Outcome-Based AI Budgets: Moving From Token Spend To Cost Per Useful Result In UK Firms

ROI & Cost Optimisation

10 July 2026 | By Ashley Marshall

Token spend is a useful invoice line, but it is a weak management metric. UK firms need AI budgets that show what each workflow costs per resolved ticket, approved document, qualified lead, checked invoice or avoided manual hour.

Token Spend Is A Meter, Not A Budget Strategy

Token spend is where many AI finance conversations start because it is visible, measurable and easy to put in a spreadsheet. Finance can ask how many input tokens and output tokens a team used last month. Engineering can compare OpenAI, Anthropic, Google Gemini, Azure AI Foundry or Amazon Bedrock rates. Procurement can negotiate discounts. That is all useful. It is also incomplete. A token is a unit of model consumption, not a unit of business value.

The Office for National Statistics gives the UK business context. In its 2025 article on management practices and AI adoption in UK firms, the ONS reported that AI was adopted by 9 percent of UK firms in 2023 and was projected to be adopted by 22 percent in 2024. It also found that the most common barriers to AI adoption were difficulty identifying activities or business use cases at 39 percent, cost at 21 percent and AI expertise or skills at 16 percent. Those figures matter because they show the real budget problem: firms do not only need cheaper tokens. They need clearer use cases and better evidence of value.

Outcome-based AI budgeting starts from the other end of the ledger. Instead of asking how much did the model cost, ask what did the workflow produce that the business can recognise as useful? A customer service assistant might be measured by cost per correctly resolved enquiry or cost per clean handoff. A finance agent might be measured by cost per invoice checked and approved without rework. A sales research workflow might be measured by cost per qualified account brief accepted by the sales team. A policy search tool might be measured by cost per grounded answer that staff actually use.

What this means in practice is that token cost becomes one input inside a larger unit economics model. The budget should include model calls, retrieval, orchestration, logging, human review, exception handling, vendor seats, cloud infrastructure and remediation time. That may sound more complex, but it is the only way to see whether AI is lowering the cost of work or just moving spend from salaries and SaaS into a new API invoice. Internal pieces such as AI FinOps unit economics already point in this direction: the metric that matters is not the cheapest prompt, but the cheapest reliable result.

Define Useful Results Before You Approve More Spend

The biggest budgeting mistake is to fund AI projects before the useful result is defined. A department asks for budget to use generative AI for productivity, customer service, sales enablement or operations, and the approval is based on plausible savings. That may be enough for a small experiment. It is not enough once usage spreads across teams and invoices become recurring. The first budget question should be: what result would make this cost worth paying every month?

A useful result is not always a completed automation. In regulated, sensitive or high-trust workflows, the result may be a better prepared human decision. For example, an HR assistant may produce a compliant interview pack for review rather than make a recruitment decision. A legal support tool may produce a clause comparison for a solicitor. A finance workflow may flag invoice exceptions for approval. A customer support co-pilot may draft a reply that a human sends. The common thread is that the result is specific, accepted by the workflow owner and measurable after the fact.

GOV.UK's AI Opportunities Action Plan is useful because it treats AI adoption as an economic and public service productivity issue, but it also says government purchasing power needs leadership and radical change, especially in procurement. It also recommends expanding the AI Research Resource by at least 20 times by 2030 and ensuring a range of hardware providers to avoid vendor lock-in and ensure value for money. That value-for-money language is exactly the mindset private firms need. A budget line is not justified because AI is strategic. It is justified because the firm can say what result it bought.

What this means in practice is to write a result definition for every funded AI workflow. Use a simple sentence: this workflow is successful when it produces X accepted outputs per month, with Y percent passing quality checks, at no more than Z cost per useful result. For support, X might be resolved enquiries. For operations, it might be completed exception checks. For marketing, it might be briefs that reach human approval. For finance, it might be invoices matched without manual rework. The budget then has a target and a stop rule. If cost per useful result rises above the threshold, the workflow is tuned, routed to a cheaper model, reduced in scope or paused.

Build The Cost Stack Behind Each Result

Cost per useful result is only credible if the firm understands the full cost stack. Token pricing is one visible layer, but AI workflows often include hidden layers: document ingestion, vector databases, embedding calls, retrieval queries, orchestration services, monitoring tools, staff review, security review, exception queues, failed runs and support time. A workflow can look cheap on tokens and expensive in operations if every third answer needs manual rescue.

The FinOps Foundation's FinOps for AI guidance is clear that AI brings new usage metrics such as cost-per-token, volatile costs and GPU scarcity, while requiring regular review of AI costs and usage, quotas, tagging, GPU optimisation and alignment of real-time financial monitoring to business outcomes. It also notes that non-traditional groups such as product, marketing, sales and leadership now contribute directly to AI-driven expenses. That is the governance challenge for UK firms: AI spend is no longer only an IT line. It is operational spend created wherever staff can launch, connect or prompt a tool.

Vendor pricing shows why the full stack matters. Google's Gemini API pricing lists paid input and output prices per million tokens, includes context caching and offers Batch API pricing with a 50 percent cost reduction. Azure AI Foundry documentation says Azure model availability varies by region and cloud and that models come with different capabilities and price points. Those are not just procurement details. They change whether a workflow should run live, batch, cached, routed to a smaller model or restricted to premium cases.

Start with a cost ledger for each workflow. Include model input, model output, embeddings, retrieval, reranking, storage, observability, tool calls, human review minutes, failed attempts, escalation handling and vendor platform fees. Then divide total monthly cost by accepted useful results, not total attempts. If 10,000 calls produced 6,000 accepted outputs, your denominator is 6,000. If 20 percent of outputs needed heavy rework, add the rework cost. This is where the conversation becomes honest. A model route that looks expensive per token may be cheaper per accepted result if it reduces review time and failure rates. Equally, a cheap model may be expensive if it floods staff with poor drafts.

Measure Quality And Risk As Budget Variables

The common misconception is that outcome-based budgeting is only about financial efficiency. It is not. Quality and risk change the cost of a result. A support answer that is quick but wrong is not a useful result. A recruitment summary that exposes sensitive personal data is not a useful result. A board report that looks polished but cannot cite its sources may increase risk rather than reduce work. The budget metric must exclude outputs that fail quality, compliance or approval rules.

The ICO's artificial intelligence guidance hub points organisations to detailed guidance on applying UK GDPR principles to AI systems and an AI and data protection risk toolkit for assessing risks to individual rights and freedoms. For UK firms, that means AI budgets cannot be separated from data governance. If a workflow processes personal data, the result is only useful if it is lawful, fair, secure, proportionate, explainable where required and auditable enough for the organisation's risk profile.

Quality gates should be budget gates. For each workflow, define what counts as accepted: grounded in approved sources, within policy, correctly formatted, reviewed where required, no personal data leakage, no prohibited advice, no missed escalation, no hallucinated facts and no unacceptable tone. Then calculate cost per accepted result after those gates. This stops teams celebrating cheap volume. If an AI sales assistant produces 1,000 account briefs but only 300 are accepted by sales managers, the cost per useful result is based on 300. If an invoice checker flags too many false positives, staff review time belongs in the cost stack.

There is a practical UK procurement angle too. Even outside the public sector, boards and finance teams increasingly want evidence trails. Who approved the use case? Which data can enter the system? Which model route was used? How were outputs tested? What human review applied? Which metrics show value? A firm does not need an enterprise governance office to answer those questions. It needs a simple operating record. Pair each monthly budget report with a quality report: useful result count, accepted rate, failure categories, average review time, critical incidents, data concerns and model changes. The budget then reflects both money and trust.

Use Routing And Quotas To Control Unit Cost

Once useful results and cost stacks are visible, the firm can control spend with routing and quotas. Without them, every workflow tends to drift towards the easiest default: the model already integrated in the tool, the most capable model in the demo, or the route chosen by the first engineer. That default is rarely the best budget choice. Different work deserves different cost envelopes.

A practical routing policy splits AI work by value, risk and latency. Premium reasoning models can be reserved for complex analysis, high-value sales strategy, legal review support or difficult exception handling. Mid-tier models can handle drafting, classification, summarisation and routine staff assistance. Small language models or rules can handle extraction, tagging, simple routing and templated responses. Batch routes can process non-urgent work overnight. Cached context can reduce repeated prompt cost. Retrieval can narrow the context window. Human review can be applied where the business value or risk justifies it.

This is where outcome-based budgeting connects to model routing cost control and AI ROI before committing. The route should be chosen because it delivers the target result at the best acceptable cost, not because it looks neat architecturally. A customer-facing complaint workflow may justify a higher model cost if it lowers escalation errors. A back-office tagging workflow may need a cheaper model even if the output is less elegant. An executive briefing workflow may need a premium route only for the final synthesis, not for every document extraction step.

Quotas make this enforceable. Set monthly budgets by workflow and by route. Give each team a useful result target, a maximum cost per result, a premium model allowance and an exception approval route. Use tags such as department, workflow, customer segment, environment, data class, model route and owner. The FinOps Foundation's guidance on tagging, quotas and business-outcome alignment is especially relevant here because AI expenses are now created by teams outside IT. A quota should not merely say you have spent 80 percent of tokens. It should say whether cost per useful result is within target, whether quality is holding and whether the premium route is being used for the work it was approved to serve.

Report AI Budgets In A Language The Board Can Use

The board does not need a token dashboard as its primary view. It needs a short, reliable account of what the firm is buying, what changed, what risk is attached and what decision is required. Token dashboards are useful for FinOps practitioners and technical owners. They are not enough for leaders deciding whether to scale, pause, renegotiate, replace a vendor or change the workflow.

A good monthly AI budget report has five lines per workflow. First, useful results delivered: accepted support resolutions, approved summaries, qualified leads, reviewed invoices, compliant case notes or hours released. Second, cost per useful result: all-in cost divided by accepted outputs. Third, quality and risk: accepted rate, critical failures, review time, data concerns and escalation misses. Fourth, routing: which model tiers were used, where premium models were consumed and whether cheaper routes were tested. Fifth, action: keep, tune, expand, cap, pause or retire. That is a board conversation.

The counterargument is that this is slower than simply buying licences and watching adoption. It can be slower at the start. It is also cheaper than discovering six months later that usage was high, staff liked the tool, invoices grew, but no one can prove the business result. Adoption alone is not value. Staff may use AI because it is convenient, interesting or bundled into Microsoft 365, Google Workspace, Salesforce, ServiceNow, HubSpot, Notion or a vertical SaaS platform. That does not mean the workflow is improving margin, risk, speed or customer experience.

What this means in practice is to approve AI spend like an operating portfolio. Green workflows have proven useful result economics and can scale within guardrails. Amber workflows show promise but need tuning, better data, better routing or clearer review rules. Red workflows have poor cost per accepted result, weak ownership, excessive rework or unresolved compliance concerns. This language lets finance, operations, IT, data protection and department heads work from the same evidence. It also protects good AI projects from blanket cuts. When budgets tighten, the workflows with documented cost per useful result survive because they can explain what they do in business terms.

Frequently Asked Questions

What is an outcome-based AI budget?

It is an AI budget that measures spend against useful business results rather than raw consumption. Examples include cost per resolved support enquiry, cost per reviewed invoice, cost per accepted sales brief, cost per compliant case note or cost per staff hour released.

Why is token spend not enough?

Token spend shows how much model capacity was consumed, but it does not show whether the work was useful, accurate, accepted, compliant or cheaper than the previous process. It is a meter, not a value metric.

How should a UK firm calculate cost per useful result?

Add the full monthly workflow cost, including model calls, embeddings, retrieval, orchestration, monitoring, platform fees, review time, failed runs and rework. Divide it by accepted outputs that passed the workflow's quality and risk gates.

Which AI workflows suit outcome-based budgeting?

Customer support, invoice checking, CRM enrichment, proposal drafting, policy search, compliance review support, sales research, document triage and operational reporting are good candidates because useful outputs can be counted and reviewed.

Does this mean always using the cheapest model?

No. The cheapest model per token may be more expensive per accepted result if it creates errors, rework or missed escalations. Outcome-based budgeting chooses the route that delivers the useful result at the best acceptable cost and risk level.

How does UK GDPR affect AI budget measurement?

If personal data is involved, outputs should only count as useful when they respect data protection controls such as lawfulness, fairness, security, minimisation, transparency and appropriate human review. Failed or risky outputs belong in the cost stack, not the value count.

What should finance teams ask for each AI project?

Ask for the useful result definition, expected monthly volume, target cost per accepted result, quality gates, model routing policy, owner, data classes, review rules and stop condition if the economics do not work.

How often should AI cost per useful result be reviewed?

Review active workflows monthly and after major model, pricing, prompt, retrieval, data or process changes. Fast-moving or customer-facing workflows may need weekly monitoring until the economics are stable.