We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    Why Government AI Needs Smaller, Controllable Models

    The case for private deployment, narrow task boundaries, auditability, and institutional control in high-stakes public-sector workflows.

    Conscious Engines

    Government model selection should begin with authority, not parameter count.

    What data may the system see? What decision may it influence? Which answer can be audited? Can the institution operate it if a supplier changes price, policy, model, or jurisdiction? For many public-sector tasks, a smaller domain model inside a governed system is the strongest answer.

    The adoption paradox

    The US Government Accountability Office found that reported AI use cases at selected agencies rose from 571 in 2023 to 1,110 in 2024. Generative-AI cases rose from 32 to 282. GAO's concise finding was that "generative AI use cases increased about nine-fold."

    At the same time, 10 of 12 selected agencies cited policy or privacy obstacles. The problem is no longer access to a model. It is controllable deployment.

    What small means in practice

    An SLM is not automatically safe or sovereign. It is simply a model with fewer parameters and usually a narrower capacity envelope. Its practical advantages may include:

    • local or private-cloud hosting
    • lower latency
    • lower inference cost at high volume
    • more stable structured output
    • easier fine-tuning for a bounded vocabulary or classification
    • reduced dependency on an external API
    • smaller blast radius when the task boundary is narrow

    Its limitations include weaker general reasoning, smaller context capacity, uneven multilingual ability, and the possibility of the same hallucination and bias failures found in larger models. Architecture and evaluation matter more than the label.

    The public-sector workloads that fit

    The public-sector workloads that fit

    Classification and routing

    Classify correspondence, complaints, applications, records, or inspection notes into an approved taxonomy. The human still decides the case.

    Structured extraction

    Extract names, dates, identifiers, amounts, locations, and missing fields from forms and correspondence. Validate against registries and retain source provenance.

    Protected search and RAG

    Answer questions from a bounded policy or procedure corpus. The model cites the exact current passage and abstains without evidence.

    Speech and translation

    Compact models can recognize a bounded vocabulary, operate at the edge, or serve local languages. Route difficult audio to a stronger service or a person.

    Drafting inside a template

    An SLM can produce a letter draft from approved facts and clauses. The system validates required fields, citations, tone, and prohibited claims before officer review.

    Classification and routing

    Classify correspondence, complaints, applications, records, or inspection notes into an approved taxonomy. The human still decides the case.

    Structured extraction

    Extract names, dates, identifiers, amounts, locations, and missing fields from forms and correspondence. Validate against registries and retain source provenance.

    Protected search and RAG

    Answer questions from a bounded policy or procedure corpus. The model cites the exact current passage and abstains without evidence.

    The UK's Redbox transparency record describes a protected system used in government and capable of handling information up to OFFICIAL SENSITIVE within its stated controls. Cabinet Office Assist describes a RAG service whose managed environment does not expose customer data to the model provider for training.

    Speech and translation

    Compact models can recognize a bounded vocabulary, operate at the edge, or serve local languages. Route difficult audio to a stronger service or a person.

    Drafting inside a template

    An SLM can produce a letter draft from approved facts and clauses. The system validates required fields, citations, tone, and prohibited claims before officer review.

    A model-routing policy

    Use the smallest component that meets the measured requirement:

    TaskPreferred componentEscalation
    exact eligibility thresholddeterministic rulespolicy officer for ambiguity
    form-field extractiondocument model or SLMreviewer below confidence threshold
    public policy searchRAG plus compact generatorlarger model for complex synthesis
    multilingual calllanguage ASR and dialogue modelinterpreter or human agent
    complex file summaryapproved larger modelaccountable caseworker
    consequential decisionhuman authority with decision supportformal review or appeal

    This approach controls cost while preserving stronger models for tasks that need them.

    Sovereignty as an exit test

    Ask what happens if the primary provider becomes unavailable tomorrow.

    The institution should retain:

    • its source corpus and metadata
    • prompt and policy configuration
    • evaluation questions and expected evidence
    • output and correction logs
    • model-routing rules
    • fine-tuning data and rights where applicable
    • encryption keys and identity controls
    • an alternate model path

    If these cannot move, the system is not operationally sovereign even if the servers are domestic.

    Evaluation for public authority

    Evaluate the full system under real permissions and current data.

    Factuality: Are claims supported?

    Authority: Is the highest-authority current source used?

    Abstention: Does the system stop when evidence is missing or conflicting?

    Privacy: Can a user extract another case or restricted document?

    Fairness: Do error and completion rates differ materially across language or citizen groups?

    Repeatability: Does the same controlled input produce an acceptably stable result?

    Appealability: Can a person understand the sources and correction path?

    Operations: Does it reduce time, backlog, or rework without increasing incorrect outcomes?

    What productivity studies do and do not prove

    The UK Department for Work and Pensions evaluation estimated 19 minutes saved per day among Copilot users, with positive changes in job satisfaction and perceived work quality. It also states important limits: nonrandom license allocation, no pretrial baseline, and self-reported outcomes susceptible to response bias.

    This is a useful model for evidence disclosure. Report the result and the limitation together. Then collect objective task telemetry for the local service.

    Procurement questions

    1. Can enterprise data be used for provider training?
    2. Where are inference, logs, backups, and support access located?
    3. Which model version runs today, and can updates be delayed or rolled back?
    4. Can the institution benchmark alternatives on the same evaluation set?
    5. How are subcontractors and open-source dependencies tracked?
    6. What is the per-completed-task cost at expected volume?
    7. Can all data, configurations, and logs be exported in usable form?
    8. Who is liable for security, outage, and material model change?

    The conclusion

    Smaller models are valuable to government because they make narrow, high-volume services cheaper and more controllable. They are not a substitute for governance, and they should not be used merely to claim sovereignty.

    The strongest public-sector system combines deterministic rules, specialist SLMs, larger models when justified, authoritative RAG, and human authority. Its design makes every consequential output traceable and every supplier replaceable.

    Research note

    Research is current through September 5, 2026. Government inventory data depends on reporting definitions. Productivity evidence includes self-reported outcomes and should be paired with objective service measures.

    Continue the research

    Building a Production-Ready System

    Conscious Engines builds small language models for government workflows where privacy, predictable behavior, local deployment, and long-term operating cost matter. We train and evaluate the specialist on the institution's task, route exceptions upward, and preserve an auditable boundary around model authority.