AI infrastructure is often described in technical terms: models, data platforms, agents, workflows, and tools. But public services face another infrastructure problem that may be just as important: meaning.
A public-service AI agent does not only process data. It processes institutional language. Words like eligible, urgent, referred, reviewed, closed, and escalated may sound clear. Across agencies, vendors, forms, and databases, they may not mean the same thing.
The risk is not that AI misunderstands a word. The risk is that an AI agent may take action based on the wrong meaning of that word.
A “referral” might mean a name was sent to another office. In another system, it might mean the person was accepted for service. In another, it might mean the service was completed. If an AI agent treats all three meanings as the same, it may close a gap that is still open.
This connects directly to PSA’s earlier post on “An AI stack when personal Agents are involved”. The point there is that agents are different from ordinary chatbots because they can touch data, call tools, invoke applications, and eventually execute work. That makes the “System of Intelligence” — the governed layer of data, context, rules, actions, and business logic — central to whether agents can act safely.
That is why public services need a meaning layer before agents are allowed to act across systems.
A semantics analyst could help map where important terms appear, who defines them, what action they trigger, and where human review is required. This is practical governance, not academic word work.
NIST’s AI Risk Management Framework is useful here because it emphasizes mapping, measuring, managing, and governing AI risk before systems are trusted in real settings. OMB’s federal AI guidance also stresses agency governance and risk management for AI uses that affect public rights and safety.
If the meaning layer is unclear, AI will scale the confusion.
Before public-service agents act, agencies need to define what their own words mean, who has authority to interpret them, and what must happen when meaning is uncertain.
AI agents should not be allowed to act on public-service terms that the organization has not clearly defined.
Source references
- PSA, An AI stack when personal Agents are involved
- NIST AI Risk Management Framework
- OMB Memorandum M-24-10 on federal AI governance
This post was drafted with GPT assistance and reviewed through a PSA public-service lens focused on governance, accountability, and real-world implementation.