Domain-Specific AI Models Gain Ground as Enterprise Copilot ROI Lags

The shift from bigger models to better-fitting ones

For about three years, enterprise AI strategy mostly meant choosing a frontier model and handing employees a chat window. That approach is now being reconsidered. Analyst forecasts, vendor results and adoption surveys all point to companies moving spending and engineering effort toward models built for a specific industry, function or task. General-purpose large language models are not being thrown out. They are being moved into a narrower role.

The forecasts behind this shift come mostly from Gartner. The firm included domain-specific language models (DSLMs) in its Top Strategic Technology Trends for 2026. Its argument is that generic LLMs often fall short on specialized work, and that models trained or fine-tuned on industry or process data offer better accuracy, lower cost and easier compliance.9 Gartner expects that by 2028 more than half of the generative AI models enterprises use will be domain-specific.9 Separately, it predicts that by 2027 organizations will use small, task-specific models at least three times as much as general-purpose LLMs, measured by usage volume.1

The forecasts differ in their exact numbers. At a Gartner conference in March 2025, analysts said more than 60% of enterprise generative AI models would be domain-specific by 2028.5 The later trends release said "over half."9 The difference is small, but it shows that the timing and size of the shift are still estimates, not settled fact. Every version of the forecast points the same way.

Why generic models are falling short

Analysts mostly give three reasons for the shift: accuracy, cost and context.

Accuracy. Gartner's Sumit Agarwal has said that the range of tasks in business workflows, combined with the need for accuracy, is pushing companies toward models fine-tuned on particular functions or domain data.1 Gartner adds that a general LLM's response accuracy drops on tasks that depend on specific business context.10 Gartner's Roberta Cozza put the buyer's view more bluntly: enterprises have finished piloting and now want to see a return, and a generic system that doesn't understand a company's processes and content isn't really helping.2

Cost. Gartner's Annette Zimmermann has said some domain models the firm studied were about four times more efficient than LLMs on cost and latency.5 Gartner's 2026 trends material reportedly claims development costs up to 50% lower than with general-purpose models.25 Info-Tech Research Group's Thomas Randall makes the practical case: for narrow, repetitive, high-volume work such as customer-service triage, paying for a trillion-parameter generalist doesn't make sense, and running a large frontier model on routine workflows at scale produces cloud bills companies can't sustain.7

Context and data. Gartner says that once companies customize models through retrieval-augmented generation (RAG) or fine-tuning, their own data becomes the main differentiator. That makes data preparation, quality checks and versioning essential.1 Gartner also links DSLMs to AI agents. It argues that agents built on domain models handle industry context better, which matters more as companies move from assistants to systems that act on their own.9

Better small models also make specialization easier. Cozza notes that the major model providers now offer compact models that perform well, and that open-source releases let companies start from a base model, shrink it and train it on their own domain data.2 CIO Dive lists the many small-model choices technology leaders now face, including Google's Gemma, Microsoft's Phi and OpenAI's mini models.3

Copilots are widely deployed but deliver uneven value

The copilot market explains why this matters now. Microsoft 365 Copilot is the main way most large companies first used generative AI, and it is now in very wide use. Microsoft says more than 90% of the Fortune 500 use it.12 Several organizations have rolled it out to six-figure employee counts, including Barclays, Infosys, TCS and Wipro.12

Seat counts differ between trackers, and the gaps are large. One roundup puts paid Microsoft 365 Copilot seats at 15 million in Microsoft's fiscal second quarter of 2026 and more than 30 million by fiscal Q4. On a commercial base of roughly 464 million seats, that is about 6.5% penetration.17 Other trackers still report around 3.3% penetration based on the older 15 million figure.20 Some third-party estimates are much higher, such as claims of 160 million enterprise-licensed users.14 Those figures sit badly against Microsoft's own disclosed seat numbers and should be treated with caution.

The usage data is less flattering. A Recon Analytics survey of more than 150,000 US respondents found that only 35.8% of employees with Copilot access actually use it, compared with 83.1% for ChatGPT.16 The same research gives Copilot an accuracy net promoter score of -19.8, and reports that 44.2% of lapsed users cite distrust.17 A Morgan Stanley/RSM survey found that 79% of enterprises had deployed Copilot, yet 74% of companies using AI tools could not show tangible business value.11

The positive numbers mostly come from Microsoft or from studies it paid for. These include a Forrester Total Economic Impact study estimating 116% ROI for a composite organization, and Work Trend Index findings that most users reach first drafts faster.17 Some large customers do report heavy use. Infosys, for example, reported monthly active usage above 91%.12 Taken together, the data suggests a horizontal copilot works well for general office tasks like summarizing meetings and drafting email.16 It is weaker where accuracy and specialist knowledge decide whether the tool is worth paying for.

This is where domain models come in. Low trust in accuracy and unclear ROI are the same problems DSLM supporters say specialization solves. Microsoft also appears to see the gap. Copilot Studio lets companies build custom agents connected to internal data, and one tracker counts more than 120,000 such agents deployed by early 2026.14 In effect, the general-purpose copilot is becoming the delivery channel, while company-specific data and models carry more of the value.

Legal and regulated industries are moving first

The clearest commercial evidence comes from regulated, document-heavy fields. Gartner says healthcare, financial services, manufacturing and automotive are leading domain-specific adoption, because accuracy matters most there. It projects the domain-specific AI market will reach $11.3 billion by 2028.2 Legal AI is often given as the strongest example. Harvey reportedly reached $300 million in annual recurring revenue by May 2026 and counts most AmLaw 100 firms as customers.23 Other documented deployments include IBM's DSLM trained on legal documents, which helped German courts pre-sort filings in large class-action cases.5 AI Magazine reports that JPMorgan built its own financial text analysis model and that GSK developed specialized models for drug discovery.6

Some performance claims in circulation should be handled carefully. Consultancy and vendor material often cites precise hallucination rates, accuracy ranges and cost savings.2126 These figures usually lack disclosed methods and come from parties selling the shift they describe. The pattern across sources is more convincing than any single figure.

Companies are building model portfolios, not replacing LLMs

The most useful finding is that analysts largely agree this is not a simple replacement. Info-Tech's Randall calls it a better division of labor: routers send well-defined queries to small specialized models and send complex ones to large models.7 He notes that general LLMs still lead on open-ended reasoning and breadth of knowledge, and that small models break down on multi-step reasoning across unfamiliar domains or when long context is needed.7 Gartner advises piloting small models where LLMs have been too slow or not accurate enough, and using multi-model setups where a single model falls short.1 Commentary on the 2026 trends uses the same portfolio framing: general models handle drafting and broad questions, and domain models handle compliance, regulatory documents and other work where mistakes are costly.25

The more extreme "death of generic AI" headlines go further than the evidence supports. The general-purpose frontier is still moving. As of mid-2026, closed models such as GPT-5, Claude Opus 4 and Gemini 2.5 Pro still lead on hard reasoning and tool use. Open-weight models are reportedly about a generation behind and much cheaper to run on a company's own infrastructure, which makes them practical starting points for fine-tuning.

What happens next

Over the next two years, the main question for enterprise AI is likely to be fit rather than model size. That changes what buyers need to get right. Specialization depends on clean, well-managed proprietary data. Gartner says companies need to invest in data preparation and train staff across engineering, risk, procurement and business teams.10 More than a third of technology leaders have already delayed AI projects by at least three months because of compute, budget or skills shortages, which suggests many companies will struggle to put this into practice.3

Gartner also expects some companies to sell access to their specialized models, including to customers and competitors.1 If that happens, a company's proprietary data could become a source of revenue as well as an internal advantage.

The overall picture across the coverage is consistent. General-purpose copilots reached most large companies first, and domain models are where companies now expect to see returns. The organizations that benefit most will probably use both: general models for broad, everyday work and specialized models where errors are expensive.