Defensibility in the age of AI: Rethinking competitive advantage

How far the AI frontier moved while our inquiry was underway

Around June 2025, prominent public offerings included OpenAI’s GPT-4.1, Anthropic’s Claude 4 family, and Google’s Gemini 2.5 family.  By 2026, another generation of frontier models had arrived, including GPT-5.6 and Claude Opus 5.5.  The point is not model genealogy. It is that the technical baseline moved materially during the life of the inquiry.

Other developments reinforced the same point from different directions. In September 2026, OpenAI announced an AI-assisted result on the Navier–Stokes Millennium Prize Problem.  In August, it reported a cybersecurity incident arising from July internal evaluations conducted with reduced safeguards.  In June, Anthropic reported suspending access to Fable 5 and Mythos 5 after a US government directive and later restoring access after the controls were lifted. Capability, containment, and access governance were all moving over the life of the inquiry; none of these events, by itself, demonstrates erosion of a particular business advantage.

This is why we sought the judgment of practitioners: people directly involved in building, operating, backing, or evaluating businesses and technologies. Their views can identify perceived pressures, disagreements, and questions worth investigating. They complement — rather than replace — evidence from observed company, customer, and competitor behavior over time.

Our 64 completed submissions were received from 5 June to 23 August 2026. This is the period represented by the responses, rather than the later administrative closure date. Our survey was self-selected and was not designed to be statistically representative. Detailed information on recruitment, identity verification, denominators, anonymity, and evidence handling appears in the methodology section.

Respondents were deliberately asked to answer at a general level and not to disclose confidential, proprietary, commercially or competitively sensitive, or third-party information. Our study therefore sought broad practitioner judgment rather than detailed proprietary company case studies.

Responses are reported anonymously, and only comments for which explicit permission was given are eligible to appear as quotations. The methodology section explains how optional contact information, identifiers, and quotation permissions were handled.

The report moves from current sources of advantage to perceived weakening, estimated copying times, the commercial work beyond a product, and the limits of a speed-only response. It then brings the main tensions together, poses questions for leaders, and closes with questions the research leaves open. The methodology explains how the responses were analyzed and checked.

Where respondents see their strongest competitive advantage today

Before examining pressure on competitive advantage, it helps to understand where respondents currently locate it. Our survey asked for the single strongest source in their own organization and, separately, the strongest source in their market.

For their own organization, the most frequent choices were deep technical or scientific intellectual property (22%) and brand and trust (21%). At market level, distribution and partnerships led at 20%. The complete distributions follow so that the leading choices can be read alongside the alternatives.

These two results answer different questions. The organization question asks respondents to choose the single source that most distinguishes their own organization; the market question asks which source is strongest in the market more generally. The answer lists and valid response counts also differ. The results therefore should not be read as evidence that advantage has shifted from one source to another, or that organizations are “underinvesting” in a category simply because it is selected more often at market level.

Choosing a single strongest source does not mean that the organization lacks the others. The answer choices also do not map exactly across the two questions: the organization question includes community or ecosystem effects and an option for not yet having strong defensibility, while the market question includes “No single source dominates.” A category that was not offered in one question must not be treated as though respondents selected it zero times.

A respondent can identify an advantage as the strongest source in an organization today and still believe that it has become less reliable over the previous two to three years. Relative strength today and perceived change in reliability are different judgments, not two ends of a single scale.

Pure feature differentiation and access to AI were most often selected as sources of advantage that have become less reliable

Pure feature differentiation and access to AI models or tooling were the most frequently selected sources of advantage becoming less reliable. Asked about the previous two to three years in their business or organization, 42% selected pure feature differentiation and 38% selected access to AI models or tooling.

Access to an AI model is not the same as the ability to apply it effectively, and visible product features are not the same as deep technical or scientific intellectual property. These answers describe perceived reliability rather than changes in revenue, profit, or market position. The finding is therefore narrow but useful: respondents expressed concern about two particular sources of advantage, not about technology-based differentiation as a whole.

Question 6 required respondents to select between one and three sources and did not offer a “none” option. It therefore shows which sources were most often identified as becoming less reliable; it does not tell us what proportion of respondents believed defensibility had weakened overall.

A source selected less often is not thereby proven durable: a respondent may not possess it, may not rely on it, or may have used the available selections for other concerns.

The founder and startup-operator contrast remains large after a response-pattern check

Pure feature differentiation was selected by 70% of founders and senior startup operators, compared with 18% of respondents in other roles. This is a large descriptive difference in the response set; the comparison group includes several different roles, not investors alone.

Founders and startup operators also tended to select more options on this question. When both groups were re-expressed using the same mix of one-, two-, and three-option respondents, the gap fell from about 52 to 39 percentage points. Selection behavior therefore affects the size of the contrast but does not erase it. The founder/operator group was also much more concentrated in pre-seed and seed businesses, so role and company stage remain intertwined.

Within this sample, founders and startup operators were substantially more likely to report weakening feature differentiation. Because role and company stage are entangled, this is a pattern to test again rather than an estimate of the wider founder population.

The respondents’ free-text comments, when provided, also push against the idea that all technology-related differentiation is becoming interchangeable. One respondent focused on the difficult work of designing and refining a product so that it is genuinely intuitive and satisfying to use:

This judgment qualifies, rather than contradicts, the concern about weakening feature differentiation. A source of advantage can come under pressure without every instance becoming easy to reproduce or losing all value.

Other free-text comments described AI as something that could strengthen capabilities rather than merely commoditize them. Two respondents put that possibility directly:

These comments separate access to a tool from the organizational ability to combine it with experience, domain expertise, and technical capability. Our survey does not measure the performance effect of those combinations; it does show that respondents did not treat simple access and effective application as the same thing.

Taken together, the structured answers and contrasting comments support a precise conclusion: feature differentiation and mere access to AI are perceived pressure points in this response set, while technology-related differentiation can still arise from how capabilities are designed, combined, and applied.

Download the full report below.