The precision pivot: The case for local, lean and governed AI

For years, the dominant narrative in artificial intelligence has been one of scale. Bigger models, broader databases, more parameters, under the assumption that being that size alone would unlock transformative capability. 

Governments and regulated industries have watched this race unfold, and many have invested heavily in the promise that general-purpose AI would deliver equally general-purpose results.

Despite record investment, large language models are proving too expensive to scale in resource-constrained environments, too opaque to satisfy audit requirements and too generic to perform reliably in the specific, high-stakes contexts that define public sector work. 

What is emerging in response is not a retreat from AI but rather a shift toward more precise, context-aware systems, designed to be locally accountable, efficiently governed and fit for purpose.

The Case for Fit-for-Purpose Data  

This shift begins by rethinking data as something to be curated with intent instead of consumed in volume. 

Representative over raw: A model’s value is defined by the context it inhabits. For public institutions, that means grounding AI on data that reflects actual operational realities such as local demographics, regional service patterns and jurisdiction-specific regulatory frameworks. Generic models tend to reflect the populations best represented in large-scale internet data, which rarely mirror the actual populations and communities governments serve.  

The Role of Synthetic Data: One of the more significant advances in fit-for-purpose AI is the use of synthetic data, which is algorithmically generated datasets that replicate the statistical properties of real-world information without exposing sensitive records. Using end-to-end platforms, agencies can model citizen services and simulate infrastructure scenarios without compromising personal information. Used wisely, synthetic data can level the playing field for under-resourced agencies. Used poorly, it can amplify the gaps it was meant to close, particularly where underlying datasets are already biased or unrepresentative.

Strategic intent by design: The most effective public sector AI deployments begin with clearly defined use cases and working backward to the data required. By designing pipelines with intent, agencies ensure their AI is built on verifiable, high-quality inputs that can be demonstrated to auditors, oversight bodies and the public.  

Agentic AI as an Efficiency Enabler  

Fit-for-purpose data enables a fundamentally different architecture and movement from massive, centralized processing toward “edge AI,” agentic systems that operate at the point where data is generated.

In practice, this means AI monitoring environmental sensors, processing field-collected health data or supporting local emergency response, without the latency, bandwidth costs or connectivity dependency of a centralized model. 

The gains are already being documented: research from SAS and the Global Center on AI Governance examining AI deployment in the Global South found that agentic systems, when designed for efficiency and deployed at the edge, could become “catalysts for sustainability instead of new sources of the strain.” This finding is directly relevant for any resource-constrained public sector environment, not just emerging economies. 

Domain-specific models also carry a structural efficiency advantage: they do not require the continuous data ingestion that sustains general-purpose systems. For agencies operating under strict budget constraints, leaner models consume less compute, lower operating costs and can be maintained without the resource intensity of re-training a foundational model. 

In an era of fiscal pressure, efficiency becomes a critical enabler. Importantly, these approaches address systemic challenges in AI more broadly, from sustainability to representation, and offer models that may define best practice globally.

The Public Sector Blueprint and Governance as the Backbone  

No sector faces greater scrutiny around AI than government, and no sector is better positioned to model what responsible deployment looks like. The framework underpinning trustworthy public sector AI rests on four pillars: operations, oversight, culture and compliance. These are not independent variables. Each pillar reinforces the other, and their combined presence is what generates genuine institutional trust.  

As AI becomes more distributed, governance can no longer function as a compliance checkbox. It must be the framework upon which the entire AI lifecycle is built. AI governance encompasses the initial ingestion of real or synthetic data, through model training and validation, to the final autonomous action taken by an edge agent. Every decision must be traceable and continuous, not just when something goes wrong.  

Critically, strong governance is an enabler. Organizations that trust their data and processes can move faster, innovate more freely and extend AI into higher-stakes decisions with confidence. In regulated environments, governance is what transforms AI from a technical capability into a genuine institutional asset. 

Precision Will Define the Next Era  

The organizations that will lead the next phase of AI adoption are not necessarily those with the largest models or the deepest training budgets. They are the ones asking the hard question, “How precisely can we deploy AI?”  

Precision means data designed for purpose. Intelligence distributed to where decisions are made, and governance embedded from the beginning, not retrofitted. For government and regulated industry, this is a practical and increasingly necessary path. to AI that is fast enough to be useful, accountable enough to be trusted and efficient enough to be sustained.  

The precision pivot has begun. The question for public sector leaders is not whether to make the shift, but how quickly, and how well.

Josefin Rosén is a Trustworthy AI specialist, working globally in SAS’ AI Ethics, Governance and Social Impact team. She acts as a counselor on AI strategy to business and government and has been named as one of the most influential voices on AI in Sweden by Tech50. Her specialty is ensuring that human factors such as ethics, diversity and transparency are factored into AI solutions. She has more than 20 years of experience working with AI and advanced analytics and holds a doctoral degree from the Faculty of Pharmacy at Uppsala University in Chemometrics.