AI Tokens: The Hidden Commercial Risk In Enterprise AI Deals

For many organisations, AI procurement discussions continue to focus on functionality, i.e. whether an artificial intelligence (“AI”) system can summarise contracts, draft reports, analyse data or assist or even replace employees. While these capabilities remain important, a more fundamental issue is beginning to emerge beneath the surface: the economics of AI infrastructure and specifically the role of AI tokens. Tokens are rapidly becoming one of the most significant commercial, operational and legal risk drivers in enterprise AI deployments. Increasingly, sophisticated customers are recognising that many AI systems are not truly ‘fixed fee’ products, but instead operate much like cloud consumption models, where cost, performance and scalability are directly linked to usage patterns that are often opaque.

This shift has particular significance in sectors such as banking, healthcare, telecommunications, mining, higher education and large-scale professional services, where AI systems are deployed across high-volume environments involving extensive document processing, data analysis, or user interaction. In these contexts, seemingly minor variations in usage can translate into substantial downstream cost and performance implications.

AI systems process information in a fundamentally different manner to human users. Rather than interpreting language holistically, they break text into smaller computational units known as “tokens”. These tokens may represent a full word, part of a word, punctuation, numbers, or formatting characters. The practical effect is that even relatively small pieces of text can translate into hundreds or thousands of tokens, while enterprise-scale datasets may involve billions. Every interaction with an AI system, whether uploading documents, querying data, generating outputs, or running automated workflows, consumes tokens. In many commercial arrangements, these tokens function as the underlying unit by which customers are charged.

This represents a significant departure from traditional software licensing models, which have historically been based on predictable metrics such as users, devices, or fixed annual subscriptions. AI systems instead increasingly adopt variable consumption pricing, where costs fluctuate based on factors such as token usage, model complexity, context window size, throughput and computational intensity. This introduces a layer of unpredictability that many organisations have not yet fully internalised.

One of the most immediate risks is consumption unpredictability. An organisation may implement an AI system on the assumption of moderate usage, only to discover that employee adoption accelerates rapidly, automated processes generate substantial background activity, or large-scale initiatives such as due diligence or compliance reviews drive unexpected spikes in token consumption. The resulting cost profile can closely resemble the ‘bill shock’ experienced in early cloud computing adoption. This risk is compounded by the fact that much of the token consumption is not visible to the end user.

Behind a single user interaction, an AI system may execute multiple backend processes, including orchestration across different models, retrieval of data from knowledge bases, summarisation steps, and reasoning layers. Each of these processes consumes tokens, often without clear visibility to the customer. As a result, a seemingly simple query can generate disproportionately high consumption, making it difficult for organisations to predict or control costs effectively.

Another emerging concern is token inflation. There is no uniform standard for how different AI models and providers tokenise content which may result in the same document generating materially different token counts depending on the model or provider. In addition, model upgrades or changes to underlying prompts may alter token consumption profiles over time. This creates a risk analogous to tariff escalation in utility services, where customers face increasing costs without corresponding increases in value or transparency.

Curiously, an unintended consequence of tokenised pricing is that, as many organisations are beginning to discover, the costs of tokens might outweigh the efficiencies and cost-savings gained by replacing human beings with AI systems. A large part of this is attributed to the fact that AI systems are notoriously expensive to create and sustain and the cost of these will inevitably have to be borne by customers. This places an interesting spin on the debate as to whether AI will replace human beings.

The implications of these dynamics extend beyond procurement and into broader legal and operational considerations. Token economics directly affect budget forecasting, operational resilience, vendor dependency and governance frameworks. For example, a legal AI system may become commercially unviable during periods of intensive litigation activity, a healthcare system may experience performance degradation under heavy diagnostic workloads, or a financial institution may struggle to forecast AI expenditure in the face of fluctuating usage patterns. In this sense, token governance is evolving from a technical issue into a strategic and, increasingly, board-level concern.

In response, organisations are beginning to adopt a more structured approach to contracting for AI systems, drawing on lessons learned from cloud computing and telecommunications. A central theme in these arrangements is transparency. Customers require detailed visibility into token consumption, including how usage is calculated, how it is allocated across different workflows or business units, and how backend processing contributes to overall consumption. Without this level of insight, meaningful governance is not possible.

Pricing protections are also critical. Customers are increasingly seeking to limit the vendor’s ability to unilaterally increase token prices, alter tokenisation methodologies, or introduce cost changes through model substitution. This may be achieved through fixed pricing periods, controlled escalation mechanisms, or benchmarking rights. Closely linked to this is the need to address context window dependencies, ensuring that where large-scale processing is integral to the solution, the associated capabilities, performance levels, and pricing structures are contractually secured. Financial governance mechanisms have also become essential. Organisations are negotiating spend caps, usage thresholds and automated alerts to prevent uncontrolled cost escalation. In parallel, audit and verification rights are being used to ensure that billing methodologies are accurate and that token calculations can be independently validated. These protections are particularly important in complex environments where multiple AI models or services are orchestrated together.

Token-based pricing models also give rise to a new form of vendor lock-in. AI solutions often rely on components such as prompts, embeddings, vector databases, and workflow orchestration layers, all of which may be difficult to migrate. This has prompted increased focus on exit and portability provisions, including requirements for data export, migration support, and continued service during transition periods. Model substitution rights are also under scrutiny, with customers seeking to ensure that any changes to underlying models do not adversely affect performance, cost, or compatibility.

Looking ahead, it is likely that AI-related disputes will increasingly focus on consumption economics rather than purely technical issues. While risks such as hallucinations, privacy breaches, and intellectual property concerns will remain relevant, disputes are likely to centre on opaque billing practices, unpredictable cost profiles, performance degradation, and vendor dependency. In many respects, this reflects a familiar trajectory from earlier waves of cloud and telecoms adoption, albeit with greater volatility and less transparency.

AI procurement is therefore evolving into a complex hybrid of cloud infrastructure contracting, data governance, operational resilience planning, and utility-style consumption management. At the centre of this shift are tokens. Organisations that fail to understand and govern token economics risk exposing themselves to significant commercial and operational challenges. Ultimately, the most significant risk in enterprise AI deployments may not lie in the model itself, but in the commercial architecture that underpins it.

The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.