Your Next AI Contract Will Bill You by the Token. Is Your Finance Team Ready?

ROI & Cost Optimisation

14 April 2026 | By Ashley Marshall

Most UK IT leaders are still negotiating AI contracts the way they bought Office licences in 2012. Vendors have moved on, and the next renewal cycle will make that gap very expensive.

The Pricing Shift That Caught Enterprise Buyers Off Guard

For most of the past decade, buying enterprise software was straightforward. You counted your users, agreed a price per seat, and knew roughly what you would pay each year. IT finance could plan around it. Procurement could benchmark it. Legal could review it once and file it away.

That model is collapsing under the weight of AI. According to a recent report from IDC, by 2028 pure seat-based pricing will be obsolete as AI agents replace repetitive tasks with digital labour, forcing 70 per cent of vendors to refactor their value proposition into entirely new pricing models. That is not a distant forecast. It is already happening in the contracts being signed right now.

Salesforce launched Agentforce in late 2024 priced at two US dollars per conversation. ServiceNow moved to a credit-based system for its AI capabilities. SAP introduced AI Units as an underlying consumption currency sitting beneath its seat licences. Workday's CEO told analysts the company was focused not just on seats but on revenue per seat, a signal that usage-based premiums are coming. Microsoft Copilot, priced at 30 US dollars per user per month on top of an existing Microsoft 365 licence, has made the blended cost of AI-enabled productivity tools significantly higher than many organisations budgeted for when they first signed up.

What makes this particularly difficult for UK buyers is the opacity of the new units. Tokens, conversations, credits, AI units, Flex Credits, interactions and events are not interchangeable terms. Each vendor defines them differently, and many contracts are deliberately vague about exactly what triggers a billing event. As one analyst quoted in CIO magazine put it: these units are designed to obscure the value exchange, giving vendors a margin floor while customers absorb the risk of overconsumption.

The shift is not theoretical. According to SaaS management firm Zylo, 78 per cent of IT leaders they surveyed in early 2026 reported unexpected charges on their SaaS contracts due to consumption-based or AI pricing models. The same research found that organisations spent an average of 1.2 million US dollars on AI-native applications in 2026, a 108 per cent year-on-year increase. Both figures point to the same underlying problem: spend is accelerating faster than governance.

What Usage-Based Pricing Actually Means in Practice

Understanding the mechanics matters before you can manage the risk. Consumption-based AI pricing comes in several flavours, and vendors often layer them on top of each other within the same contract.

Token-based pricing is the foundational model for large language model interactions. You pay per thousand tokens of input and output. OpenAI, Google, and Anthropic all operate this way at the API level. The catch is that token consumption is highly variable depending on prompt length, context window size, and the complexity of the task. A simple summarisation workflow might cost fractions of a penny per run. An agent that pulls context from multiple sources, reasons across long documents, and generates a structured output can consume orders of magnitude more. Without visibility into which workflows are running and how often, finance has no basis for forecasting.

Conversation-based pricing, used by Salesforce Agentforce in its original form, charges per resolved or attempted interaction. The problem identified by procurement advisers at UpperEdge is that many customers never clarified what constitutes a conversation in their specific use case. Does a single customer service interaction that escalates through three agent handoffs count as one conversation or three? Is a failed resolution still billed? Contracts that lack this precision leave buyers exposed when the invoice arrives.

Credit-based systems, used by SAP, ServiceNow, and increasingly others, wrap token or interaction costs into a proprietary currency. You buy a bundle of credits, use them across AI features, and top up when they run out. The opacity here is twofold: you cannot easily compare credit costs between vendors, and the credit consumption rate for different features is rarely published clearly upfront. ServiceNow seat licences, for example, come with a provisioned limit of credits. Exceed that limit and you are topping up at additional cost.

Hybrid models, which combine a per-seat base with usage overages, are the most common structure in enterprise agreements right now. They look predictable on the surface, but the usage ceiling is where the risk lives. Jason Andersen, principal analyst at Moor Insights and Strategy, notes that vendors will typically offer a large upfront token bundle at a lower per-token rate, then charge at a higher rate for anything beyond that threshold. For organisations that are still learning how their employees use AI, exceeding the bundle in the first year is close to inevitable.

What this means in practice is that finance teams need to understand the billing mechanics of every AI tool in their stack, not just the headline price. That requires close collaboration between IT, procurement, and finance at the point of contract negotiation, not after the first overage notice.

The Counterargument: Is Seat-Based Pricing Actually Coming Back?

It would be unfair not to address the argument that the market is already correcting itself. In December 2025, The Register reported that Salesforce CEO Marc Benioff had effectively confirmed a partial retreat from pure consumption pricing, with the company's Agentic Enterprise License Agreement (AELA) offering a seat-based package with reusable credits bundled in. Benioff acknowledged that customers had pushed back hard on per-conversation pricing in favour of more flexibility and predictability.

Gartner's senior software licensing expert Jan Cook told The Register that customers were still cautious about GenAI precisely because of pricing unpredictability, and that the return to seat-based structures was a market response to that caution. His observation deserves weight: if seat shrinkage from AI-driven headcount reduction were really happening at scale, we would have seen contract renegotiations by now. We have not, which suggests most organisations are deploying AI to augment rather than replace workers, and seat counts are holding steady.

The honest answer is that the pricing landscape is genuinely hybrid and still evolving. Pure usage-based billing has proven difficult to sell to procurement teams who need budget predictability. Pure seat-based billing does not work for agentic AI because digital workers do not map to headcount. The market is converging on a middle ground: seat-based or user-based licences with embedded usage allowances, above which consumption charges kick in.

This middle ground does not reduce the FinOps challenge. It reframes it. The risk is no longer that you will receive an entirely consumption-based invoice you cannot forecast. The risk is that you will sign what looks like a familiar licence agreement, miss the embedded usage cap buried in the order form, and be surprised by overage charges six months later. The hidden ceiling is the new blank cheque.

Sanchit Vir Gogia, chief analyst at Greyhound Research, put it starkly: vendors are transferring the cost volatility of AI compute to customers while monetising customer-side productivity gains as margin. Whether that volatility is expressed as pure consumption pricing or as overage clauses in a hybrid seat deal, the underlying dynamic is the same. Procurement must develop fluency in AI mechanics, system telemetry, and the behavioural signals that trigger spend. Without it, contracts will outpace control and budgets will unravel fast.

What UK Organisations Are Getting Wrong at Renewal

UK enterprise buyers face several compounding difficulties that their US counterparts do not always share. Sterling-denominated budgets are exposed to dollar-denominated AI pricing, meaning that exchange rate movements affect the real cost of AI contracts even when the nominal terms have not changed. The weakness of sterling against the dollar over the past three years has already made US cloud and SaaS contracts materially more expensive in pound terms without any renegotiation. Adding consumption variability on top of that currency risk creates a genuinely difficult forecasting environment.

Beyond currency, the UK regulatory context adds layers of compliance spend that are not always factored into AI budgets. Organisations operating under the UK GDPR and the Data Protection Act 2018 have obligations around how personal data is processed by AI systems, including systems operated by third-party vendors. The ICO has published guidance on AI and data protection, including expectations around transparency, data minimisation, and the right to explanation for automated decisions. Compliance with these obligations often requires additional tooling, audit logging, and data residency controls that carry their own cost. If your AI vendor processes personal data on US infrastructure and you need UK or EEA residency to satisfy your data protection obligations, that premium needs to be in the contract and in the budget.

The second common mistake is treating AI procurement as a technology decision rather than a financial governance decision. Many UK organisations are signing AI contracts at the departmental level, with line-of-business leaders buying tools on expense accounts or small procurement thresholds that do not trigger central review. Zylo's research describes this as shadow AI, and it is expanding spend and risk simultaneously. When consumption-based costs from shadow AI tools eventually land on a corporate card or expense claim, they are often unrecognised until they are already significant.

Third, organisations are renewing existing vendor relationships without renegotiating the pricing basis. If you signed a Microsoft 365 enterprise agreement two years ago and have since adopted Copilot as an add-on, the pricing terms for that add-on may not have been negotiated with the same rigour as the original agreement. Adam Mansfield of advisory firm UpperEdge has documented cases where enterprises exceeded Salesforce Agentforce conversation thresholds faster than expected and found their order forms void of the necessary detail to resolve the dispute. The vendor defines the unit; the customer absorbs the overrun.

Finally, there is the agentic AI problem. Once you deploy autonomous AI agents that operate without human initiation of each individual action, you lose the natural throttle that human usage provides. A human user logs off at the end of the day. An agent does not. If your consumption pricing includes agent-initiated interactions, your AI spending can accrue overnight, at weekends, and across time zones in ways that your existing budget cycles are not designed to monitor. This is not hypothetical: Jason Andersen from Moor Insights noted that usage monitoring systems need to be coded to trigger an alarm when usage hits a threshold, with policies built into the agents themselves capping token usage. Most UK organisations have not yet implemented either.

Building a FinOps Practice for AI Spend Before Your Next Renewal

FinOps, defined by the FinOps Foundation as the cultural practice of bringing financial accountability to technology spend through cross-functional collaboration between engineering, finance, and business teams, has strong heritage in managing cloud costs. Extending it to AI spend is not a conceptual stretch. The challenge is institutional: most UK organisations built their FinOps capability around AWS, Azure, or Google Cloud, where billing is relatively transparent and tooling is mature. AI spend management is still catching up.

The first step is achieving visibility. You cannot manage what you cannot see. That means deploying SaaS management tooling such as Zylo, Torii, or Productiv that can discover AI-enabled applications across the estate, including those being expensed outside central IT. It means requiring vendors to provide usage dashboards as a contractual term, not an optional feature. And it means instrumenting your own AI workloads with observability tooling at the application level, tracking token consumption, agent invocations, and credit burn per use case, per team, and per business unit.

The second step is establishing a FinOps function or extending an existing one. J.R. Storment, Executive Director of the FinOps Foundation, has drawn an explicit parallel between the current AI spend explosion and the early cloud adoption wave a decade ago, when individuals started signing up for AWS accounts on credit cards until the bills hit a critical threshold. The solution to uncontrolled cloud spend was not a blanket ban on cloud usage. It was shared ownership of costs, with engineering teams understanding the financial consequences of their architectural choices. The same logic applies to AI: developers and product teams need unit economics visibility, not just access to powerful tools.

Third, involve finance and procurement in AI tool selection from the beginning, not just at the point of renewal. Usage-based pricing structures require finance to model scenarios based on anticipated usage volumes, with upside and downside cases. That modelling requires engineers and product managers to provide honest estimates of how frequently workflows will run, how many tokens a typical interaction consumes, and what the growth trajectory looks like. The collaboration required to produce those estimates is itself valuable: it forces a shared understanding of the economics of AI deployment that most organisations currently lack.

Fourth, use the period between now and your next renewal to baseline actual consumption. Jason Andersen of Moor Insights recommends asking vendors for a stay of execution of at least a year before a new pricing model goes into effect, giving your team time to measure current usage patterns before committing to a volume commitment. Even if you cannot renegotiate the pricing model itself, twelve months of usage data puts you in a far stronger position to negotiate sensible ceilings and realistic overrun terms.

Tooling worth evaluating for AI-specific FinOps includes CloudHealth, Apptio Cloudability, and the native cost management dashboards provided by major cloud platforms. For vendor-specific AI usage, tools like Datadog, Dynatrace, and OpenTelemetry-based observability stacks can capture model invocations and token counts at the infrastructure level. None of them solve the problem on their own, but visibility is the prerequisite for everything else.

What to Negotiate Before You Sign Your Next AI Contract

Procurement strategy for AI contracts needs to evolve beyond the tactics that worked for per-seat software. Here is what experienced advisers are recommending for the current environment.

Define every billing unit in writing. Before signing, ask your vendor to document precisely what constitutes a token, conversation, credit, interaction, or AI unit in the context of your specific deployment. If you are deploying a customer service agent, establish in the contract whether a multi-turn dialogue counts as one conversation or multiple. If credits are consumed by background processes running after business hours, get that confirmed explicitly. As Aaron Perkins of Market-Proven AI advises: ask the hard questions, because a usage model means different things to different people.

Negotiate hard ceilings with written permission requirements for overruns. Rather than accepting open-ended consumption, push for a contractual cap on usage spend, above which the vendor must contact you in writing before continuing to bill. If the vendor will not accept an absolute ceiling, push for automatic alerts at 70 per cent and 90 per cent of the agreed volume, with a 48-hour pause before additional charges accrue. This shifts the onus for managing overruns from you to the vendor.

Include fraud protection clauses for agentic workloads. If your AI deployment involves autonomous agents, include contract language that protects you from charges generated by external attacks. As Jason Andersen of Moor Insights has noted, if a competitor or malicious actor bombards your customer-facing AI agent with automated requests to drive up your usage bill, you need contractual protection against paying for traffic you did not generate. This is a specific and growing risk for organisations with publicly accessible AI interfaces.

Require data residency and processing location commitments in writing. UK GDPR obligations require you to know where personal data is processed and to ensure adequate protections are in place for any transfers outside the UK. In a consumption-based model where compute is dynamically allocated, the vendor's data processing agreement needs to specify whether UK or EEA residency can be guaranteed and under what conditions processing might shift. Do not assume a generic standard contractual clause covers the specifics of an AI workload.

Ask for multi-year price protection on per-unit rates. Consumption-based pricing exposes you to rate changes at renewal as well as volume overruns during the term. Negotiate fixed per-token or per-credit rates for the duration of your agreement, with any increases capped at a defined percentage indexed to UK CPI. This is a harder ask for smaller organisations, but mid-market and enterprise buyers have enough leverage to get it into the order form if they push for it early.

Finally, build in a right to audit. For significant AI spend commitments, include a contractual right to request a usage report in vendor-neutral format, broken down by product, feature, and billing unit, no less than quarterly. This is standard practice in cloud agreements and should be standard in AI licensing as well.

Frequently Asked Questions

How does usage-based AI pricing differ from traditional per-seat software licensing?

Per-seat licensing gives you a fixed annual cost per user regardless of how much they use the software. Usage-based pricing charges you for what is actually consumed, measured in tokens, conversations, API calls, or vendor-specific credits. The cost varies with activity levels, making it harder to forecast and easier to exceed budget without realising it.

Is Salesforce actually moving back to seat-based pricing for Agentforce?

Partly. Salesforce introduced its Agentic Enterprise License Agreement (AELA) in late 2025, which bundles seats and reusable credits in a single package. However, Gartner analysts note that most vendor seat licences now include embedded usage limits, above which consumption charges apply. It looks like a seat deal but still carries usage overrun risk.

What is FinOps and why does it matter specifically for AI spend?

FinOps is the practice of bringing financial accountability to technology spend through shared ownership between engineering, finance, and operations teams. It originated in cloud cost management. AI spend now requires the same discipline because consumption-based pricing creates the same dynamic: costs that vary with usage and accrue without visibility unless actively managed.

How does UK GDPR affect AI procurement decisions?

UK GDPR requires organisations to know where personal data is processed, to minimise data collection, and to be able to demonstrate lawful basis for automated processing. AI contracts involving personal data need explicit data processing agreements, data residency commitments, and clarity on automated decision-making. The ICO has published specific guidance on AI and data protection that procurement teams should review before signing.

What tools can help UK businesses track AI consumption across their vendor estate?

SaaS management platforms such as Zylo, Torii, and Productiv can discover AI-enabled applications and provide usage visibility. For workload-level observability, Datadog, Dynatrace, and OpenTelemetry-based stacks can capture token consumption and model invocations. Native dashboards from cloud providers also provide some visibility for AI services consumed through AWS, Azure, or Google Cloud.

What is shadow AI and why is it a budget risk?

Shadow AI refers to AI tools adopted by employees or departments without central IT approval, typically expensed individually or purchased below procurement thresholds. Because these tools often carry consumption-based pricing, their costs can grow significantly before they are visible to finance or IT governance teams. Zylo research found shadow AI is one of the fastest-growing sources of unexpected SaaS spend.

Can autonomous AI agents generate costs overnight without human intervention?

Yes. Unlike human users who log off at the end of the day, autonomous AI agents can run continuously and generate billable interactions around the clock. For consumption-based contracts, this means spend can accrue outside normal business hours without any human initiating the activity. Usage monitoring systems with threshold alerts and token caps built into the agents themselves are needed to control this.

How should UK SMEs approach AI pricing negotiations if they do not have the leverage of large enterprises?

Smaller organisations have less leverage on per-unit rates, but can still negotiate usage ceilings, written overage notification requirements, and quarterly usage reports. Working with an independent software licensing adviser or procurement consultancy gives smaller buyers access to benchmark data and negotiating tactics that are difficult to develop in-house without volume purchasing history.