What do you think of when you hear “the Borg”? If you are a Star Trek fan, you know that the Borg are a technologically advanced species that goes around the universe assimilating every technology and species it can find. It absorbs intelligence, erases individuality and “standardizes” everything into its own image.
Sound a bit like how we think about modern LLMs?
Today, we often ask the question, “Which LLM do you use?” There are political and technical fights over which LLMs are allowed and which ones can be trusted. Some of this is because the capabilities of these platforms are dangerous in the wrong hands; but, the far more common concerns are trust and cost considerations. Do we really know what the cloud AI providers are doing with our data? Have you ever been surprised by an unexpected token usage bill?
For government agencies, those concerns affect privacy, compliance, mission continuity and the public’s trust in how AI is used.
Consumer Tools as Enterprise Signals
Apple has taken a lot of flak for being behind in the AI race; but Apple has always been a user experience, privacy-first company. What if they are playing the long game and planning for on-device-first AI capabilities? Newer hardware is certainly powerful enough to run local, purpose-built AI engines. What if they aren’t behind but are on the forefront of where AI is really headed: local AI engines that we can control and can trust?
We are seeing this in other consumer apps and platforms like Google’s AI Edge Eloquent, which can perform on-device voice transcription and cleanup. It also has an option to go to the cloud for more advanced features.
OpenClaw is another example, essentially your own personal agentic AI engine that can run local models and connect to the tools and workflows you already use. Initial security and guardrail concerns aside, these tools point toward a broader trend of AI moving closer to the user, closer to existing workflows, and closer to the data.
Enterprise Platforms Are Moving Toward More Control
Amazon Bedrock is a great example of a broader enterprise version of this trend by giving organizations access to multiple foundation models, including open-source and custom models, within a governed cloud environment.
While Bedrock is not “on-device” AI, it reflects the same shift toward model choice, tighter controls, and fit-for-purpose deployment. Agencies and enterprises increasingly want the ability to decide which model is used, where it runs, what data it can access, and how it scales.
I see this trend accelerating over the next few years. The frontier models will continue to advance and have the most powerful features. But for most use cases, we don’t need access to 100% of the latest and greatest features. You likely only need about 20% to meet specific use cases, and these can be run locally without the need to use metered online platforms.
Practical Government Use Cases
For government, this could apply to several practical use cases.
- Intelligent OCR extraction: Agencies process enormous volumes of documents. A local AI model could help extract key fields, identify missing information, classify document types, and prepare materials for human review without sending sensitive content outside approved boundaries.
- Image interpretation and analysis: From geospatial images to medical imaging, AI can help interpret what is in the image, summarize the contents, and identify specific, relevant details.
- Field and frontline assistance: Many government missions happen in the field, in low-bandwidth areas. A local AI assistant could help draft notes, summarize interactions, retrieve approved policy language, or guide a user through a standard process even when network connectivity is limited or when data sensitivity requires tighter controls.
The Benefits and Limits of Local Control
Open-source models already exist for local usage. Implementing these models in local environments provides greater control over where data lives, how costs are managed, and how models are tested before deployment to production. It can also reduce unnecessary data movement, support offline or edge use cases, and give agencies more flexibility to tune or constrain models for specific missions. Of course, local does not automatically mean safe. These models still need governance, security controls, monitoring, and human oversight.
Start with the Mission, Not the Model
Usage of the major cloud AI providers has its benefits and will continue to be part of an intentional, mission-focused AI strategy. Combining that usage with a healthy mix of localized AI instances will help improve security, cost, and dependability.
As with most technology, there is not a one-size-fits-all, Borg-like solution for every use case. An intentional approach starts with the mission objective, understanding the data and risk, and then choosing the AI approach that fits. Sometimes that will be a frontier cloud model; sometimes it will be a local model; and increasingly, it will be both.