The Regime Investor· ·8 min read
Why cheaper AI tokens do not prove a company is saving money
A practical way to measure enterprise AI cost per accepted outcome, separate cash savings from capacity, and test management’s ROI claims.
A lower model price can produce a larger total bill
The price of an AI model is only one input into the cost of an AI-enabled workflow. A business can pay less for every million tokens and still spend more overall if employees use the model more often, prompts and responses become longer, agents call several models for one task, or the company adds retrieval, security and monitoring around the model.
This is the same arithmetic seen in many efficient technologies. Unit cost falls, usage expands, and the total bill depends on which effect is larger. Stanford’s 2025 AI Index illustrated how quickly inference prices had fallen: the cost of querying a model at roughly GPT-3.5 performance dropped from US$20 per million tokens in November 2022 to US$0.07 by October 2024. That price decline is valuable, but it does not say how many calls a company now makes or whether the output is accepted.
The right unit is therefore not a token. It is a completed business outcome: a customer issue resolved correctly, a compliant document accepted, a software change deployed without rework, or a sales lead converted. Every model call, integration and human intervention needed to reach that outcome belongs in the cost numerator; accepted outcomes form the denominator.
Start with the workflow that existed before AI
An AI return calculation needs a baseline. Before deployment, record how many cases the team completes, the elapsed time per case, paid labour time, error and rework rates, customer or employee waiting time, and any losses caused by delay or mistakes. Without that baseline, a faster-looking demonstration cannot establish incremental value.
Next, define the boundary. If AI drafts a response but a specialist must check every fact, rewrite half of it and enter the result into another system, the workflow is not “draft generation”. It runs from the original request to an accepted result in the company’s official system. Moving work from an employee to an unmeasured reviewer does not remove the cost.
The same rule applies to revenue. More generated leads, messages or product ideas are intermediate outputs. Revenue benefit begins only when the company can attribute a higher conversion rate, better retention, increased volume or improved pricing to the AI-assisted process, after subtracting sales merely shifted from an existing channel and the cost of delivery.
Count four layers of cost
The first layer is direct consumption. It includes model input and output tokens, data-indexing and retrieval calls, image or audio processing, agent retries, evaluation calls and idle reserved capacity. A multi-step agent may make many hidden calls before presenting one answer.
The second layer is the supporting application and data stack. Databases used to retrieve company information, cloud compute, licences, identity controls, system monitoring, data cleaning, connectors and vendor-management work can exceed the marginal inference bill. Self-hosting replaces an API invoice with AI chips, the risk of paying for idle capacity and an operations team; it does not make compute free.
The third layer is human work. Staff need training, workflow redesign and support. Domain experts may review outputs, investigate exceptions and repair errors. Security, legal, risk and compliance teams also spend time approving and monitoring the system. NIST’s generative-AI profile explicitly calls for operator training, real-world evaluation, structured feedback, ongoing monitoring and incident handling. These are necessary controls, not optional overhead that can be ignored in a return calculation.
The fourth layer is failure and change. Wrong answers create rework, customer remediation, delayed decisions and sometimes regulatory or security exposure. Models and vendor prices change, prompting fresh tests and migrations. A system that saves five minutes on successful cases but doubles the number of escalations may make the queue slower.
CBA’s disclosure shows both progress and the measurement limit
Commonwealth Bank provides a useful Australian example because it reports both technology investment and measured AI benefits. In its FY2026 results, CBA showed gross technology investment spend of A$2.4 billion, up from A$2.3 billion in FY2025. The increase funded technology-infrastructure refresh as well as generative-AI capability, including tools, employee training, AI-ready infrastructure and use cases.
CBA separately reported about A$200 million of measured gross benefits from AI use cases in FY2026, including capacity that was reinvested rather than removed from the cost base. About A$100 million was incremental during the year. The bank expected gross benefits to double in FY2027 and exceed investment, but that is a forward-looking company expectation.
It would be wrong to divide A$200 million by A$2.4 billion and call the result CBA’s AI return. The A$2.4 billion is a broad technology budget, not a disclosed AI-only cost. The A$200 million is gross benefit, includes reinvested capacity and is not necessarily cash saved. The two figures have different boundaries.
The operational evidence is still meaningful. CBA reported more technology changes, fewer major incidents, faster restoration, broad staff engagement with AI and an 86 per cent resolution rate for agentic messaging, meaning conversations resolved without a human-assisted service path. These measures show where benefits may arise. Investors still need to learn how much of the improvement is incremental to AI, how the benefit is valued, what ongoing cost supports it and when it reaches operating expenses, revenue or risk losses.
Measure accepted outcomes, not generated output
A useful monthly scorecard begins with volume: eligible cases, AI-assisted cases and accepted completions. It then records direct model and platform cost, paid specialist checking time, exception handling, rework and the cost of incidents. Divide that total by accepted outcomes, not by prompts or drafts.
Suppose a team handles 10,000 cases a month. Before AI, each accepted case costs A$12 in labour and systems. After AI, model and platform use costs A$1.20 per case and ordinary labour falls to A$7. If review, rework and exception handling add A$2.50, the new total is A$10.70 and the saving is A$1.30 per accepted case. At 10,000 cases, that is A$13,000 a month before fixed implementation cost.
If implementation and integration cost A$390,000, a simple payback estimate is 30 months: A$390,000 divided by A$13,000. But if volume doubles to 20,000 cases and rework adds another A$2 per case, cost rises from A$10.70 to A$12.70—A$0.70 above the pre-AI baseline, or A$14,000 more a month before fixed costs—even though token prices fall. A lower acceptance rate would make the result worse. The example is illustrative; each business needs its actual volumes, wages, error costs and cash timing.
Quality must remain paired with speed. Track the share accepted without material correction, defects found after release, escalation rate, customer outcome and cycle time from request to final acceptance. A productivity gain exists only when the business produces at least the same quality and risk outcome with fewer scarce resources, or produces a better outcome worth the additional cost.
Separate capacity from cash savings
AI often releases minutes rather than employees. If a team uses those minutes to handle more demand, improve service or reduce a backlog, the benefit is real but appears as capacity, not an immediate fall in operating expense. Calling all released time a cash saving overstates short-term profit.
Keep three benefit columns. The first is cash removed from the cost base, such as a cancelled licence or avoided contractor expense. The second is capacity redeployed to measured work, such as additional resolved cases. The third is risk or quality improvement, such as fewer losses or incidents. Do not add them together unless each uses a consistent valuation rule and avoids double counting.
Revenue benefits need the same discipline. Attribute only the incremental margin from customers or transactions that would not otherwise have occurred, and allow for the time between deployment, adoption and cash collection. A two-year payback forecast should show monthly cash flows, not merely a cumulative productivity estimate.
What investors should ask management
Ask which workflows produced the claimed benefit and what the pre-AI baseline was. Request cost per accepted outcome, not only token price or user adoption. Clarify whether reported benefit is gross or net, realised cash, reinvested capacity, avoided future cost, revenue or risk reduction.
Then ask about durability. Does usage growth raise the bill faster than model prices fall? How much review and rework remains? What happens when the vendor changes a model, price or data policy? Are employees using the tool voluntarily because it helps, or is licence activation being counted as adoption?
Finally, reconcile the scorecard to the accounts. Cash savings should eventually appear in operating expense or avoided capital expenditure. Revenue should appear with a margin and collection period. Reinvested capacity should show up in higher throughput, shorter queues or better outcomes. If a benefit never reaches an operating measure or a financial statement, its valuation deserves a haircut.
That buyer-side test complements the supplier-side question of where AI infrastructure spending becomes durable profit. The AI chipmaker and platform cash-flow comparison follows that separate path through capital expenditure, depreciation, margins and free cash flow.
When the return claim becomes credible
The conditions that would change an unproven return claim into a credible one are a decline in cost per accepted outcome over several periods, quality and incident rates that hold or improve, and a reconciliation from gross benefit to cash, throughput or avoided loss. The evidence is stronger when the result survives a model change and does not depend on one unusually enthusiastic team.
Counter-signals that weaken the view are usage and multi-step agent calls expanding faster than unit prices fall, persistently high specialist checking, error costs emerging after deployment, or adoption that displaces no existing expense. A narrow AI benefit compared with a broad technology budget, or a gross capacity estimate presented as net profit, also fails the reconciliation test.
Cheap tokens are an input advantage. They become an enterprise return only when the whole workflow delivers more accepted outcomes for less cash after allowing for error and risk costs.
This article provides general information only and does not take account of your objectives, financial situation or needs.
Sources
More through this lens