Why AI procurement is a governance decision, not just a commercial one

Organisations buying AI tools need to confront and uncomfortable reality – buying AI is not like buying traditional software. While all software procurement decisions rely on internal rules around data sovereignty, liability, regulatory compliance, and long-term operational dependency, the dynamic nature of the AI landscape means that procurement professionals cannot rely on these rules as easily when purchasing AI. This article explains why standard procurement processes are not built for that task, and what organisations need to do differently. 

This article will cover: 

• Why AI procurement carries governance consequences that standard buying processes cannot handle 

• The structural asymmetry between AI vendors and buyers, and what it means in practice 

• Four risk areas organisations are currently underestimating: supply chain accountability, data and IP exposure, supplier lock-in, and agentic AI 

• What ISO/IEC 42001 and the EU AI Act require of buyers 

• The practical shift needed to treat procurement as a governance function 

Buying AI is not like buying software 

Traditional software procurement has identifiable risks. They are relatively stable, reasonably well understood, and most procurement teams have frameworks for managing them. AI procurement is different in kind, not just in degree. AI systems, including traditional software integrating AI features, learn and evolve; their behaviour at the point of contract may not reflect their behaviour six months into deployment. Their supply chains span foundation models, infrastructure providers, APIs, and application layers, and accountability can be genuinely difficult to establish across that stack. 

Dependency is also a different problem. Unlike traditional software, AI systems incorporate data and learning over time, which makes switching suppliers significantly more complex than it initially appears. Supplier ‘lock-in’ is not always visible at the point of purchase. It tends to become apparent when an organisation tries to exit, but by that stage the governance exposure is already embedded. 

An uneven playing field 

The AI procurement landscape has a structural asymmetry built into it. Vendors are incentivised to sell. Executives are under pressure to adopt at pace. Procurement teams, meanwhile, are frequently under-equipped to interrogate AI systems with the rigour the situation demands. The gap between the risk organisations is taking on and their capacity to manage it is growing. 

That asymmetry shows up in contract terms. Liability terms in AI vendor contracts are frequently weighted against buyers, with 38% of AI vendors applying customer liability caps compared to 44% in standard SaaS contracts. Buyers are routinely taking on risk they cannot see, for which they have not planned, and which standard processes were not designed to surface. Third-party AI risk management begins not after a contract is signed, but in how that contract is approached in the first place. 

The risks organisations are underestimating 

AI supply chain risk falls into four interconnected areas, each frequently underestimated. 

Supply chain accountability is complicated by the multi-layered architecture of most AI systems and traditional software integrating AI features. Three companies account for 88% of enterprise LLM API usage, which means buyers transacting with a specialist vendor are often, in practice, dependent on a handful of foundational infrastructure providers. Smaller vendors may be built on top of those same foundational models. Mid-contract acquisitions can alter arrangements without warning. Establishing where responsibility lies requires a level of due diligence that standard procurement templates are not built to provide. 

Data and IP risks are among the most significant and most overlooked. Business data may be used to train AI models without explicit disclosure. IP ownership over model outputs, metadata, and derived data often remains unresolved at the point of contract. In 2026, organisations on lower-tier GitHub Copilot contracts discovered that interaction data from private repository sessions was being used for model training because AI-specific data processing protections had not been included in their agreements. It is the kind of gap that only becomes visible after the fact. 

Supplier dependency compounds this. Limited interoperability between platforms restricts the ability to move when needed. Contractual arrangements should secure the right to interrogate, exit, and transfer AI configurations, but that conversation needs to happen before an organisation signs, not after it has become operationally reliant on a vendor. 

Agentic AI sits in a category of its own. These systems traverse multiple datasets and platforms without human authorisation at each step, and the governance implications of that aggregated access are not yet well understood by most buying organisations. Some vendors are already excluding key AI-related harms from liability clauses in agentic AI contracts. Buyers are carrying exposure they may not know is there. 

What good practice looks like 

Organisations do not need to build AI procurement governance from scratch. ISO/IEC 42001 includes specific controls for AI supplier management Clause 8.2, which covers AI supply chain risk management, requires formal processes for evaluating supplier risk, setting minimum contractual standards, and ensuring ongoing monitoring of third-party AI providers. That is a reasonable starting checklist for any procurement team before an AI contract is signed. 

The EU AI Act adds obligations on the buyer side. For high-risk AI systems, organisations must verify that a conformity assessment has been completed, confirming that risk management, quality control, and performance monitoring processes are in place. Where an organisation customises or white labels a third-party system, it may become a provider under the Act, with documentation and registration obligations that depend on having received sufficient information from the original vendor at the procurement stage. These are not peripheral compliance considerations. They affect the terms on which a purchase should proceed. 

What this means for organisations 

Buying an AI system means more complex, and dynamic monitoring of, choices about data sovereignty, liability, regulatory compliance, and long-term operational dependency. Those choices need governance and compliance stakeholders involved from the outset, not brought in after the commercial decision has been made. 

Your organisation needs to design frameworks, contract terms, and internal accountability structures to manage the specificities of AI systems, including traditional software integrating AI features.   

Trilateral's full report on AI Procurement and supply chain risk sets out a practical framework to get you started.