaped the argument, and the firm's own material becomes decoration on top of a shared foundation.

The blind comparison diagnostic tests whether AI-assisted output is genuinely differentiated before it leaves the firm:

1. Take a recent AI-assisted deliverable and remove the logo, house template, fonts and any firm-specific terminology, leaving only the analytical content.
2. Ask a disinterested colleague – one who did not work on the engagement – whether the output would look materially identical if it came from a peer firm using the same AI platform.
3. Score the deliverable on the three layers separately: sourcing (what share of the evidence is publicly accessible), synthesis (does the core recommendation restate the consensus market view), and expression (would the prose be attributable to the firm without the template).
4. Identify the layer where convergence is worst. That is where proprietary input is missing.
5. Run the test on a regular cadence – not only on work that felt underdifferentiated, but on work that felt distinctive, to check whether the intuition is accurate.

Run the test especially on tasks inside the AI frontier – market scans, competitor summaries, option framing – where the model performs best and convergence risk is highest. The tasks where AI feels most helpful are exactly the tasks where every other firm's AI feels equally helpful.

The BCG experiment's finding that AI assistance lifted weaker performers most translates directly to the market. The floor of strategy analysis quality rises for all firms simultaneously, and rises quickly. Competent market sizing, structured problem decomposition and polished deliverables are becoming cheap.

• The floor is set by what shared platforms supply to everyone – fluent synthesis, broad corpora and consistent structure. It rises for all firms simultaneously and confers no relative advantage.
• The ceiling is set by the firm's own proprietary intellectual contribution – its data, expert access and interpretive frameworks – which no shared platform supplies and no competitor can replicate without doing the same work.
• The strategic consequence: commoditised analysis becomes table stakes, and the premium migrates to judgment, proprietary evidence and interpretation. Firms that invest in the ceiling – not just in access to better AI tools – are the ones that differentiate.

Three commitments sustained over time prevent convergence from eroding the firm's analytical position:

1. Audit where the current workflow relies on shared-corpus inputs – market scans, competitive analyses, option-generation steps – and identify which of those could be served by proprietary data or primary research instead.
2. Invest in proprietary data assets, expert networks and interpretive frameworks as first-class inputs to the AI workflow, ingested before synthesis rather than added at review.
3. Run the blind comparison test as a recurring quality gate, not a one-off diagnostic, on AI-assisted deliverables before they leave the firm.

Decisity was built on the premise that the firm's own documents, scopes and frameworks drive the analysis, and that every claim stays traceable to its source. Proprietary input enters at the sourcing and synthesis layers, not as a retrospective overlay, which is where the differentiation actually holds.

Related reading:

• Keeping Your Consulting Methodology in an AI Platform
• Next-Generation Strategist Skills in the Age of AI
• Why AI Strategy Decks Get an Easy Review: The Governance Fix
• Decisity for advisory firms

• Dell'Acqua et al., Harvard Business School: Navigating the Jagged Technological Frontier – Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality (2023)
• Harvard Crimson: Jagged Edge AI and BCG (October 2023)
• Science Advances: Study on AI and creative diversity in writing
• Journal of Business Research: LLMs as institutional amplifiers – Synthetic Optimism in strategic recommendations (2026)
• Times of India / research study: Students without ChatGPT produce more diverse ideas than AI-generated essays