If all companies have access to the same models, the competitive advantage can't lie in the model itself. It lies in what only you possess: your data, your domain knowledge, and the quality with which your systems are connected. In other words, it lies in the engineering.
During the peak of enthusiasm, the opposite was assumed: that AI would level the playing field and that the difference would shift to marketing or speed of adoption. Data from the last two years points in the opposite direction. Adoption has leveled out, but results have not.
McKinsey, in its 2025 global survey of 1,993 participants in 105 countries, found that 881% of organizations use AI in at least one function, but only 391% report any impact on EBIT at the company level and barely 61% exceed 51% of attributable EBIT.
That difference is the news. If the 88% has access and the 6% gains a significant advantage, access isn't the variable. The factor McKinsey identifies as most correlated with impact is the profound redesign of workflows, which only the 21% of the adopters had undertaken.
Redesigning workflows, integrating them with existing systems, and sustaining them in production has a name. It's not "AI strategy": it's engineering.
Asset | You can buy it | It is being built |
|---|---|---|
Language model | Yes, to several providers. | — |
Development tools | Yeah | — |
Structured and reliable proprietary data | No | Years of operational discipline |
Defined and measured processes | No | Business and Engineering Work |
Integration between proprietary systems | No | Architecture and contracts |
Coded domain knowledge | No | Experience turned into rules |
The first three rows have become commoditized at a remarkable rate: Stanford HAI documented that the inference cost of a system equivalent to GPT-3.5 was divided by more than 280 between the end of 2022 and the end of 2024. What gets cheaper ceases to differentiate.
The last four don't become cheaper because they aren't products: they are the accumulated result of internal decisions. And they are precisely what determine whether a generic model generates value within a specific organization.
There is an interesting market confirmation. Reuters analyzed in August 2026 why established European companies —SAP, Capgemini, Sopra Steria, OVHcloud— have become unexpected winners in the AI cycle: because the challenge for companies is no longer choosing the best model, but making it work with the software, data, and processes they already have.
The pessimistic interpretation would be that only the big players can compete. It's the opposite, and for a specific reason: The capabilities that now differentiate them are more accessible to a medium-sized company than to a multinational corporation..
A mid-sized manufacturer knows its process better than any consulting firm, has fewer systems to integrate, can redesign a workflow without eighteen months of committee meetings, and has proprietary data about its niche that no one else possesses. All of that works in its favor.
What's usually lacking isn't size: it's consistent senior technical expertise. That's why the model works well. External or fractional CTO [internal link], A senior engineer would work one or two days a week, supporting committees, supervising the team, and contributing to architectural decisions. This is a way to develop sound judgment without the structure of a large corporation.
Put your own data in order. A single source of truth based on data, shared definitions, and a consistent history. It's an unassuming task, but it's the foundation of everything else. Applying a model to contradictory data yields no advantage: it produces certain and contradictory answers.
Define and measure the processes that provide margin. Not all of them, just the two or three that explain your profitability. Define them, measure them, and then decide what to automate.
Build the integration layer. Contracts between systems, permissions, traceability. This is what allows you to incorporate any new capability—one model today, another tomorrow—without rebuilding the entire system. It's also what prevents your business logic from becoming trapped within a vendor's configuration.
All three share a common and unusual property: Its value does not depend on which technology wins.. Whatever the dominant models may be in three years' time, a company with clean data, defined processes, and well-integrated systems will be better positioned than one that lacks them.
There's an argument that's rarely framed in technical terms but carries significant weight with the board. In a technology due diligence process, a buyer or investor assesses precisely these things: whether the code and intellectual property belong to the company, whether architectural documentation exists, whether the system can be maintained by another team, and whether there's a critical vendor dependency.
A company whose system is only understood by its supplier receives an implicit discount on every transaction. A company with documented architecture and clear ownership has an asset. It's the same investment, valued in two different ways depending on how it was made.
The conversation about AI will become normalized, just as the conversation about the cloud or having a website did. When that happens, the question won't be which model each company uses, because they'll probably all be using the same ones.
The question will always be the same: who has built a system that understands their business, that can change when the business changes, and that they own?.
That's not a fad. It's a profession.
Because models and tools have become commoditized and no longer differentiate: 881% of organizations use AI according to McKinsey (2025), but only 61% achieve a significant impact on EBIT. The difference lies in proprietary data, defined processes, and the quality of system integration.
Four: structured and reliable proprietary data, defined and measured processes, integration between proprietary systems, and domain knowledge codified in rules. These are the cumulative result of internal decisions, not products available on the market.
Yes, and with advantages: they have a better understanding of their process, fewer systems to integrate, can redesign workflows without lengthy committee cycles, and possess niche data that no one else has. What they typically lack is not size, but ongoing senior technical expertise.
This is a senior engineer who acts as chief technology officer one or two days a week: supporting the committee, overseeing the team, and making architectural decisions. This makes sense when the company needs ongoing technical expertise but doesn't warrant a full management structure.
Three: organize your own data with a single source of truth, define and measure the processes that explain the margin, and build the integration layer with contracts, permissions, and traceability. Its value doesn't depend on which technology ultimately prevails.
A technology due diligence process assesses whether the code and intellectual property belong to the company, whether architectural documentation exists, whether another team could maintain it, and whether there is a critical vendor dependency. A system that only its vendor understands introduces an implicit discount into the transaction.
Do you have ongoing senior technical expertise? Our external CTO service puts a senior engineer on your committee one or two days a week: architecture, team supervision, and decisions that cannot be delegated to a vendor. Let's talk → |
Español Español 
English
Français
العربية
Português
Deutsch
Italiano
