AI Productivity in the UK: Why Token Usage Is Not the Same as AI ROI

Discover how UK enterprises can measure AI productivity, optimise AI costs, and move beyond token usage to measurable business outcomes and ROI.

AI adoption is expanding across UK businesses, particularly among larger organisations.

The question for enterprise leaders is no longer simply whether their organisation should use AI. It is whether increasing AI adoption is actually producing measurable business value.

That distinction is becoming more important as companies move from individual AI assistants and generative-AI experiments to enterprise-wide automation, AI agents, and AI-enabled business processes.

According to the Office for National Statistics, around 29% of UK businesses reported using at least one AI technology in June 2026, rising to 49% among businesses with 250 or more employees. The ONS also found that improving business operations is the most common reported purpose for AI adoption.

At the same time, UK government research shows a more complicated picture. Among businesses already using AI, 75% reported improved workforce productivity and 57% reported new or improved processes or operations, but 77% had not yet seen a change in revenue. The same research identified high costs, unclear regulation, data complexity, limited skills, and integration challenges as important barriers to scaling AI.

This creates a new challenge for enterprise leaders:

How do you scale AI without scaling AI cost faster than AI value?

A recently emerging term, ‘tokenmaxxing’, describes a particularly visible version of this problem.

Tokenmaxxing describes the idea of treating high AI-token consumption as a proxy for productivity. But tokens measure model usage, not business outcomes.

For UK enterprises, a more useful way to frame this shift is:

AI Consumption → AI Productivity → Business Outcomes → Financial ROI

This article explains why that shift matters, which AI metrics UK businesses should track, how the economics change with AI agents, and how SAP-centric enterprises can connect AI usage to measurable business value.

What Is Tokenmaxxing?

Tokenmaxxing is an informal, emerging term used to describe the tendency to treat high AI-token consumption as a proxy for productivity. In practice, it reflects a broader tendency in enterprise AI: when usage is easy to measure, organisations can start using usage itself as evidence that an AI initiative is successful.

The idea is simple:

More AI usage = More productivity

The problem is that this assumption confuses activity with outcomes.

Tokens tell an organisation how much model processing has taken place. They do not tell leadership whether:

• a useful task was completed;
• the result was accurate;
• an employee actually saved time;
• a process became cheaper;
• customer service improved;
• inventory was reduced;
• revenue increased;
• errors fell.

The distinction is especially important for enterprises because AI is increasingly embedded in business workflows rather than used purely as a conversational tool.

An AI agent may call several models, retrieve enterprise data, invoke applications, perform multiple actions, and request human approval before completing one business process.

The business does not care primarily how many tokens were consumed.

It cares whether the process was completed faster, better, cheaper, or at greater scale.

Why Token Consumption Is a Poor Enterprise Productivity Metric

Token consumption is attractive because it is easy to measure.

It can be tracked in dashboards, compared across teams, and included in AI budgets. But it has the same fundamental weakness as many activity metrics: it can increase without increasing business value.

Imagine two finance teams.

Team A uses AI heavily, generating large numbers of model calls. Employees still manually validate most outputs, correct data, and move information between systems.

Team B uses AI less frequently, but its workflows automatically identify invoice exceptions, extract relevant information, route approvals, and update the appropriate systems.

Team A may consume more tokens.

Team B may create more value.

The difference is that Team A is optimising for AI activity, while Team B is improving a business process.

This is why enterprise AI measurement needs to move beyond usage alone. Organisations should evaluate AI across several levels: adoption, task performance, process impact, business outcomes, and financial value.

Token consumption is useful for understanding AI usage and operational cost, but it should not be treated as a standalone productivity metric.

A business can increase AI usage without making a process faster or cheaper. Conversely, a relatively small amount of AI activity can create significant value when it automates a high-volume, high-cost, or time-sensitive business process.

A more meaningful measure of AI productivity is therefore not how much AI is being consumed, but how much useful business work is being completed and what changes as a result.

Token consumption helps explain the cost of AI. Process and business metrics explain whether that cost is justified.

The Gap Between AI Productivity and Financial Impact

The UK data makes this distinction particularly relevant.

According to UK government research on AI adoption, many businesses already using AI report improvements in workforce productivity and business processes, while changes in revenue appear less widespread. These findings are based on businesses’ self-reported experiences and should therefore be interpreted as reported outcomes rather than as evidence of direct causation.

This does not mean AI is failing.

It means productivity and financial value are not the same measurement.

An organisation can produce more output with the same workforce. It can shorten a process without increasing revenue. It can reduce employee workload without immediately lowering operating expenditure. It can reduce risk or improve decision quality without generating a visible line item in the income statement.

Office for National Statistics (ONS) data for June 2026 shows that AI adoption is becoming increasingly significant across UK businesses, particularly among larger organisations. This makes the question of how AI value is measured more important as adoption moves from experimentation into core business operations.

For UK enterprises, AI business cases should therefore distinguish between:

Cost Reduction - reducing the cost of completing an existing process.

Capacity Creation - enabling the same workforce to handle more work without a proportional increase in resources.

Revenue Growth - improving conversion, retention, sales capacity, or other revenue-generating outcomes.

Risk Reduction - reducing errors, compliance exposure, operational disruption, or other business risks.

Service Improvement - improving customer experience, response times, service levels, or employee experience.

These are different forms of value and should be measured differently. A successful AI initiative does not necessarily have to reduce headcount or immediately increase revenue. Its value may come from creating additional capacity, improving operational performance, or reducing risk.

The important point is to define which type of value an AI initiative is expected to create before measuring its success.

What UK Enterprises Should Measure Instead

A mature AI productivity framework should look beyond usage and token consumption. The most useful metrics cover four connected areas:

1. Adoption - Are employees and business processes actually using AI?

2. Operational Productivity - Is AI helping teams complete work faster or at greater scale?

3. Quality - Are productivity gains being achieved without increasing errors, rework, or human-review effort?

4. Business Outcomes - Is AI improving measurable business performance and financial results?

These four areas help organisations distinguish between AI activity and AI value.

1. AI Adoption

This measures whether people are actually using the capability.

Examples include:

• active users;
• AI-enabled workflows;
• percentage of eligible employees using AI;
• frequency of AI-assisted tasks;
• number of business processes using AI.

These metrics are important during adoption, but they are not enough to demonstrate ROI.

2. Operational Productivity

This measures how the process itself changes.

Examples include:

• cycle time;
• cost per transaction;
• cases processed per employee;
• manual steps removed;
• automation rate;
• average handling time;
• throughput.

This is where AI begins to become an operational investment rather than simply an IT experiment.

3. Quality

Speed without quality can create hidden costs.

Track:

• error rate;
• rework;
• exception rate;
• human-review rate;
• defect rate;
• escalation rate;
• customer-impacting mistakes.

An AI solution that reduces processing time by 50% but doubles rework may not have created meaningful value.

4. Business Outcomes

This is where executive-level ROI begins.

Depending on the process, metrics may include:

• cost per completed process;
• working capital;
• inventory;
• order fulfilment;
• customer retention;
• margin;
• revenue;
• forecast accuracy;
• procurement savings;
• service-level performance.

The key is to connect AI-enabled improvements to KPIs that already matter to the business.

For example, a finance team may care less about how many invoices AI has processed than about cost per invoice, touchless-processing rate, exception volume, and invoice-cycle time.

Similarly, a supply chain team may care less about the number of forecasts generated than about forecast accuracy, inventory levels, stock-outs, and planner workload.

The objective is therefore to move from measuring AI usage to measuring useful work, process improvement, and business results.

The Right AI KPI Is Usually the Business Process

One of the biggest mistakes in enterprise AI is measuring the technology instead of the process.

AI is rarely deployed for its own sake. A business does not invest in a model because it wants more model calls, more generated content, or more AI interactions. It invests because it expects something in the business to work better.

That “something” is usually a process.

Invoices should be processed faster. Customer requests should be resolved with fewer escalations. Buyers should spend less time on routine tasks. Planners should respond to exceptions earlier. Sales teams should spend more time with customers and less time preparing information.

This makes the business process the most useful unit for measuring AI value.

It also changes how an organisation builds an AI business case. Instead of starting with:

“What can this AI model do?”

the better starting point is:

“Which business process are we trying to improve, what does it cost today, and which KPI should change?”

This approach creates a direct connection between technology investment and operational performance. It also makes it easier to establish a baseline, measure improvement, identify hidden costs such as human validation and rework, and calculate the financial impact.

For enterprise AI, that connection is especially important because many workflows cross multiple applications, teams, and approval stages. An AI capability may perform extremely well at one step while creating additional work elsewhere in the process.

A useful AI business case should therefore consider the end-to-end process, not just the step where the AI is introduced.

Take invoice processing.

A weak AI KPI is:

Number of invoices analysed by AI

A stronger KPI is:

Percentage of invoices processed without manual intervention

An even stronger business KPI is:

Total cost per invoice + processing time + exception rate

The same principle applies across the enterprise.

Procurement

Instead of measuring:

Number of AI supplier interactions

measure:

Procurement cycle time + compliance + negotiated savings + buyer capacity

Sales

Instead of:

Number of AI-generated recommendations

measure:

Conversion rate + sales-cycle time + seller administration time

Supply Chain

Instead of:

Number of AI forecasts generated

measure:

Forecast accuracy + inventory + stock-outs + planner effort

Finance

Instead of:

Number of AI-generated financial insights

measure:

Close-cycle time + manual effort + exception volume + reporting accuracy

Customer Service

Instead of:

Number of AI responses

measure:

First-contact resolution + handling time + escalations + customer satisfaction

In many enterprise scenarios, the most useful AI metric is connected to a business process owner. That gives the AI initiative a clear baseline, a measurable target, and someone accountable for translating technical performance into business value.

Practical recommendation: Start every AI business case with one clearly defined business process and its existing KPIs. Establish the baseline before introducing AI, then measure the change in cycle time, cost, quality, human effort, and business outcomes. Keep AI usage and token consumption as supporting operational metrics rather than the primary definition of success.

AI ROI Should Include the Full Cost of the Workflow

Enterprise AI cost is more than model pricing.

The cost of an AI initiative is determined by the entire workflow required to produce a usable business outcome. A model may be inexpensive to run, but the surrounding technology, data, governance, and human effort can become a much larger part of the total investment.

A realistic AI total-cost calculation can therefore include:

• Model and inference cost - token consumption, API calls, model routing, and usage volume;
• AI platform and infrastructure - hosting, compute, storage, orchestration, and supporting services;
• Integration - connections with SAP and non-SAP applications, APIs, workflows, and enterprise systems;
• Data preparation - data cleansing, enrichment, retrieval, classification, and ongoing data management;
• Security and governance - access controls, privacy, monitoring, evaluation, compliance, and audit requirements;
• Monitoring and operations - observability, performance monitoring, incident management, and AI FinOps;
• Human validation - review, approval, exception handling, and quality control;
• Change management - employee training, process redesign, adoption, and organisational change;
• Maintenance - model updates, prompt and workflow optimisation, integrations, testing, and ongoing support.

This matters because AI can shift costs instead of eliminating them.

For example, an AI-enabled finance workflow may generate a recommendation in seconds, but the process may still require an employee to validate the output, resolve an exception, obtain approval, and update the relevant enterprise system.

The model itself may represent only a small part of the total cost. Human effort, integration, governance, and downstream process work may account for much more.

That is why an AI business case should not stop at:

How much does each AI request cost?

It should ask:

How much does it cost to complete the business process successfully?

This distinction is particularly important for enterprise workflows that span multiple systems or involve human approval. The relevant metric is therefore not simply cost per AI request, but total cost per successful business outcome.

For UK enterprises, this approach also makes AI investments easier to compare with existing operational costs. Instead of treating AI as a separate technology expense, organisations can assess whether the AI-enabled process is actually cheaper, faster, more accurate, or more scalable than the process it replaces or augments.

Practical recommendation: Build the AI business case around the end-to-end workflow. Measure model cost alongside integration, human validation, governance, rework, and maintenance, then compare the total cost with the value of the completed business process.

The Hidden Cost of AI Validation

Human oversight is particularly important in enterprise environments, where AI outputs can affect financial transactions, customer interactions, operational decisions, and other business-critical processes.

According to UK government research on AI adoption, 84% of businesses using AI reported at least some human input or checking of AI outputs or decisions, while 67% reported significant human input or checking.

This is not necessarily a weakness.

For many enterprise processes, human review is an appropriate control. A finance team may need to approve a payment-related decision. A procurement specialist may need to review a supplier recommendation. A planner may need to validate an AI-generated forecast before changing inventory parameters.

The issue is not whether a human is involved.

The issue is whether human involvement is included in the economics of the AI-enabled process.

Consider a simple example.

Traditional process: 30 minutes of manual work

AI-enabled process: 5 minutes of AI processing + 10 minutes of human review + 5 minutes of correction

At first glance, the AI solution appears to save 25 minutes because the AI completes its part of the task in five minutes.

The end-to-end process saves 10 minutes.

That difference matters when the workflow runs thousands or millions of times per year.

For example, saving 10 minutes across 100,000 transactions represents more than 16,000 hours of annual capacity. The same calculation also makes clear why a small amount of human review can materially change the economics of an AI use case.

Human validation should therefore be measured alongside:

• review time;
• approval time;
• exception handling;
• correction and rework;
• escalation rates;
• percentage of outputs requiring intervention.

A useful calculation is:

Net Time Saved = Previous Process Time − AI Execution − Human Review − Rework

The same principle can be applied to cost:

Net Process Cost = AI Cost + Human Effort + Infrastructure + Rework + Governance

This provides a much more realistic baseline for evaluating enterprise AI productivity.

The goal is not necessarily to remove humans from the process. In many cases, the higher-value objective is to move people from routine processing to exception handling, judgement, and higher-value decisions.

Practical recommendation: Do not model an AI use case on automated execution time alone. Measure the complete human-in-the-loop workflow and define in advance which decisions should remain subject to human review. This will produce a more credible ROI estimate and a more sustainable operating model.

AI Agents Change the Economics

AI agents introduce another layer of complexity because they can turn a single business request into a sequence of model interactions, tool calls, system actions, and human approvals.

Traditional AI assistance might involve a single interaction: a user asks a question and receives an answer.

An agentic workflow can be much more involved. An agent may:

1. interpret an instruction;
2. retrieve enterprise data;
3. reason about the task;
4. call one or more business systems;
5. validate the information;
6. invoke another tool or workflow;
7. update a record;
8. identify an exception;
9. request human approval;
10. complete or escalate the process.

As the workflow becomes more autonomous, the number of model interactions can increase significantly.

That is not inherently inefficient.

The value of an AI agent should be judged by what the complete workflow achieves, not by how few model calls it makes. A workflow that uses several model interactions but eliminates hours of manual work can be economically attractive. Conversely, a single low-cost model call has limited value if employees still have to perform most of the process manually.

This changes the unit of measurement.

For a traditional AI assistant, the organisation may focus on the cost of an interaction.

For an AI agent, the more meaningful question is:

What does it cost to complete the business process successfully?

This becomes especially important when agents interact with enterprise systems such as SAP. An agent supporting procurement, finance, supply chain, or customer service may access several data sources and applications before completing a task. Model consumption is only one component of that workflow's economics.

For every AI agent, enterprises should therefore be able to answer:

What business process does the agent execute?

How often does the workflow run?

How many model calls does each completed workflow require?

Which models are being used, and why?

Which enterprise systems and tools does the agent access?

How often is human intervention required?

How often does the workflow fail or require rework?

What is the total cost of one successfully completed workflow?

What measurable business outcome does that workflow create?

This is the foundation of sustainable AI FinOps.

It also introduces a broader management principle: enterprises should optimise AI agents for cost per successful outcome, rather than simply cost per token or cost per model call.

The objective is not to make every agent as inexpensive as possible. It is to ensure that the intelligence, automation, infrastructure, and human oversight involved in the workflow generate enough business value to justify the total cost.

AI FinOps Is Becoming a Business Discipline

AI FinOps is often discussed in terms of cloud consumption, model pricing, and controlling AI infrastructure spend.

For enterprises, that definition is too narrow.

As AI becomes embedded in business applications and autonomous workflows, finance and technology leaders need to understand not only how much AI costs, but also what the organisation is getting for that cost.

The goal should be to connect:

AI Consumption + Infrastructure + Human Effort + Business Outcome

That means monitoring AI economics across several dimensions:

• model;
• application;
• user group;
• workflow;
• business unit;
• business process.

This level of visibility can reveal an important distinction between expensive AI and valuable AI.

A high-cost workflow may still generate an attractive return if it automates a large amount of manual work or improves a high-value business process. A low-cost workflow may deliver little value if employees continue to perform most of the work themselves.

For example, instead of reporting:

£X spent on LLM inference this month

an enterprise should increasingly be able to report:

£X spent processing Y invoices

£X spent resolving Z service cases

£X spent supporting N planning decisions

£X spent automating Y procurement workflows

The next step is to connect those figures to business performance:

£X cost → Y transactions → Z hours saved → N% fewer exceptions → £Z business value

This changes the conversation with CFOs, CIOs, and business leaders.

AI becomes measurable in the same language as other operating investments: unit cost, productivity, capacity, quality, risk, and return.

It also makes AI optimisation more sophisticated. Instead of simply trying to reduce token consumption, teams can identify where to:

• route tasks to more economical models;
• reduce unnecessary model calls;
• redesign inefficient workflows;
• automate repetitive validation;
• improve prompts and context management;
• remove unnecessary human intervention;
• focus investment on the use cases with the strongest business economics.

In other words, the objective of AI FinOps is not minimum AI spend.

It is optimal AI economics: achieving the required quality and level of automation at a cost that makes the underlying business process economically attractive.

Practical recommendation: Measure AI costs at the level of the business workflow, not only at the infrastructure or model level. Give finance and business owners visibility into cost per completed process, human effort, quality, and resulting business value so AI spending can be managed as an operating investment rather than simply an IT expense.

Model Selection Is Part of the ROI Strategy

Using the most capable model for every task is rarely the most economical strategy.

Enterprise AI workloads are not all equally complex. Some tasks require advanced reasoning and broad context. Others involve relatively straightforward activities such as extraction, classification, summarisation, matching, translation, or structured transformation.

Treating all of these workloads in the same way can create unnecessary cost and latency without delivering additional business value.

A mature AI architecture should therefore select the appropriate approach for each task based on:

Complexity + Accuracy Requirement + Risk + Cost

This means model selection should be considered as part of the AI business case, not as a purely technical decision.

A typical decision logic might look like this:

For example, using a highly capable reasoning model to classify routine invoice fields may add little value. Likewise, relying on a lightweight model for a complex financial or supply-chain decision may create more review effort and operational risk than it saves.

This creates an important principle for enterprise AI:

The most powerful model is not necessarily the most appropriate one. The better choice is the model or combination of technologies that reliably delivers the required business outcome at an appropriate cost.

Model selection also affects the wider architecture. Organisations need to consider routing, fallback models, prompt and context management, monitoring, data access, latency, security, and human-approval mechanisms.

The economics therefore become:

Task Requirements → Architecture Choice → Model Selection → Cost + Quality → Business Outcome

And sometimes the highest-ROI solution is not an LLM at all.

A deterministic calculation may be better handled by traditional software. A workflow may be more efficiently automated with business rules. A high-risk decision may require AI support combined with human judgement.

The objective is not to maximise the amount of AI used.

It is to use the right combination of AI, automation, rules, and human expertise for each part of the process.

UK Enterprises Also Need to Account for Data and Governance

AI economics do not exist separately from enterprise governance.

An AI use case can look inexpensive during a pilot and become significantly more complex when it moves into production. Once an AI workflow starts accessing sensitive business data, interacting with enterprise systems, making recommendations, or taking actions automatically, the organisation also needs to account for security, privacy, oversight, monitoring, and operational controls.

This is particularly relevant in the UK, where businesses continue to identify cost, skills, data complexity, integration, and regulatory considerations as barriers to wider AI adoption. According to UK government research on AI adoption, larger organisations are more likely to report concerns around the cost and complexity associated with scaling AI.

For enterprise leaders, governance should therefore be viewed as part of the AI operating cost and risk model, not as a separate compliance exercise.

A sustainable AI operating model should address:

• Data access — which data the AI can access and under what conditions;
• Privacy - how personal and commercially sensitive information is protected;
• Security - authentication, authorisation, encryption, and secure integration;
• Model governance - which models are approved and how their use is controlled;
• Human oversight - where human review or approval remains mandatory;
• Auditability - how decisions, inputs, outputs, and actions can be traced;
• Model evaluation - how accuracy, reliability, bias, and performance are assessed;
• Monitoring - how cost, quality, failures, and unusual behaviour are detected;
• Vendor management - how third-party models, platforms, and services are assessed;
• Change control - how model, prompt, workflow, and system changes are tested and governed.

The UK regulatory environment also continues to evolve. Rather than relying on a single AI-specific compliance checklist, enterprises need to consider the regulations and sector requirements that already apply to the underlying activity, particularly where AI processes personal data or influences decisions.

This is why governance can have a direct impact on ROI.

A poorly governed AI workflow may require more manual review, additional controls, repeated testing, or expensive remediation before it can be deployed at scale. By contrast, governance designed into the architecture from the beginning can make it easier to move successful use cases from pilot to production.

The objective is not to slow AI down.

It is to make AI safe, measurable, and economically sustainable enough to scale.

Why SAP Customers Have a Different AI Opportunity

For an SAP-centric enterprise, AI should not be treated as a disconnected layer sitting above the ERP.

The strongest opportunity is often to embed intelligence directly into business processes.

SAP's current Business AI strategy increasingly centres on AI assistants and agents that are grounded in business data and process context and can execute complex workflows across SAP and non-SAP environments.

That creates a different AI ROI model.

Instead of asking:

How many people are using an AI assistant?

an SAP customer can ask:

How much faster is the process?

How many exceptions are automatically resolved?

How many transactions require human intervention?

How much working capital is affected?

How many planning hours are removed?

How much manual effort has been eliminated?

This is particularly relevant in:

• finance;
• procurement;
• supply chain;
• manufacturing;
• sales;
• customer service;
• asset management;
• HR.

AI ROI in Finance

Finance is one of the clearest areas for outcome-based AI measurement.

Potential use cases include:

• invoice processing;
• accounts receivable;
• dispute management;
• financial close support;
• reconciliation;
• management reporting;
• anomaly detection.

Instead of measuring the number of documents processed by AI, measure:

The technology becomes valuable because it changes the economics of finance operations.

AI ROI in Procurement

Procurement teams can use AI to support:

• supplier discovery;
• sourcing analysis;
• contract review;
• spend analysis;
• supplier communications;
• purchase-order exception management.

Useful KPIs include:

Procurement cycle time

Buyer hours per transaction

Compliance rate

Supplier response time

Negotiated savings

Contract-processing time

This is much stronger than reporting how many AI recommendations were generated.

AI ROI in Supply Chain

Supply chain is particularly suitable for outcome-based AI because operational metrics are already relatively concrete.

Possible AI use cases include:

• demand forecasting;
• inventory optimisation;
• exception management;
• order prioritisation;
• supplier-risk monitoring;
• logistics planning.

The business case can connect AI to:

Inventory

Service levels

Stock-outs

Expediting

Planner workload

Forecast accuracy

An AI model that generates excellent forecasts but does not change planning decisions may have limited value.

The real value appears when the forecasting improvement changes inventory, service, or planning outcomes.

AI ROI in Manufacturing

Manufacturers can connect AI to:

• production planning;
• predictive maintenance;
• quality management;
• engineering support;
• root-cause analysis;
• operator assistance.

Useful metrics include:

• downtime;
• OEE;
• scrap;
• first-pass yield;
• maintenance cost;
• planning effort;
• production throughput.

Again, the correct question is not how much AI was used.

It is:

What changed in the manufacturing process because AI was introduced?

AI + SAP BTP: Measure the Full Solution

Many enterprise AI use cases will require orchestration across SAP and non-SAP systems.

SAP BTP can provide the foundation for applications, integration, automation, data, and AI capabilities around the SAP core.

That introduces another important principle:

AI ROI should be measured across the whole solution architecture, not just the model.

An AI-enabled workflow may involve SAP S/4HANA, SAP BTP, integrations, an AI model or agent, external applications, and human approval. Every layer can affect the final economics.

The model may be inexpensive while integration, data preparation, monitoring, or human validation drive much of the total cost. Conversely, a more expensive solution may deliver stronger ROI if it removes significant manual effort or scales a high-value process.

For SAP customers, the right question is therefore not:

“How cheaply can we run the AI model?”

It is:

“What is the most economical architecture for delivering the required business outcome?”

Practical recommendation: Evaluate AI, BTP, integration, automation, and human effort as one connected investment, and compare the total cost of the workflow with its measurable business value.

From Pilot ROI to Enterprise ROI

Many AI pilots look economical because they operate at a small scale.

The real test begins when the use case moves into production and transaction volumes increase.

A pilot may process:

1,000 transactions per month

An enterprise deployment may process:

1,000,000 transactions per month

At that point, the economics can change. Model usage, infrastructure, integration, monitoring, support, governance, exception handling, concurrency, and data volumes all increase. Human validation may also become a significant operating cost.

A workflow that looks attractive at pilot scale is therefore not automatically economical at enterprise scale.

The business case should be tested at three levels:

Pilot Economics

Does the use case work technically and deliver measurable value?

Production Economics

Does it remain reliable, accurate, and cost-effective at realistic business volumes?

Enterprise Economics

Does the unit economics remain attractive when the solution is scaled across users, processes, locations, and transaction volumes?

This last step is critical. The goal is not simply to prove that AI can create value in a controlled pilot, but to prove that the value remains sustainable as adoption grows.

Practical recommendation: Before scaling an AI pilot, model its unit economics at expected production and enterprise volumes. Include model costs, infrastructure, integration, human validation, support, governance, and exception handling, then compare the projected total cost with the value created per completed business process.

A Practical AI ROI Formula for UK Enterprises

A useful starting point is:

AI ROI = (Annual Business Value − Annual AI Operating Cost) / Total AI Investment

Total AI Investment should include the implementation costs required to put the solution into production, such as integration, data preparation, testing, deployment, and change management.

The business value itself can come from several sources.

Cost Savings

Direct reductions in operating cost, such as:

• reduced manual processing;
• lower support effort;
• reduced external-service costs;
• less rework and exception handling.

Capacity Value

Additional productive capacity without proportional increases in resources, such as:

• more transactions processed;
• greater employee capacity;
• faster response times;
• higher throughput without equivalent headcount growth.

Revenue Value

Additional or protected revenue generated through:

• higher conversion;
• shorter sales cycles;
• improved customer retention;
• better service and response.

Risk Value

Financial or operational value created by reducing:

• errors;
• compliance exposure;
• fraud or anomalies;
• operational disruption.

For some UK enterprises, capacity and risk value can be as important as direct cost savings. An AI initiative does not necessarily need to eliminate jobs or produce immediate revenue to justify investment. It may instead allow the business to handle growth without proportional resource increases, improve service levels, or reduce operational risk.

The exact calculation will vary by use case, but the principle remains the same: establish the baseline before implementation and measure the change against it after deployment.

A credible business case should therefore define:

Current Process Cost → AI-enabled Process Cost → Net Business Value → Total Investment → ROI

Without a reliable baseline, AI ROI quickly becomes an estimate rather than a measurable business result.

A 90-Day Framework for Proving AI Value

UK organisations do not need to wait for a multi-year transformation programme before measuring AI ROI.

A practical 90-day approach can be structured as follows.

Days 1–30: Establish the Baseline

Choose one business process.

Measure:

• volume;
• cycle time;
• labour effort;
• error rate;
• cost;
• exceptions;
• current system dependencies.

Do not introduce the AI solution before understanding the baseline.

Days 31–60: Introduce and Instrument AI

Deploy the AI capability with:

• usage monitoring;
• quality measurement;
• human-review tracking;
• model-cost tracking;
• workflow monitoring.

The objective is to understand the entire process, not just model performance.

Days 61–90: Prove Economic Value

Compare:

Before AI

versus

After AI

Measure:

• net time saved;
• automation rate;
• quality;
• cost per completed task;
• human intervention;
• business KPI;
• estimated annual value.

At the end of the exercise, leadership should be able to answer:

Should we scale this use case, redesign it, or stop it?

That is a much stronger decision than simply asking whether employees like the AI tool.

The Enterprise AI Scorecard

A practical executive scorecard can combine technology, operational, and financial metrics.

This creates a much more balanced view of AI performance.

What UK Boards and CFOs Should Ask About AI

AI investment is increasingly becoming a board-level issue. As UK businesses move from experimentation to broader deployment, leadership needs to evaluate AI with the same commercial discipline applied to other major technology and transformation investments.

The questions should therefore move beyond:

How many employees are using AI?

A CFO, CIO, or board should ask:

What business process are we improving?

What is our cost per successful AI-enabled transaction?

Which AI use cases have a quantified business case and a clearly accountable owner?

What measurable value has each production use case delivered?

What percentage of AI pilots have reached production, and why did the others stop?

Which initiatives have demonstrated cost savings, additional capacity, revenue impact, or risk reduction?

How much human validation and rework does each AI workflow still require?

How do the unit economics change as usage and transaction volumes scale?

Are we using the most appropriate model for each workload, or simply the most capable one?

What data, security, governance, and human-oversight controls are in place?

Which AI investments should we scale, redesign, or stop?

These questions shift the discussion from AI adoption to AI portfolio management.

This is particularly important because an AI initiative can perform well technically while remaining commercially unattractive. A successful pilot may not retain its economics at enterprise scale, and a highly adopted tool may still have limited measurable impact on the underlying business.

The final question is therefore particularly important.

An effective AI strategy is not only about identifying where to invest.

It is also about knowing where to invest more, where to redesign, and where to stop.

Practical recommendation: Review AI initiatives as a portfolio, not as isolated technology projects. Give each use case a business owner, baseline, target KPI, total-cost model, and clear scale-or-stop criteria. This allows the board and finance team to direct investment toward AI initiatives that demonstrate sustainable business value.

What Tokenmaxxing Gets Wrong

The problem with tokenmaxxing is not that high AI usage is inherently bad.

AI can deliver significant productivity gains, and complex use cases may legitimately require substantial model processing. The problem starts when usage becomes the measure of value.

The assumption is simple:

More AI usage = More value

Enterprise economics are rarely that simple.

A better way to evaluate an AI initiative is to look at what happens after the model is used:

AI Usage → Work Completed → Process Improvement → Business KPI → Financial Value

Each stage matters.

If AI generates more content but increases review and editing time, the productivity gain may be limited.

If AI generates more software code but creates additional testing and technical-debt costs, the overall value may be negative.

If AI processes more invoices but employees still handle most exceptions manually, the reduction in cost per invoice may be much smaller than expected.

On the other hand, an AI agent may consume significantly more tokens because it retrieves data, reasons through several steps, interacts with multiple enterprise systems, and completes a workflow that previously required several employees.

That can still be a strong business case.

The relevant question is therefore not how much AI is being consumed, but what the organisation gets for that consumption.

This is why enterprise AI optimisation should consider three variables together:

Cost + Quality + Business Outcome

A cheaper model is not necessarily better if it produces lower-quality results and increases human review. A more expensive model may be more economical if it reliably completes a high-value process with minimal intervention.

The same principle applies to AI agents, copilots, automation platforms, and SAP-enabled AI workflows.

Practical recommendation: Do not optimise for the cheapest AI request or the highest AI usage. Optimise for the best business outcome at an economically sustainable cost, taking into account model consumption, human effort, quality, process performance, and the value created.

How LeverX Helps UK Enterprises Turn AI Into Business Value

LeverX helps UK organisations move from AI experimentation to scalable, measurable enterprise AI.

Our AI consulting services cover AI strategy, use-case discovery, readiness assessment, solution design, implementation, integration, governance, deployment, and optimisation.

For SAP-centric organisations, LeverX combines SAP, AI, data, integration, and engineering expertise to connect AI with the business processes where measurable value is created.

We help organisations:

• identify and prioritise high-value AI use cases;
• establish process baselines and ROI targets;
• assess data, integration, and technology readiness;
• design AI-enabled business workflows;
• select appropriate models and architectures;
• integrate AI with SAP and non-SAP systems;
• build governance, security, and monitoring into the solution;
• move successful pilots into production;
• measure and optimise business value over time.

Our focus is not on deploying AI for the sake of adoption or maximising model consumption. It is on helping UK enterprises determine where AI can create measurable business value, how much that value costs to deliver, and how successfully the solution can scale.

Conclusion: The Next AI Advantage Is Economic Discipline

The first phase of enterprise AI focused largely on experimentation.

The next phase has increasingly focused on adoption.

The phase ahead is likely to place much greater emphasis on economics and measurable business value.

As UK businesses move from AI pilots to broader deployment, the competitive advantage will increasingly depend not on how much AI they use, but on how effectively they turn AI investment into measurable business value.

The questions are therefore changing.

Not:

How many prompts were sent?

How many tokens did we consume?

How many employees use AI?

But:

How much useful work did AI complete?

How much manual effort did it remove?

Did quality improve or decline?

What is the true cost of the AI-enabled process?

What measurable business outcome changed?

And ultimately:

What value did the organisation create for every pound invested in AI?

This matters particularly for UK enterprises as AI adoption expands while businesses continue to face challenges around cost, skills, data, integration, and governance. The ability to prove value will become increasingly important as AI moves from isolated productivity tools into core operational processes.

The organisations that build lasting AI advantage will not necessarily be those that consume the most AI.

They will be those that know where AI creates value, how much that value costs to deliver, and how to scale it without losing economic discipline.

That means connecting AI to the right processes, the right data, the right operating model, and the right business KPIs.

The future of enterprise AI is unlikely to be defined by maximum consumption.

It will increasingly be defined by how much useful business value organisations can generate from their AI investment.

Frequently Asked Questions

What is tokenmaxxing?

Tokenmaxxing is the practice of treating high AI-token consumption as a proxy for productivity. Token usage measures model consumption, but does not directly show whether the AI completed useful work, improved a process, or generated financial value.

Why is tokenmaxxing a problem for UK businesses?

Because high AI usage can increase model and infrastructure costs without generating proportional improvements in productivity, quality, or revenue. UK businesses already identify cost, skills, data complexity, integration, and governance as important barriers to scaling AI.

Is high AI token usage always bad?

No. Complex reasoning and agentic workflows can legitimately consume significant model resources. The problem is using token consumption as the primary measure of productivity rather than connecting AI usage to completed work and business outcomes.

What should businesses measure instead of token usage?

Businesses should measure process cycle time, automation rate, human-review effort, error rates, cost per completed task, and business KPIs such as revenue, inventory, service levels, procurement savings, or working-capital impact.

How do you calculate AI ROI?

Start with a baseline for the business process, calculate the full cost of the AI-enabled workflow, measure quality and human intervention, and compare the resulting business value with total investment.

What is the most important AI productivity metric?

There is no universal KPI. In most enterprise scenarios, cost and time per successful business outcome are more meaningful than model usage because they connect AI directly to the process being improved.

Why does human validation matter for AI ROI?

Human review can significantly affect the economics of an AI workflow. UK government research found that 84% of businesses using AI reported at least some human input or checking of AI outputs or decisions.

Can AI increase productivity without reducing headcount?

Yes. AI can create value by increasing employee capacity, allowing teams to process more transactions, respond faster, reduce backlogs, or focus on higher-value activities without reducing workforce size.

What is AI FinOps?

AI FinOps is the discipline of managing the financial and operational economics of AI. It includes monitoring model usage, routing workloads, controlling cost, analysing unit economics, and connecting AI spending to business outcomes.

Why are AI agents different from traditional AI assistants?

AI agents can execute multi-step workflows, use tools and enterprise systems, reason over context, and take actions rather than only generate responses. This can increase both business value and AI consumption, making workflow-level ROI measurement more important. SAP positions Joule Agents and Joule Assistants around context-aware, end-to-end workflow execution across business processes.

How does SAP Business AI change enterprise AI economics?

SAP Business AI is increasingly designed to connect AI with enterprise data, business-process context, and applications. This makes it possible to evaluate AI based on business-process outcomes rather than standalone model usage.

What role does SAP BTP play in AI?

SAP BTP can provide the platform layer for integrating applications, data, automation, AI services, and extensions around the SAP core. This is useful when an AI workflow needs to interact with multiple SAP and non-SAP systems.

How should UK businesses choose AI use cases?

Prioritise use cases with a clear business owner, measurable baseline, meaningful operational impact, accessible data, manageable risk, and a credible path to positive ROI.

What is the biggest mistake when measuring AI productivity?

The biggest mistake is measuring AI activity instead of business outcomes. High usage can demonstrate adoption, but it does not prove that the organisation is becoming faster, cheaper, more accurate, or more profitable.

How can UK enterprises move from AI pilots to production?

Start with measurable business processes, establish baselines, validate data and integration requirements, define governance, instrument AI usage and quality, prove unit economics, and scale only use cases that demonstrate sustainable value.

Disclaimer: The information in this article is provided for general informational purposes only and does not constitute financial, legal, regulatory, or professional advice. AI capabilities, model pricing, product functionality, regulatory requirements, and market conditions can change over time. Any AI ROI, productivity, cost, or business-value examples presented are illustrative and should be validated against an organisation’s own data, processes, technology landscape, and operating costs before making investment decisions.