01Why do usage metrics fail the CFO?
Usage metrics fail the CFO because they measure how much AI is used, and the CFO answers for what each result costs. Logins can rise while cost per claim rises too, and the usage report will never show it.
The research shows how wide the gap is. Gartner's survey of 2,002 CIOs, published on 19 October 2026, found only 13% say their enterprise has realised significant or transformational value from AI, and the likelihood of a use case succeeding is still one in five. KPMG's AI Pulse for Australia (October 2026) found 75% of organisations say AI will stay a top investment priority, and only 2% report adoption has progressed to established ROI.
Cost visibility trails too. KPMG found 58% of Australian organisations can see their total AI cost. Deloitte's State of AI in the Enterprise survey of 3,500 leaders (October 2026) found 71% track total AI inference and API spend, 51% track token volume and 44% track per-unit inference cost. EY's global CEO Outlook (October 2026) found 16% of CEOs have comprehensive, real-time visibility of AI costs and high confidence in assessing the return.
AI ROI measurement starts with what one outcome costs, before and after.
02What is AI cost per outcome?
AI cost per outcome is the total monthly cost of running a process with AI, divided by the number of finished outcomes it produced that month, compared with the same figure from before AI.
An outcome is a unit of finished work the business already recognises, such as an invoice paid. A drafted email that a person rewrites is a step, and counting it inflates the result.
The cost side is the AI total cost of ownership for that process, in six parts:
- Model and token charges. Usage-based fees for every call the AI makes, including retries and long prompts.
- Platform and licences. The AI platform, vector stores, orchestration and any seat licences.
- Integration. The build that connects AI to core systems, spread over its expected life, plus the cost of keeping it running.
- Review effort. People's time checking, correcting and approving AI outputs, and handling the exceptions AI passes back.
- Governance and monitoring. Model testing, output sampling, risk review, audit records and incident handling.
- Remaining process cost. Existing systems and the people still doing the parts of the work AI does not touch.
The AI return on investment then follows from the unit figure: the saving per outcome, multiplied by volume, less the one-off investment, over a stated period. Gartner found 51% of CIOs expect AI to increase the total cost of ownership of technology solutions, and its analyst Alicia Mullery noted that "AI costs are probabilistic and determined at run time." So measure the cost side every month.
03How do you baseline the work before AI?
Baseline the work by measuring four things for the process in scope before anything changes: volume, cycle time, error rate and cost per outcome today, agreed with finance and frozen.
- Define the outcome. Write down what counts as one finished unit, and what does not.
- Pick a fair period. Use at least three months, so peaks and quiet weeks both appear.
- Count volume. Outcomes completed per month, from system records.
- Time the work. Cycle time from start to finished outcome, and people's hours by activity, from rosters, workflow logs or a short time study.
- Measure quality. Error, rework and complaint rates, using the same definitions you will use after go-live.
- Cost it. Loaded labour cost, the systems the process uses and any outsourced work, divided by volume.
- Sign it off. Finance agrees the method and the number before the build starts. Nobody revisits it once results arrive.
Finance signing the baseline before the build is what settles the argument when the first results reach the board.
04How do you build the cost per outcome number, step by step?
Build it by costing the redesigned process one component at a time each month, dividing by finished outcomes, and setting the result beside the frozen baseline.
Illustration only, using hypothetical round numbers. A service team resolves 10,000 requests a month. Before AI, each costs $40: $34 of people time, $4 of systems and $2 of rework. That is $400,000 a month.
After the process is redesigned, AI resolves 70% of requests with a person reviewing each one, and people handle the other 30% in full.
| Monthly cost | Before AI | With AI |
|---|---|---|
| People doing the work end to end | $340,000 | $102,000 (3,000 requests at $34) |
| People reviewing AI outputs | none | $42,000 (7,000 reviews at $6, about five minutes each) |
| Model and token charges | none | $15,000 |
| Platform and licences | none | $20,000 |
| Integration, $600,000 build spread over 24 months | none | $25,000 |
| Governance and monitoring | none | $15,000 |
| Existing systems | $40,000 | $40,000 |
| Rework | $20,000 | $20,000 |
| Total | $400,000 | $279,000 |
| Cost per outcome | $40.00 | $27.90 |
The gap is $12.10 per outcome, or $121,000 a month. Two sensitivities show where the number is fragile. If each review takes ten minutes, review cost doubles and cost per outcome rises to $32.10. If the token bill doubles, it rises to $29.40. Review effort moves the number more than the model does, so measure it from the first week.
05Who owns the capacity AI releases?
The executive who owns the process owns the capacity AI releases, and they decide before go-live where it goes. In the illustration, people time falls from $340,000 to $144,000 a month. That $196,000 is capacity. It becomes a saving or revenue only when someone acts on it.
It has three honest destinations:
- Redeployed. People move to backlog, growth or customer work with its own measure, such as more applications processed or faster complaint resolution.
- Absorbed. Volume growth is handled without new hiring, recorded as cost avoided against an agreed growth forecast.
- Banked. Roles are not backfilled as people leave, and the budget is reduced, so the saving appears in the cost base.
Record the choice for every affected role, with a date, and report any capacity with no destination yet. Gartner found organisations with explicit plans to redesign roles for a human and agent workforce are 2.4 times more likely to be in its top-performing group, and only 4% of CIOs have such plans. Our AI workforce and role redesign page sets out how roles are redesigned with the people who hold them.
Explore AI workforce and role redesign →
06How do you prove AI value to the board?
Prove AI value to the board with one page per use case each quarter, showing cost per outcome against the frozen baseline, the full cost of AI by component, and where the released capacity went.
The one-page measure set:
- Cost per outcome: baseline, this quarter and target.
- Volume and cycle time: against the baseline.
- Quality: error, rework and complaint rates, with human override rates.
- Full AI cost: each of the six components, with the token trend per outcome.
- Capacity: released, redeployed, absorbed, banked and still unassigned.
- Net value to date: against the costed value case approved at the start.
- Risk: incidents, audit findings and any change to the risk rating.
- Decision sought: scale, fix or stop.
Usage figures can sit in an appendix as context. The board decides on lines 1, 5 and 6.
07When should an AI use case be stopped?
Stop an AI use case when it misses the stop rules agreed before go-live and the cause cannot be fixed in a set time. Gartner found 31% of CIOs have guidelines for when a pilot should not move to production, and 17% have guidelines for abandoning low-value AI already in production.
Write the stop rules into the value case, for example:
- Cost per outcome is still above the baseline after two quarters of live operation.
- Quality falls below the baseline and stays there for a month.
- Review effort per outcome is not falling quarter on quarter.
- Token cost per outcome grows faster than volume for two quarters.
- Released capacity stays unassigned for two quarters.
When a rule trips, find the cause first: the process, adoption, data or the model. Fix it within a set time or stop, and move the spend to the next opportunity in the value case.
08What are the common traps in measuring AI ROI?
The most common trap is counting time saved that never turns into lower cost or more revenue. Five more are close behind:
- Ignoring review effort. Checking AI output is real work, and it can be the largest new cost.
- Rising token bills. Longer prompts, retries and agents calling other agents raise cost per outcome without anyone noticing, unless token cost is tracked per outcome.
- Counting steps as outcomes. A draft is a step. Only finished work goes in the denominator.
- Moving the baseline. Rebuilding the baseline after results arrive undoes the comparison.
- Crediting AI with everything. Process changes, staff changes and seasonal volume all move the number, so note them beside the result.
09Where does cost per outcome fit in an AI program?
Cost per outcome belongs in the AI business case before the build starts, and in every report after go-live. In the Five Levels of Enterprise AI, Level 3 is where return becomes measurable, because AI is wired into a redesigned process with value baselined. At Level 2, usage is all there is to measure.
Catalyst+, our AI transformation product, builds this in from the first week. Define baselines the work and the AI spend, and ends with a costed value case: where AI pays, how much, what it will cost and in what order to go after it, with an owner for each opportunity. Activate measures value from the first live AI agent against that case, and Scale funds each wave on the evidence from the last. Our AI value measurement service runs this inside every engagement, and AI governance keeps the records that make the numbers auditable.
AI ROI measurement works best when finance helps design it at the start.
Explore AI value measurement →Explore Catalyst+ →Read The Five Levels of Enterprise AI →AI governance →
Notes and sources
- 13% of CIOs report significant or transformational value from AI; one-in-five use case success; 51% expect AI to raise total cost of ownership; "AI costs are probabilistic and determined at run time" (Alicia Mullery, VP Analyst); 31% and 17% with pilot and production stopping guidelines; 4% with human-agent role redesign plans, 2.4 times more likely to be vanguards. Gartner, 2027 CIO and Technology Executive Agenda survey of 2,002 CIOs and technology executives (May to July 2026), press release 19 October 2026
- 75% say AI will remain a top investment priority; 2% report established ROI; 58% report visibility of total AI cost: KPMG Australia, AI Pulse Survey, Australian insights, October 2026
- 71% track total AI inference and API spend, 51% token volume, 44% per-unit inference cost: Deloitte AI Institute, The State of AI in the Enterprise, survey of 3,500 leaders in 25 countries (May to June 2026), press release 29 October 2026
- 16% of CEOs with comprehensive, real-time visibility of AI costs and high confidence in assessing ROI: EY, Global CEO Outlook Survey, October 2026 (1,200 CEOs in 21 countries, including Australia), published 1 October 2026
- The worked example in section 04 is an illustration using hypothetical round numbers. It describes no client.