In the AI economy, domain expertise is becoming scarcer—and therefore more valuable—than the underlying models themselves. For us in the software industry, this may well be the most important strategic shift of this decade. Those who ignore it will build sophisticated apps on a commodity foundation; those who embrace it will build platforms with lasting competitive advantage.
Three years ago, model size was the main factor. Now we’re seeing the opposite. Generic models are quickly becoming comparable, open-source is catching up to commercial providers, and margins on pure “model access” are under pressure. What remains as a differentiator is how well a platform understands its customers’ business. That’s a challenge that we, as platform builders, can only solve through industry-specific design.
Why Generic AI Stops at the First Specific Question
Any peer provider who has conducted demos in regulated sectors recognizes the pattern. The first demo is spectacular. So is the second one. And then comes the first truly specific question from the client: a legal case citing a specific section of the law, an accounting rule that requires nuance, or an advisory situation that deviates from the standard case.
At that point, it becomes clear what generative AI can and cannot do. It understands language, but not the subject matter. It hallucinates, refers to the wrong context, and provides an average answer when a correct one is needed. That’s useful for writing emails, but useless for real work. Combined with compliance requirements, it’s also irresponsible. As an industry, we need to be honest about that limit.
Domain knowledge consists of three layers, and only the first layer digitizes automatically
A common assumption in our industry: “Train the model on customer data.” As if domain knowledge were just a matter of uploading data. In our experience, that doesn’t work. Domain knowledge consists of three layers, each of which requires its own design choices.
The first layer is explicit knowledge: legal provisions, regulatory frameworks, and product specifications. This information can be digitized and is relatively easy to incorporate into a model.
The second layer is implicit knowledge: how an experienced specialist reads between the lines, which combination of signals sets off alarm bells, and when the letter of a rule differs from its spirit. This knowledge resides in people’s minds and is rarely written down.
The third layer is practical, on-the-job knowledge: how a profession actually works on a day-to-day basis, what exceptions are common, and what customer signals really mean. This knowledge is embedded in workflows and years of experience.
A platform that uses only layer one is just a search engine with a nicer interface. A platform that combines all three becomes a “colleague in software.” You consciously design that difference into the architecture.
Our choice: in-house experts in product development
We build AI agents in collaboration with specialists from the industries we serve, not solely with software developers. This is a deliberate choice with commercial implications: longer development cycles, deeper investment per industry, and greater organizational complexity.
The alternative—generic agents with an industry-specific skin—can be launched more quickly. But in our experience, this runs into the same roadblock that our peers face: the first serious customer request that goes beyond the standard use case. For software leaders, this is a strategic trade-off: do you prioritize time-to-market or time-to-stickiness? We use both, depending on the complexity of the solution.
Three questions we ask ourselves regularly
Are you evaluating your own AI platform strategy? Here are three questions we also use to keep ourselves on our toes.
“Who incorporated the knowledge into our model?” Are they mainly developers or domain experts? The answer determines whether you’re building a product or a tool.
“Can our platform handle exceptions?” In regulated sectors, exceptions are the rule. Simply functioning well in standard cases solves the wrong problem.
“How do we keep our domain knowledge up to date?” Regulations evolve, and market practices change. A platform that doesn’t adapt will be obsolete within a year. That includes ours, if we don’t actively invest in it.
Conclusion: Depth as an Industrial Strategy
In the coming years, the differences between the largest generative models will blur. They will all become good enough for general language tasks. What will remain as a distinguishing factor is domain knowledge and the architecture in which it is embedded.
For fellow providers in the business and financial services sector, this is a strategic choice. Domain expertise is the foundation of our platforms—and the competitive advantage that you can build as a player in this industry, provided you’re willing to invest in what it truly requires.
Frequently Asked Questions
Because models are quickly becoming a commodity. Open-source models are catching up to proprietary models in terms of quality, and margins on pure model access are declining. What remains as a structural differentiator is how well a platform understands its customers’ business. That requires years of investment and is much harder to replicate than a model upgrade.
A sector prompt guides a generic model toward a specific domain, but it remains dependent on what was included in the training data. Industry-specific AI combines explicit industry knowledge (curated), implicit expertise (from domain experts), and practical experience (from product feedback). This results in output that is fundamentally more reliable in terms of nuance, exception recognition, and source attribution.
By systematically involving domain experts in product design, from architectural decisions to feature validation. In our experience, external consultation alone does not work. This requires roles such as domain product leads, cross-functional teams with one industry specialist per squad, and an explicit knowledge retention strategy for when experts leave. It is primarily an organizational choice; the technical stack is secondary.
Yes. But that doesn’t happen simply by “adding an extra sector” as a feature. Sector-specific depth requires dedicated teams, partnerships with industry experts, and a willingness to accept longer development cycles. For players who want to do this, the first strategic question is: Which sector is large and relevant enough to justify that investment?
By establishing a knowledge management process in which legal, tax, or industry updates are systematically translated into platform updates. Combine automated regulatory monitoring with human curation and clear release cycles for knowledge updates. Without this process, even the best platform will quickly become outdated.
