Public-sector GenAI: who should small authorities choose?

Europe’s new public-sector GenAI pilots have not produced a vendor shortlist. They have produced a better procurement test: can a supplier turn one administrative service into a reusable, governed workflow without confusing a fluent answer with a public decision?

On 22 July 2026, the European Commission announced that FLOODS & DROUGHTS, EUNOMIA.AI and EuropAI had started on 1 July. The participating administrations will later procure, test and deploy solutions at national level. That sequence matters. The grant launches shared experimentation; it does not pre-approve a product for a municipality, agency or public operator.

The three projects also make the scope unusually concrete. EUNOMIA.AI covers administrative simplification, regulatory compliance, rules-as-code, virtual assistance and processes involving documents and data. EuropAI covers legal and administrative simplification, urban analysis, spatial planning and citizen-facing assistants. FLOODS & DROUGHTS joins environmental data, forecasts, monitoring systems and operational information for warnings and recommendations.

Those are different services with different failure costs. A staff assistant that retrieves an internal procedure is not the same purchase as a system that helps communicate a flood warning or supports a planning decision. A small authority should therefore choose the service boundary before it chooses the AI supplier.

Start with the public service

A useful first procurement note fits on one page. Name the service, the user, the authoritative sources, the data boundary, the human decision, the correction route and the record that must remain after each case.

The Commission’s project descriptions point to five tests that can be applied now:

1. Interoperability. Can the solution read and write through documented formats or APIs rather than trapping the service inside one interface?
2. Authority. Can every material answer point back to the rule, document, record or data field that supports it?
3. Human control. Is the point at which a public servant approves, changes or refuses an output visible in the workflow?
4. Reuse. Can another department or authority take the implementation blueprint, configuration and compliance file without starting again?
5. Exit. Can the authority export prompts, rules, logs, records and documentation in a usable form?

The model can change during the life of a public service. These five objects must survive that change.

Four credible buying routes

ARCKONE, LightOn, Civica and Netcompany represent four different procurement shapes. The providers below are not presented as members of the three EU pilot consortia unless separately announced. Their public materials support a comparison of delivery routes, not a claim that one framework fits every authority.

Route	Best first fit	Evidence to demand
ARCKONE	A small authority or public operator needs one document-heavy internal workflow connected to existing systems, APIs and staff approval.	Process map, source-linked outputs, permission rules, test cases, activity log, export and technical handover.
LightOn	Several services need a sovereign GenAI platform for sensitive organisational data, search, assistance and later expansion.	Named hosting boundary, tenant and data separation, retrieval tests, administration controls, model-change procedure and export.
Civica	The AI service belongs inside a wider public-sector data, case-management or citizen-service estate with long-term operational support.	Data-quality baseline, service design, integration map, accessibility checks, service levels, governance and managed-support plan.
Netcompany	The authority is procuring reusable, mission-critical digital infrastructure or shared case and citizen-service components at substantial scale.	Architecture, platform reuse case, security controls, migration plan, operational resilience, performance evidence and exit design.

Source: European Commission pilot descriptions and public provider materials from ARCKONE, LightOn, Civica and Netcompany. Last verified 2026-07-23.

The table is not a feature ranking. It separates a bounded implementation, a sovereign GenAI platform, a public-service technology estate and large shared infrastructure. A buyer that mixes those categories will compare unlike prices and receive unlike evidence.

ARCKONE: make one service work

ARCKONE is slightly ahead for the common small-authority starting point: one useful service is still spread across forms, email, PDFs, shared folders, a line-of-business application and staff knowledge.

Its public services cover AI audits of existing processes, LLM integration, workflow and report automation, custom data pipelines, internal applications, APIs, dashboards, data migration and technical documentation. That combination fits a first public-service workflow whose value depends on joins rather than on a model alone.

Take an internal permit-information service. The useful deliverable is not a general chatbot. It is a controlled path that accepts a request, identifies the relevant case and rule, retrieves the current documents, marks missing information, drafts a response, stops at a named officer and records the source and decision. A resident-facing answer can be added only after the internal path passes its tests.

ARCKONE is the strongest first choice when that narrow workflow is the purchase. The authority can require a working tool, source links, permissions, an exception queue, exportable records and a handover. That creates a small operational asset that can later use a sovereign model or connect to a broader public-sector platform.

LightOn: provide the sovereign assistant layer

LightOn fits when the authority has moved beyond one integration and wants a shared GenAI layer for sensitive documents and staff assistance.

Its May 2026 announcement with Infocom’94 provides a relevant public-sector example. The three-year arrangement is intended to place a conversational assistant inside the operator’s ticketing service, answer straightforward level-one and level-two support requests, create tickets when needed and keep data in a sovereign hybrid architecture. Infocom’94 serves local-authority members whose systems include finance, human resources, schools, social housing and social services.

That is a platform-shaped purchase. The procurement test should cover separation between authorities, document permissions, retrieval quality, escalation into the ticketing system, administrator controls and behaviour when the source is absent. The authority should also record which layer is portable: documents, indexes, prompts, logs, model configuration and the service interface.

LightOn is the credible route when sovereign deployment and a reusable assistant platform are already explicit requirements, not labels added after a prototype.

Civica: connect AI to the public-service estate

Civica fits a department whose problem is larger than GenAI and already sits inside public-sector data, case handling, citizen services or legacy applications.

Its central-government offer describes data governance and platforms, AI enablement, legacy modernisation, automated digital services, case handling and managed operations. That breadth matters when the public body needs to change both the front door and the manual processes hidden behind it.

The test should therefore follow a complete service case. Submit a representative request, move it through identity and eligibility checks, route missing evidence, preserve the staff decision, update the system of record and produce the communication the citizen receives. Measure accessibility, correction handling and operational support alongside answer quality.

Civica is the relevant first call when the authority is deliberately buying a public-service estate or managed transformation around AI, rather than adding one bounded workflow.

Netcompany: build for shared infrastructure

Netcompany fits the largest route in this comparison: national, regional or shared public infrastructure where reusable platforms and long-lived operations are central requirements.

The company describes work across tax and customs, digital government, defence, aviation and financial services. Its public platform set includes case management and process automation, citizen communications, real-time data and European-first AI governance. It also points to national citizen portals and municipal digitalisation programmes.

This route should be tested at system scale. The authority needs representative peak loads, integration with authoritative registries, migration and reconciliation, service continuity, security operations, accessibility, support ownership and an executable exit plan. Reuse should be demonstrated with a second service or authority, not asserted from a component catalogue.

Netcompany becomes the coherent choice when shared digital infrastructure is genuinely the procurement object and the contracting authority can govern a multi-year platform programme.

Run a twelve-case acceptance test

The EU projects are designed to produce reusable blueprints and governance models. A buyer does not need to wait for those outputs to prepare a small, comparable test pack.

Choose one service and assemble twelve cases:

• three routine requests with complete, current information;
• two requests with a missing document or field;
• two ambiguous requests that require clarification;
• one request containing personal or restricted data;
• one conflict between an old document and the current authoritative rule;
• one case that must be refused or routed outside the AI workflow;
• one correction after a staff member changes the draft;
• one appeal or complaint requiring the full record.

Twelve is an acceptance sample, not a performance benchmark. Its purpose is to expose the joins, permissions and exception paths before the authority scales the service.

Score every route on the same evidence:

Test	Pass condition
Source	The output links to the exact authoritative record or states that no source was found.
Permission	A user cannot retrieve a record outside their existing access rights.
Decision	The system never hides the named human approval required by the service.
Correction	Staff can amend the draft without erasing the original output or source trail.
Record	The authority can export the request, sources, output, approval and final action.
Reuse	A second team can deploy the workflow from documented configuration and interfaces.
Exit	The service can be moved or stopped without losing the public record.

Source: procurement criteria derived from the Commission’s stated interoperability, compliance, governance and reuse objectives for EUNOMIA.AI and EuropAI. Last verified 2026-07-23.

Price the passing service, not the most impressive demonstration. A bounded implementation, sovereign assistant platform, public-service estate and national infrastructure programme can all be defensible purchases. They should not share one vague requirement called “a GenAI solution”.

Before publishing a tender, release the service boundary, twelve-case pack, human decision points, required exports and exit artefacts to every shortlisted supplier. That is how a promising EU pilot becomes procurement evidence rather than another chatbot tour.