Government Decision-Support AI and Institutional Dependency

Introduction

Government decision-support AI refers to artificial-intelligence systems used by public authorities to analyse information, predict outcomes, rank cases, allocate resources, detect fraud, assess risk, recommend enforcement action, or otherwise assist officials in making governmental decisions. Examples include AI used for welfare-fraud detection, tax administration, immigration screening, policing, public procurement, healthcare allocation, regulatory inspections, and judicial or administrative case management.

The competition-law problem is not simply whether government use of AI is efficient. A deeper concern arises when a government becomes institutionally dependent on a small number of AI vendors, models, cloud providers, datasets, or technical infrastructures. Once the administration's ability to perform a statutory function depends upon a private technological system, switching costs, interoperability barriers, proprietary data, model-specific workflows and accumulated institutional knowledge may make genuine technological competition difficult.

The issue can therefore be understood as:

AI-assisted governmental decision-making + concentrated technological supply + switching barriers = potential institutional dependency and reduced competitive neutrality.

This creates an intersection between competition law, public procurement, administrative law, constitutional principles, digital governance and public-sector technological sovereignty.

1. Meaning of Institutional Dependency

Institutional dependency exists where a public authority becomes so reliant upon a particular private technology provider that replacing the provider becomes technically, economically or administratively difficult.

Dependency can arise at several levels:

1. Model dependency – dependence upon one foundation model or proprietary AI system.
2. Cloud dependency – AI operates on infrastructure supplied by one cloud provider.
3. Data dependency – historical datasets are stored or structured in a proprietary format.
4. Workflow dependency – government processes become designed around one vendor's system.
5. Skills dependency – public officials become trained around one vendor's tools.
6. API dependency – essential government functions rely on proprietary APIs.
7. Audit dependency – only the vendor possesses sufficient information to explain or validate system outputs.
8. Procurement dependency – successive contracts effectively renew the incumbent's technological advantage.
9. Knowledge dependency – government officials gradually lose the internal capacity to reproduce or evaluate the technology independently.

The critical concern is that formal ownership of the governmental function remains public while practical technological capacity migrates toward a private supplier.

2. Why Government Decision-Support AI Is Different

Traditional procurement normally concerns a product or service that can be replaced relatively easily.

AI decision-support systems can be different because they may become embedded in institutional decision-making.

For example:

Government authority → AI vendor → model → data → recommendation → official decision

Over time, this can evolve into:

Government authority → proprietary model + proprietary data infrastructure + proprietary workflow + vendor expertise → governmental decision

The vendor therefore does not merely supply software. It may become part of the institutional architecture through which government exercises public power.

This creates a potential competition problem because the incumbent may acquire advantages from:

• historical government datasets;
• accumulated system-performance information;
• integration with government databases;
• bespoke model training;
• institutional familiarity;
• proprietary interfaces;
• long-term maintenance contracts;
• security accreditation;
• certification;
• government-specific APIs;
• switching costs.

3. Relevant Competition-Law Questions

Several competition questions arise.

A. Is the AI vendor dominant?

The relevant market could potentially involve:

• government decision-support AI;
• AI risk-assessment systems;
• public-sector analytics;
• AI fraud detection;
• government cloud-AI infrastructure;
• AI case-management systems.

Market definition should not automatically be limited to general-purpose AI.

A specialised public-sector AI supplier may possess substantial market power because government agencies require:

• security certification;
• regulatory compliance;
• interoperability;
• data residency;
• auditability;
• reliability;
• procurement approval;
• continuity of service.

4. Foreclosure Through Institutional Lock-In

A dominant AI supplier may theoretically engage in conduct that makes competing suppliers less viable.

Potential mechanisms include:

Proprietary formats

Government data may become stored in a format that competitors cannot easily use.

API restrictions

The incumbent may make it technically difficult to migrate government applications to competing systems.

Long-term exclusivity

Contracts may prevent government agencies from using competing AI systems.

Bundling

AI decision-support may be bundled with cloud infrastructure, cybersecurity, analytics and database services.

Self-preferencing

A technology ecosystem could favour its own AI model over competing models.

Data accumulation

The incumbent may obtain access to valuable operational data that improves its model and makes future competition increasingly difficult.

5. The Feedback-Loop Problem

A particularly important issue is the government AI dependency feedback loop.

It can operate as follows:

Initial contract

↓

Vendor receives government data and deployment experience

↓

System becomes better adapted to governmental workflows

↓

Government employees become trained on the system

↓

Migration costs increase

↓

Competitors face greater entry barriers

↓

Government renews incumbent contract

↓

Vendor receives more data and institutional knowledge

↓

Dependency becomes stronger

This resembles a data-and-learning network effect.

The competitive problem is therefore dynamic rather than merely static.

6. Procurement as a Source of Market Power

Public procurement can unintentionally reinforce concentration.

A government may initially select the technically strongest supplier.

However, subsequent procurements may favour the incumbent because the incumbent already possesses:

• integration knowledge;
• government-specific datasets;
• security approvals;
• historical performance information;
• trained personnel;
• existing infrastructure.

Consequently, an apparently competitive procurement process may produce a self-reinforcing incumbent advantage.

Competition authorities should therefore examine not merely whether the initial tender was competitive, but whether the architecture of successive procurements preserves contestability.

7. Switching Costs

Switching costs may be particularly severe in AI systems.

They include:

Technical switching costs

Migration of APIs, databases and models.

Financial switching costs

Retraining and redevelopment expenses.

Institutional switching costs

Employees must learn another system.

Legal switching costs

New security and compliance certification may be required.

Operational switching costs

Replacing the system could interrupt government services.

Epistemic switching costs

Officials may no longer understand how historical recommendations were generated by the old system.

The last category is especially important.

If a government has relied on one AI system for years, changing the system may create difficulties in comparing new recommendations with historical decisions.

8. Explainability and Vendor Dependency

AI systems may create an unusual form of dependency because the government may not possess the technical information necessary to evaluate the system independently.

A public authority may depend on the vendor for:

• model documentation;
• training-data information;
• error rates;
• model updates;
• calibration;
• security information;
• audit tools;
• explanations of recommendations.

This creates an audit dependency.

Competition law may become relevant if proprietary restrictions prevent alternative suppliers from independently auditing, interoperating with or replacing the incumbent system.

9. Institutional Capacity and Competition

A government that loses internal technical capacity may become structurally dependent upon private suppliers.

This can be represented as:

Externalisation of expertise

→ loss of internal expertise

→ greater reliance on vendor

→ greater information asymmetry

→ higher switching costs

→ reduced procurement contestability

→ greater vendor bargaining power

Therefore, competition policy may need to consider institutional capacity as an element of market contestability.

10. Six Important Case Laws

The following cases do not all concern government decision-support AI directly. They provide the doctrinal principles that can be applied to AI-enabled public-sector dependency.

1. United Brands Co v Commission

Case: United Brands Company and United Brands Continentaal BV v Commission, Case 27/76.

The European Court of Justice developed important principles concerning dominance and the ability of an undertaking to behave independently of competitors, customers and consumers.

Relevance to government AI

A government may be a particularly sophisticated customer, but sophistication does not eliminate dependency.

If a government agency becomes unable to realistically switch suppliers because of:

• proprietary infrastructure;
• data integration;
• technical expertise;
• security requirements;

the supplier may acquire significant bargaining power.

The case therefore provides a conceptual foundation for analysing economic dependence and supplier power.

2. Commercial Solvents v Commission

Case: Industrie des Matières Colorantes (Intercontinental) and Commercial Solvents v Commission, Joined Cases 6/73 and 7/73.

The case is significant for the principle that a dominant undertaking can infringe competition law through conduct affecting downstream markets.

AI relevance

Suppose a company supplies both:

1. government cloud infrastructure; and
2. AI decision-support systems.

It could potentially leverage infrastructure dominance into the AI layer.

Similarly, a dominant AI infrastructure provider could potentially disadvantage competing decision-support applications by restricting access to essential technical resources.

This demonstrates why government AI should not be analysed as an isolated software product where the supplier controls several adjacent layers.

3. Bronner v Mediaprint

Case: Oscar Bronner GmbH & Co. KG v Mediaprint Zeitungs und Zeitschriftenverlag GmbH & Co. KG, Case C-7/97.

The Court established demanding criteria for refusal-to-deal claims involving an alleged essential facility.

AI relevance

The case becomes relevant where a government agency claims that access to a particular:

• AI infrastructure;
• cloud platform;
• dataset;
• API;
• technical interface;

is indispensable for competing suppliers.

The case cautions that not every commercially valuable infrastructure is legally an essential facility.

For an AI infrastructure to justify intervention under an essential-facilities theory, the legal requirements must be carefully established.

4. Microsoft Corp v Commission

Case: Microsoft Corp v Commission, Case T-201/04.

This is one of the most important precedents for technology ecosystems, interoperability and leveraging of dominance.

The case concerned Microsoft's refusal to provide interoperability information and its relationship with adjacent software markets.

AI relevance

The analogy is powerful for government AI.

A dominant provider could potentially create competitive barriers through:

• proprietary APIs;
• restricted interoperability;
• closed data formats;
• technical information withholding;
• integration advantages.

If government agencies cannot move their AI workflows or data between providers, interoperability restrictions may transform technological integration into durable market power.

The case therefore provides an important framework for assessing AI lock-in and interoperability.

5. Google Shopping

Case: Google and Alphabet v Commission, Case C-48/22 P.

The case concerns Google's conduct in favouring its own comparison-shopping service within its broader search ecosystem.

AI relevance

The broader principle is relevant to government AI ecosystems where a provider controls multiple layers.

For example:

cloud infrastructure → AI model → government application → recommendation interface

If the same supplier controls all four layers, it may have incentives and opportunities to favour its own downstream services.

The competition concern becomes stronger where government procurement effectively gives the incumbent privileged access to an institutional ecosystem that competing AI suppliers cannot replicate.

6. Slovak Telekom

Case: Slovak Telekom a.s. v Commission, Case C-165/19 P.

The case concerned abusive conduct involving access to infrastructure and foreclosure of competitors.

AI relevance

The case is useful for analysing situations in which a dominant technology provider controls an infrastructure layer upon which competing AI services depend.

A government AI ecosystem may contain:

• cloud infrastructure;
• computing capacity;
• identity systems;
• databases;
• APIs;
• model hosting.

If competitors cannot realistically access a critical infrastructure layer, competition in downstream AI services can be impaired.

The case consequently supports analysis of infrastructure-based foreclosure.

11. Additional Important Authorities

Several other cases can strengthen the analysis.

IMS Health v Commission

The IMS Health litigation is highly relevant to interoperability, intellectual property and access to technologically important structures.

It demonstrates the difficult balance between:

• protecting innovation and intellectual property; and
• preventing proprietary structures from eliminating downstream competition.

This is directly relevant to proprietary AI architectures and government datasets.

Bronner

Provides the stringent essential-facilities framework.

Google Android

Google Android, Case AT.40099, illustrates how tying and ecosystem strategies can reinforce the market position of a dominant digital platform.

Intel

Intel v Commission, Case C-413/14 P, is important for analysing exclusionary conduct and the economic effects of conditional rebates.

Although not an AI case, it illustrates why the competitive assessment should examine actual or potential foreclosure effects rather than relying solely on formal contractual labels.

12. Public Procurement Dimension

Competition concerns may arise even without an Article 102-type abuse.

Government procurement rules should consider:

Multi-vendor architecture

Government systems should ideally allow multiple AI suppliers to compete.

Interoperability

Contracts should require meaningful interoperability and data portability.

Exit clauses

Government agencies should have realistic mechanisms for changing providers.

Data ownership

Government-generated data should not become unnecessarily locked into the vendor's ecosystem.

Documentation

The government should retain sufficient documentation to operate and audit the system independently.

Open standards

Where technically appropriate, procurement should favour interoperable standards rather than proprietary architectures.

13. Competition-Neutral AI Procurement

A competition-sensitive procurement framework could contain:

1. Open technical standards
2. API portability
3. Data portability
4. Model portability where technically feasible
5. Interoperability obligations
6. Transparent switching costs
7. Exit assistance
8. Restrictions on unnecessary exclusivity
9. Independent audit rights
10. Multi-provider compatibility
11. Periodic re-tendering
12. Vendor-neutral data architecture

The objective is not necessarily to prohibit long-term contracts.

The objective is to ensure that a long-term contract does not become permanent technological dependence.

14. Competition Law and Administrative Law Converge

Government AI presents a particularly important intersection between competition and administrative law.

Traditional administrative law asks:

Was the governmental decision lawful, rational and procedurally fair?

Competition law asks:

Does the structure of the technological market permit effective competition?

With AI, these questions increasingly overlap.

If only one private company possesses the infrastructure necessary to generate governmental recommendations, then the supplier's market power may indirectly affect:

• administrative discretion;
• transparency;
• accountability;
• procedural fairness;
• institutional independence.

15. Risk of Algorithmic Institutional Capture

A more advanced concern is algorithmic institutional capture.

This occurs when a private technology provider becomes sufficiently embedded in governmental decision-making that its technological assumptions influence the functioning of the public institution.

For example:

Vendor model

→ determines what data are collected

→ determines which variables matter

→ determines risk classifications

→ determines recommendations

→ officials increasingly rely on those recommendations

→ governmental practices adapt to the model

The government remains formally sovereign, but its decision-making capacity becomes partially shaped by the private technological infrastructure.

This is broader than conventional market foreclosure.

It is a form of institutional dependency produced through technological architecture.

16. Network Effects and Government Data

Government deployments can generate particularly valuable feedback data.

For example, an AI fraud-detection system may learn from:

• previous investigations;
• confirmed fraud cases;
• administrative outcomes;
• error corrections;
• enforcement results.

If the incumbent controls the feedback loop, competitors may not have access to equivalent datasets.

This can produce:

more government contracts → more data → better model → more government contracts.

Such a cycle can create a public-sector data network effect.

Competition authorities should therefore investigate whether government procurement inadvertently gives one supplier exclusive access to data that is necessary to compete in future procurements.

17. The Problem of De Facto Exclusivity

A contract may not expressly state that the government must use only one provider.

Nevertheless, practical dependency can create de facto exclusivity.

For example:

"The government may technically switch providers, but migration would require two years, extensive redevelopment and loss of historical model compatibility."

In economic terms, the government may therefore be functionally locked in.

Competition analysis should consequently examine the practical ability to switch, rather than contractual wording alone.

18. Remedies

Potential remedies include:

Structural remedies

• separation of cloud and AI businesses;
• divestiture in extreme circumstances;
• separation of infrastructure and downstream applications.

Behavioural remedies

• interoperability;
• API access;
• data portability;
• non-discrimination;
• transparent pricing;
• audit access.

Procurement remedies

• mandatory multi-vendor procurement;
• shorter renewal cycles;
• technology-neutral specifications;
• competitive re-tendering.

Institutional remedies

• government-owned data repositories;
• internal AI expertise;
• independent technical audit units;
• model validation capabilities.

Contractual remedies

• exit assistance;
• data export;
• migration support;
• source-code escrow where justified;
• documentation obligations.

19. A Competition-Law Test for Government AI Dependency

A useful analytical framework is:

Step 1 — Identify the relevant market

Is the relevant market:

• general AI;
• public-sector AI;
• government decision-support AI;
• AI fraud detection;
• government cloud AI;
• specialised administrative AI?

Step 2 — Identify dependency

Determine whether government agencies can realistically switch suppliers.

Step 3 — Identify the bottleneck

Is the bottleneck:

• compute;
• cloud;
• model;
• data;
• API;
• security certification;
• interoperability;
• technical expertise?

Step 4 — Assess market power

Consider:

• market share;
• entry barriers;
• switching costs;
• network effects;
• procurement history;
• data advantages.

Step 5 — Examine exclusionary conduct

Consider:

• tying;
• bundling;
• refusal to interoperate;
• discriminatory access;
• exclusivity;
• self-preferencing;
• contractual restrictions.

Step 6 — Examine institutional effects

Ask whether dependence reduces:

• contestability;
• procurement competition;
• administrative independence;
• technological sovereignty.

Step 7 — Design proportionate remedies

Prefer measures that preserve innovation while restoring contestable government technology markets.

20. Key Case-Law Principles — Summary

Conclusion

Government decision-support AI creates a new form of competition concern: institutional technological dependency.

The central issue is not simply whether an AI supplier has a large market share. A supplier may become powerful because government institutions progressively build their operations around its:

• models,
• cloud infrastructure,
• APIs,
• datasets,
• workflows,
• technical standards,
• expertise and
• audit systems.

Once this happens, the government may become a captive technological customer, even though it formally retains procurement freedom.

Competition law therefore needs to look beyond conventional price effects and examine contestability, interoperability, switching costs, data advantages, infrastructure control and institutional capacity.