The largest AI models are built to do almost anything. Most enterprise work does not need a generalist - and on narrow, well-defined tasks, smaller open models now do the job as well, faster and at a fraction of the cost.
The first piece in this series asked where your AI model runs. The second asked what that choice costs. This one asks the question underneath both: how much model does the work actually need?
For most organizations, the honest answer is less than they are using. The default in enterprise AI has been to send every request to the most capable model available - the large, general-purpose systems from a handful of providers, often called frontier models. It was a sensible habit when AI was new and the biggest model was the safest bet. It is no longer the only safe bet.
Most enterprise work is bounded
Look at what AI actually does inside a typical organization and a pattern emerges. Classify an incoming document. Extract the fields from an invoice or a claim. Check a contract clause against policy. Summarize a case file in a fixed format. Route a request to the right team.
These are bounded tasks: the input is known, the shape of the answer is known, and it is clear what a correct answer looks like. They also make up most of the volume. A model that can draft a strategy paper, write code and debate philosophy is capable of all of them - but none of them uses more than a sliver of that range. You are paying for a generalist to do a specialist’s job.
What changed
On exactly this kind of work, smaller open models have caught up. A model a fraction of the size of the frontier systems, connected to your own documents or trained further on a few thousand of your own examples, can now match the large generalists on the specific task - and in a number of published cases has outperformed them, because it has learned your terminology, formats and exceptions rather than approximating them.
The analysts have taken note. Gartner expects that by 2027, organizations will use small, task-specific AI models at least three times as much as general-purpose large models, observing that the accuracy of general models declines on tasks that depend on specific business context. Researchers at NVIDIA have gone further, arguing that small models should be the default inside AI agents, with large models reserved for work that is genuinely hard or open-ended.
The practical gains follow from size. A smaller model answers faster and needs a fraction of the computing power. And because it is open, it can run where your data already is, rather than your data travelling to it.
Where the frontier still earns its place
None of this makes the largest models redundant. For open-ended reasoning, analysis across many unfamiliar sources, or problems nobody has framed before, they remain clearly ahead - on complex, open-ended work such as large software tasks, specialized small models still fall well short of the frontier. The mistake is not using frontier models. It is using them for everything.
What this looks like by industry
In financial services and insurance, document checks during onboarding, transaction categorization, claims intake and the validation of regulatory reports are bounded and high-volume. A specialized model handles them in-house; the frontier model is kept for the analyst’s open question that no template anticipates.
In logistics, healthcare and the public sector, the same split appears. Reading shipping documents and matching them to orders, structuring clinical notes, sorting and routing incoming correspondence - each is repetitive, rule-shaped work where a smaller model does the job reliably, and where keeping the data on your own infrastructure is often a requirement rather than a preference.
Route by task, not by default
The organizations getting this right do not choose one model. They match each task to the smallest model that does it reliably, and escalate to a larger one only when the task calls for it. This is how Sirma.AI Enterprise is built: with access to more than 500 models - including open models brought in through its integration with Hugging Face - each task can be routed to the model that fits it by performance, cost and sensitivity, and run on-premise, in a private cloud, or on a European sovereign cloud. Trying a smaller model on an existing workload becomes a choice within the platform, not a new integration project.
Where this leaves you
The question is not whether frontier models are capable; plainly, they are. It is whether every task you send them needs that capability. For much of enterprise work, a smaller model you control will do the job as well - faster, at lower cost, and without your data leaving your hands. Size is a capability. It is not a requirement.
If you want to know which of your workloads could move to a smaller model, we can review them with you - and show where the frontier is earning its cost, and where it isn’t.
Learn more about our platform: Sirma.AI
