{{item.videoDuration}}
{{item.title}}
{{item.text}}
{{item.videoDuration}}
{{item.text}}
AI cost discipline isn’t simply about using fewer tokens. It’s about whether AI enables you to redesign work and create outcomes that justify the full cost of delivering them. For organisations, that means testing costs and performance upfront, then reinvesting value realised in further innovation.
Is your AI spending creating enough value?
When you and your competitors have access to many of the same AI models, technology alone may not provide a lasting advantage. The difference is likely to come from how effectively you use it.
AI cost discipline is therefore more than an efficiency exercise. It can give your organisation the capacity to transform operations, improve customer and employee experiences, develop new services, and reinvest in further innovation.
The objective isn’t to minimise AI use or choose the cheapest model for every task. It’s to determine where AI can transform work, and whether the value of the resulting outcome justifies its total cost.
AI use is growing, but value is not guaranteed
Many AI services are consumed through tokens, the units of text and data processed by an AI model. Prompts, responses, documents, automated actions, and the contextual information provided to a model can all contribute to token consumption.
As AI use grows across an organisation, so does the associated cost. More sophisticated workflows can also involve several models, agents, and automated steps working together. This makes the cost of completing a task harder to see and predict.
The challenge is not token consumption in isolation. It’s whether the organisation can connect that consumption, and the wider cost of using AI, to a measurable business outcome.
The value challenge is already clear in Ireland. PwC’s 2026 CEO Survey found that 17% of Irish CEOs reported higher revenue from AI in the previous 12 months, while 23% reported lower costs. PwC’s AI Agent Survey found that 53% of Irish respondents had achieved measurable productivity gains from AI agents, but only 38% said these gains had translated into tangible cost savings.
The findings underline an important distinction: adopting AI is not the same as creating value from it. Productivity improvements only produce a financial return when an organisation can capture the time or capacity released, redesign the underlying workflow, and redirect resources towards more valuable activity.
"AI cost discipline isn’t about limiting adoption. It’s about understanding what each AI workflow costs, what outcome it supports, and whether the investment is creating measurable value."
David Lee,Chief Technology Officer, PwC IrelandLook beyond the token bill
Token costs are only one part of the economics of AI. The true cost of an AI-enabled task may also include software licences, infrastructure, data retrieval, orchestration, governance, security, human oversight, and the time employees spend checking or correcting outputs. A cheaper model can become more expensive if it requires repeated attempts, creates additional review work, or produces an outcome that does not meet the required quality or risk standard.
A premium model is not always the answer either. For a relatively straightforward task, the additional capability may add cost without creating additional value.
Organisations therefore need to assess the cost per outcome, rather than concentrating solely on the price of an individual model or the number of tokens consumed. This means comparing the total cost of completing a workflow with the result it produces, such as time saved, revenue generated, service improved, or risk reduced.
This approach can give CFOs, CIOs, and business leaders a shared way to determine whether an AI-enabled process is creating sufficient value.
Understand why costs increase
AI expenditure can be difficult to manage because it behaves differently from a conventional software licence.
Costs can be distributed across a workflow
Token consumption may occur across planning, information retrieval, reasoning, tool use, orchestration, quality checks, and review. If these costs are not attributed to a particular workflow and outcome, leaders may see the overall bill but not what’s driving it.
Consumption can grow as workflows scale
An AI agent may complete several steps before returning an answer. It may retrieve information, use other tools, or delegate work to additional agents. If a result is incomplete or unsuccessful, some or all of those steps may be repeated.
Costs therefore depend not only on how frequently employees use AI, but also on how each workflow has been designed.
Model prices do not provide a complete comparison
Different models offer different combinations of cost, speed, accuracy, and capability. The lowest-priced option may lead to more rework, while the most advanced model may be unnecessary for a routine task.
The aim should be to use the lowest-cost model that can reliably deliver the required outcome, taking account of quality, risk, rework, and human oversight.
Existing processes may not be fit for purpose
Before optimising an AI workflow, leaders should ask whether the underlying process is still necessary and whether it has been designed around the right outcome.
The stronger business case may require more than making an existing task cheaper. AI can justify higher consumption when it enables an organisation to redesign how work is done, improve the outcome, or create a service that wasn’t previously viable. Leaders therefore need to compare the full cost of a redesigned workflow with the value it creates, rather than treating a rising token bill as evidence that spending is out of control.
Build an operating model for AI value
Tools such as cost dashboards, routing engines, and caching layers can support AI cost management. But tools alone won’t determine which outcomes matter, who owns them, or where savings should be reinvested.
A more complete operating model brings together four disciplines.
1. Underwrite costs and value upfront
Assess each use case before it is built or scaled. Estimate its direct and indirect costs, the expected business outcome, and the value of achieving that outcome.
Testing should explicitly examine likely consumption behaviour. It should assess how costs change at different levels of use and compare the cost and performance of suitable models. This allows the organisation to determine whether the additional cost of a premium model is justified by a better business outcome.
The assessment should establish an expected cost per outcome and a benchmark against which actual performance can be monitored. It can then support an informed decision to proceed, redesign the proposed use case, or stop it.
2. Redesign workflows and architect for efficiency
Start with the workflow, not the model. Review whether the existing process is necessary, where decisions or hand-offs create delay, and what outcome the redesigned process should support. Only then should the organisation determine where AI adds value.
Once the workflow has been redesigned, architecture can help control costs. Options may include reducing unnecessary context, consolidating calls, setting consumption limits, and routing tasks to the most appropriate model.
The objective is not to restrict useful innovation. It’s to avoid paying for complexity that does not improve the outcome.
3. Govern spending against outcomes
Give finance, technology, and business teams a common view of AI consumption, total workflow cost and business value.
This requires clear ownership. A named owner should be accountable for each priority workflow’s cost per outcome and should have the authority to investigate variances, change the workflow or model, and increase or reduce investment.
Governance should also establish acceptable thresholds for quality, risk, and human review. A lower cost is not a saving if the output fails a compliance check, weakens an important decision, or transfers additional work to employees elsewhere.
4. Reinvest savings in further innovation
Cost discipline should create the capacity to do more, not simply reduce the AI budget.
Where a redesigned workflow releases genuine savings or capacity, leaders can establish a mechanism for reinvesting some of that value in other priority opportunities. This can produce a repeatable cycle: assess a use case, improve its economics, verify the result, and redirect resources towards the next high-value opportunity.
Over time, that discipline can allow an organisation to obtain more value from the same overall investment.
Embed discipline in the workflow
Policies and dashboards are important, but they may have limited effect if controls are separate from day-to-day AI use.
Where appropriate, organisations can incorporate cost and governance controls directly into AI-enabled workflows. These may include budget thresholds, defined escalation points, approved model-routing rules, and audit trails that show how a result was produced.
Controls should be proportionate to the task. A low-risk internal activity may not require the same oversight as a customer-facing, financial or regulatory decision.
This doesn’t remove people from the process. It enables human oversight to be applied where judgement, accountability, or intervention is most valuable. The aim is a technology-enabled, human-led model in which people have the information and authority needed to make informed decisions.
"The question isn’t how few tokens an organisation can use. It’s whether AI helps it rethink the work and create an outcome worth more than the cost of delivering it. If the workflow isn't redesigned but is simply automated, the case for higher spending is much harder to make."
Amy Ball,Reinvention Leader, PwC IrelandHow to get started: Five actions for Irish leaders
The next stage of AI advantage is unlikely to come from providing the most powerful model to the greatest number of people. It’s more likely to come from applying the right technology to the right work, with a clear connection between cost and value.
1. Establish where you lack visibility
Ask whether you know the total cost of each priority AI workflow, the business outcome it supports, and whether that outcome is being achieved. Include model consumption, infrastructure, oversight, and rework. Where that information is unavailable, identify the data, ownership, and reporting gaps preventing a complete assessment. Start with a limited number of material workflows rather than attempting to measure every use of AI at once.
2. Put AI value on the CFO and CIO agendas
Treat material AI expenditure as a shared business issue, not an occasional technology update. The CFO should help define how value and savings will be measured. The CIO should provide visibility of consumption, architecture, and performance. Business leaders should remain accountable for the outcomes. Agree a small set of common measures that allows these groups to make investment decisions using the same evidence.
3. Test cost and performance before scaling
Build consumption testing into the development and approval of AI use cases. Model how costs could change at the expected level of use and compare suitable models against the task’s quality, risk, and performance requirements. Establish a baseline cost per outcome before deployment. This gives decision-makers a basis for determining whether the use case should proceed and whether a premium model creates enough additional value to justify its cost.
4. Challenge the workflow before optimising it
Don’t automatically use AI to accelerate an existing process. First assess whether the process remains necessary, whether it supports the intended outcome, and where work can be removed or redesigned. Once the workflow is fit for purpose, select the lowest-cost model or combination of models that can reliably deliver the required result. This can prevent the organisation from embedding inefficiency in a faster but more complex system.
5. Give owners the authority to act
Assign a named owner to each priority AI workflow and its cost per outcome. Give that person access to the necessary financial, technical, and operational information, as well as the authority to address underperformance. They should be able to change routing or model selection, strengthen controls, redesign the workflow, renegotiate relevant arrangements, or recommend that investment be stopped. Accountability without decision rights is unlikely to create lasting discipline.
Discipline can create capacity for growth
AI costs are unlikely to be managed effectively through blanket restrictions or by maximising adoption without regard to outcomes.
A stronger approach links each material AI investment to a business need, tests its economics before scaling, and monitors whether it delivers the intended result. It combines disciplined spending with the capacity to redesign work and reinvest proven savings.
As access to AI capability becomes more widespread, competitive advantage may depend less on which organisation has the technology and more on which organisation uses it with the greatest discipline.
We’re here to help
Creating greater value from AI requires a clear view of costs, outcomes, and investment priorities. We can work with your finance, technology, and business teams to assess AI use cases, redesign workflows, improve cost visibility, and strengthen governance. Contact us today to identify the AI investments that matter most and establish whether they’re delivering measurable business value.
Create greater value from AI
Scale AI investments around measurable business outcomes.
{{filterContent.facetedTitle}}
{{contentList.loadingText}}
{{contentList.loadingText}}
Follow PwC Ireland
Menu