Domain-Specific AI Needs Domain-Specific Knowledge to Work

Gartner predicts that by 2028, sixty percent of enterprise GenAI models will be domain-specific. Trained on healthcare claims, financial regulations, or legal precedents, not on the entire public internet. The logic sounds right. General-purpose models hallucinate too much on specialized topics. A model that has only seen FDA submission documents should answer FDA questions more accurately than GPT ever could. Smaller, cheaper, less prone to making things up. I think this is true as far as it goes. But there is a problem that the prediction does not address.

Cohen and Levinthal (1990) showed that prior related knowledge is the precondition for absorbing new knowledge. They defined absorptive capacity as the organizational ability to recognize the value of new external information, assimilate it, and apply it to commercial ends. It is path-dependent. An organization that has never invested in building domain knowledge cannot absorb new knowledge in that domain no matter what tool it buys. A hospital that has no data analytics routines, no analysts who know how to interrogate a model, and no decision processes that use quantitative evidence will not suddenly start absorbing medical intelligence just because it deployed a medical DSLM.

I think the Gartner prediction tells us about supply. Domain-specific models will be available. The prediction says nothing about demand-side readiness, which is where prior knowledge decides whether the model produces insight or noise. The DSLM is a technological artifact. The capacity to use it is organizational, and it depends on prior knowledge that the DSLM itself cannot provide.

This is the point I want to hold onto. A domain-specific model does not make domain expertise unnecessary. If anything, it makes domain expertise more essential. A general-purpose model that gives a wrong answer about drug interactions is obviously wrong to anyone who knows drug interactions. But the same model can produce a plausible-sounding wrong answer to someone who does not know the domain, and that person cannot tell the difference. A medical DSLM that hallucinates less than GPT still hallucinates. The expert catches it. The nonexpert does not, and catches nothing.

Roberts et al. (2012) provided the IS-specific operationalization of absorptive capacity through four phases: acquire, assimilate, transform, and exploit. The organization that deploys a DSLM has completed the acquire phase. It has the model. But acquisition without assimilation is the most common failure mode, and I keep seeing organizations confuse the two. Buying the model is not the same as being able to evaluate its output, integrate it into decision routines, and act on what it produces. The DSLM solves a technical problem. It does not build the organizational capacity to use technical outputs well.

This explains something I have noticed about which industries are adopting DSLMs fastest. The regulated ones. Finance, legal, healthcare. They are not adopting faster because they have more technology budget, though some do. They are adopting faster because they already have the prior knowledge structures. A bank that has been training compliance officers for decades, building regulatory interpretation routines, and maintaining audit trails can absorb what a financial DSLM produces. The bank has analysts who can validate the model, procedures that integrate validated outputs into the work flow, and a history of acting on specialized financial knowledge. The hospital system that has been running data-driven quality improvement initiatives for years can absorb a medical DSLM in ways that a hospital system with no analytics culture cannot.

The irony is that DSLMs reduce hallucination most for the organizations that need it least. The people who can already spot the model mistakes benefit from fewer mistakes. But the people who cannot spot them, the ones who stand to benefit the most from a model that fabricates less, never know whether the answer is correct or fabricated. The DSLM improves accuracy in a way that is most visible to those who could have survived without it.

I wrote about this dynamic before in the context of why tools alone fail to produce organizational learning. Absorptive capacity accounts for path dependence, and path dependence means a DSLM cannot bootstrap domain knowledge. The domain knowledge must already exist in the organization's routines, people, and decision processes for the model to be usable. The model is a magnifier, not a creator. It magnifies the value of existing expertise. It does not create it.

The organizations that will benefit most from the shift to DSLMs are the ones that already know what they are doing in their domain. They have the prior knowledge base that Cohen and Levinthal identified as the enabling condition: the analysts, the decision routines, the history of evidence-based action. The organizations that lack those things will buy the same DSLM, get the same technical accuracy improvements, and still produce worse outcomes because they lack the capacity to act on accurate information. The tool improves. The gap persists.

There is a structural explanation for this that goes beyond individual training. Expertise is not just something people carry in their heads. It lives in organizational routines that have been built over years, the way a compliance team knows which regulatory filings require which level of review, the way a medical review board has established protocols for evaluating evidence before changing a treatment protocol. Those routines constitute the prior knowledge that makes absorptive capacity real, and they cannot be installed from a model checkpoint. They have to be built through practice, iteration, and accumulated experience.

I think this matters because the DSLM narrative tends to present the problem as a technical one. The general model does not know enough about our domain, so we will give it a domain-specific training set and now it knows. But the bottleneck is not in the model. The bottleneck is in the organization that the model reports into. A financial DSLM deployed at a bank with high compliance expertise will produce better outcomes than the same model at a bank with identical technology but weaker routines. The variable is not the model. It is absorptive capacity.

This aligns with what the institutional theory literature predicts about regulated industries versus less regulated ones. Organizations that operate under regulatory pressure develop the prior knowledge structures that absorptive capacity requires. They have to. If a compliance officer signs off on a model recommendation without knowing whether it is right, the regulator holds that officer responsible. The accountability forces the expertise into the routine. Unregulated industries lack that forcing function, which means their prior knowledge is thinner, which means their absorptive capacity is lower, which means they get less from the same DSLM.

Domain-specific models are the right direction for enterprise AI. I agree with the logic behind the Gartner prediction. Specialized models will outperform general ones on specialized tasks, and organizations should invest in them. But a DSLM is a tool, not a capability. The difference between an organization that buys a tool and an organization that absorbs its value is the prior knowledge that was there before the tool arrived. That prior knowledge has to be built, not bought, and no amount of domain-specific training data can substitute for organizational routines that know what to do with accurate information once they have it.

The shift to DSLMs will close some of the gap between generic AI and expert knowledge. But it cannot close it entirely, because the gap is not primarily technical. It is organizational. The routines, the judgment, the accumulated practice of knowing what questions to ask and what answers to trust, none of that fits inside a model checkpoint. It lives in the people and the processes that use the model. That is where the real capacity lives, and it is path-dependent, which means the organizations that built it before DSLMs existed are the ones that will extract the most value from them.