How to Budget for AI When the Price Model Is Shifting

AI used to be simple to budget. You bought a licence, paid a fixed annual fee, and knew exactly what you were committing to. That model is largely gone.

Most AI capability today is priced on consumption: the number of tokens processed, requests made, or tasks completed. That means the cost base moves with how much you use the system, how it is designed, and how broadly adoption spreads across your organisation. A solution that costs a predictable amount in a contained pilot can look very different once it is embedded in operations.

The central question for budget owners is not whether AI is cheap today. It is whether total spend stays controlled as usage grows.

Key takeaways

• AI pricing has shifted from fixed licences to consumption-based models, making spend harder to forecast with traditional budgeting tools.
• Total cost includes more than the provider fee: integration, data preparation, governance, and ongoing usage all contribute.
• Unit prices may fall over time, but total spend can still rise if usage grows faster than costs per token decrease.
• The safer approach is to budget conservatively, design for efficiency from the start, and release later-phase budget only once value is proven.

Why is AI harder to budget for now?

Traditional software licences gave finance teams something predictable: a fixed price per seat or per year. AI has moved away from that model. Most providers now charge based on usage, which means the cost of your AI deployment changes depending on how much it is used, how it is built, and who is using it.

That shift introduces three compounding challenges for budget owners:

• The pricing unit has changed. Instead of paying for access, you are paying for consumption, typically measured in tokens (the units of text or data processed by the model). A longer prompt, a more complex retrieval step, or a model that generates verbose outputs will cost more than a tightly designed alternative doing the same job.
• Cost now scales with adoption. A pilot involving one team and a narrow use case may be affordable and easy to track. The same solution rolled out across claims, service, and operations looks very different on a cost basis, even if the per-token price stays the same.
• The headline price is not the total cost. Integration work, data preparation, monitoring, human oversight, and governance all sit alongside the provider fee. Deloitte's analysis of AI token economics notes that organisations routinely underestimate total AI spend because they focus on model access costs and overlook the surrounding infrastructure.

The practical implication: AI spend is now closer to managing a utility than buying software. Volume, design choices, and organisational behaviour all affect the bill.

Are AI prices really going to rise?

The honest answer is that nobody can say with certainty. What is clear is that total AI spend can rise even when unit prices fall, because usage tends to grow faster than the cost per token decreases. That is the more useful framing for budget planning.

Current AI pricing reflects an intensely competitive market where providers are investing heavily to win adoption. Whether that translates into pricing normalisation over time is a reasonable expectation, but it is not a guaranteed outcome. The prudent position is to plan for volatility rather than assume today's economics will hold.

The planning implication is straightforward: even if your provider's unit price stays flat or decreases, your total AI spend is likely to increase as the organisation uses more of the system. Budget owners who treat today's pilot costs as a stable baseline for multi-year planning are taking on more risk than the numbers suggest.

The safe assumption: budget for total spend to rise, and design the deployment to keep that rise proportionate to the value being generated.

What does this mean for financial services and insurance?

In most sectors, uncontrolled AI spend is primarily a finance problem. In regulated financial services and insurance, it is also a governance problem. The two are harder to separate than they might appear.

Regulated organisations face a specific version of this challenge. AI adoption rarely stays contained to one team or one use case. It typically starts in a pilot, proves useful, and then spreads into claims processing, customer service, documentation, knowledge retrieval, and productivity workflows. Each expansion multiplies both usage and complexity, and the cost base follows.

The sector-specific implications worth building into your budget planning include:

• Consumer Duty obligations do not pause for cost overruns. If AI is being used in customer-facing journeys, the FCA's expectation of good outcomes applies regardless of whether the deployment is still being optimised. Governance costs are not optional.
• Insurance teams face particular exposure around claims and professional indemnity. AI-assisted decisions in claims assessment or underwriting require audit trails, oversight, and records that add to the total cost of a deployment, not just the token bill.
• Fragmented adoption creates fragmented spend. When different teams procure AI tools independently, the organisation ends up with duplicated costs, inconsistent controls, and no single view of what is being spent or why.
• Operational resilience expectations mean you need continuity planning for AI-dependent workflows. That is an additional cost layer that rarely appears in a provider's headline pricing.

The regulated environment does not make AI harder to justify. It does make it harder to run informally. Budget discipline and governance discipline need to move together.

How do you keep AI cost under control?

Cost control in AI is mostly an architecture and discipline problem, not a negotiation problem. The levers available to budget owners are more powerful than most realise, and most of them sit in decisions made before a provider contract is signed.

The five cost control levers

1. Right-size the model to the task. Not every task needs a frontier model. A large language model used to classify a customer query into one of ten categories is doing the equivalent of using a combine harvester to cut a lawn. A smaller, purpose-built model can handle narrow, repetitive tasks at a fraction of the cost, with comparable accuracy for that specific job. Reserve premium models for genuinely complex work: nuanced advice, multi-step reasoning, or unstructured document analysis.
2. Design prompts and retrieval for economy. Verbose prompts, poorly scoped retrieval pipelines, and workflows that pass unnecessary context to the model all burn tokens without improving outputs. This is not a theoretical saving: architecture choices can change token consumption materially on the same task.
3. Run cost-controlled pilots with exit criteria. Before committing to wider rollout, set a usage cap, a time boundary, and a clear set of business outcomes the pilot needs to demonstrate. If it does not hit the criteria, the budget does not follow it.
4. Consolidate spend visibility. When AI tools are procured across teams without central oversight, costs fragment and become invisible. A single view of what is being spent, by whom, and on which use cases, is a basic governance requirement and a practical cost management tool.
5. Preserve portability where possible. Organisations that build deep dependencies on one provider's proprietary stack lose negotiating leverage over time. Where the architecture allows, keeping options open protects against future pricing changes.

Worked example: A mid-market insurer running an AI assistant for policy queries could use a frontier model for every interaction, or route straightforward queries to a smaller, lower-cost model and escalate complex cases to the premium tier. The routing approach can reduce token costs on routine queries significantly, without any reduction in quality for the interactions that need it. The exact saving depends on the query mix, but the principle applies broadly.

For more on how this plays out in a virtual agent deployment, the AI virtual agent implementation guide covers the architecture decisions that affect both performance and cost.

How should you plan an AI budget across a budget cycle?

The phased approach works better than a single annual commitment for AI, precisely because the cost base is variable and the evidence base is still building. Here is a practical structure for this budget round and into next year.

• Start with a baseline. Document current pilots, expected use cases, likely user volumes, and the main technical cost drivers. Do not carry forward assumptions from a fixed-licence world.
• Separate the budget lines. Platform access, implementation, data preparation, governance and oversight, and ongoing consumption should each have their own line. Bundling everything under one AI heading makes it impossible to see where money is actually going.
• Model three scenarios. A cautious case (usage stays close to pilot levels), an expected case (moderate adoption growth), and a scaled case (broad rollout across teams). The gap between cautious and scaled is usually larger than budget owners expect.
• Tie phase two budget to evidence. Define in advance what the first phase needs to demonstrate: a cost-per-outcome figure, a usage level, a governance milestone. The budget for broader rollout should follow the evidence, not the original business case.
• Build in review points. Quarterly reviews of actual versus forecast consumption allow course corrections before overruns compound.

If you are building this for the first time, the communications stack cost guide covers the broader technology cost picture that AI budgets sit within, and the CCaaS selection guide covers the platform decisions that affect what you pay long term.

Where this leaves you

The budgeting problem is manageable. What makes it harder than it needs to be is applying fixed-software thinking to a variable-cost system.

Budget for total spend to be more volatile than today's pilots suggest. Design the deployment to keep costs proportionate to the value being generated. And phase your investment so later-stage budget is earned by evidence, not assumed from the start.

The one-line version: budget conservatively, engineer for efficiency, expand only where value is proven.

If you want a structured approach to take into your next planning round, the Budgeting for AI planning guide sets out the full framework, including scenario modelling, governance considerations, and the questions to ask any provider before you commit.

Or if you would prefer to talk through the cost side of your AI plan directly, book a short, no-obligation conversation with the Fortay Connect team.

Frequently asked questions

Why is AI harder to budget for now?

AI is harder to budget for because pricing has shifted from fixed licences to consumption-based models. Cost now moves with tokens, usage volume, prompt length, retrieval design, and how widely the system is adopted across the organisation.

What is token-based pricing for AI?

Token-based pricing means you pay for the amount of text or data the model processes and generates. A short, simple interaction uses fewer tokens than a long, complex one, which means the same AI use case can cost very different amounts depending on how it is designed and how often it runs.

Are AI prices going to rise or fall?

The safer assumption for planning purposes is that total AI spend will rise even if some unit prices fall. Providers are competing hard, so individual model prices can move down, but usage tends to expand faster than costs per token decrease, which pushes overall spend up.

How much does AI cost at scale?

AI at scale usually costs more than the headline licence or access fee suggests. The total includes integration, data preparation, governance, monitoring, human oversight, and ongoing consumption. The wider the rollout, the more those surrounding costs matter relative to the provider fee.

How do you keep AI costs under control?

The main levers are right-sizing the model to the task, designing prompts and retrieval efficiently, running cost-controlled pilots with clear exit criteria, consolidating spend visibility across teams, and preserving portability so you retain leverage over future pricing decisions.