AI is getting cheaper by the token.
That does not mean the work is getting cheaper to complete.
An agent can take a different path each time it receives the same assignment. It may call more tools, repeat a search, lose context, retry a failed step, or produce an answer that requires a professional to check and repair it. The model charge may look small while the cost of producing one dependable business result keeps growing.
That is the measurement problem underneath this week’s AI ROI news.
Enterprises are buying more AI while finance leaders ask for harder proof. But many organizations can see what AI costs without consistently connecting that cost to the value it creates. They are measuring tokens, licenses, users, and runs when the unit that matters is the successful outcome.
The better question is not, “How much did the AI cost to use?”
It is, “What did it cost to complete this piece of work correctly—and how does that compare with the way we completed it before?” Here are the numbers—and what they mean for professionals trying to turn cheaper intelligence into better business economics.
The numbers worth knowing this week
The same agentic task can consume 30 times more tokens from one completion to another. McKinsey explains that an agent may choose different tools, reasoning paths, and retries each time it performs the work. It also estimates that roughly 60% of an agentic task’s cost can come from checking, repairing, and reverifying the answer. Cost per token cannot reveal those differences. Leaders need the total cost of a completed task that meets an agreed quality standard. McKinsey
Only 12% of organizations consistently assess AI’s value against its cost across the enterprise. KPMG surveyed 2,131 senior leaders across 20 countries, territories, and jurisdictions. Among organizations already reporting established returns, the figure rises to 48%. Most respondents review costs during approval or monitor them in operation; far fewer complete the harder step of judging that cost against the outcome produced. KPMG also reports average planned AI investment of $210 million over the next 12 months—not spending already incurred. kpmg.com
AI vendors now receive 8.1% of enterprise software spending in Zip’s dataset, up from 1.4% a year earlier. Yet 21% of purchase requests involving AI vendors were rejected or canceled, compared with 12% for other vendors. Zip analyzed approximately $18 billion in approved purchase-request value from consenting customers with sufficient history. The data cannot show why each request failed, so it does not prove ROI concerns caused the rejections. It does show that AI is gaining budget while facing a tougher purchasing filter. zip.com
72% of finance leaders surveyed by Esker said their organizations spent more than planned on AI initiatives. Among CFOs, 65% reported difficulty connecting AI usage to specific business outcomes. The study covered 338 finance leaders, including 137 CFOs, and was sponsored by a company selling finance-automation software. It should be read with that context. Still, the gap it identifies is clear: visibility into usage is not the same as evidence that the work was completed better or more economically. esker.com
The pattern that matters
The unit of AI value is changing.
For traditional software, leaders could often use seats, logins, or adoption as early indicators of value. For agentic work, even the same task can vary dramatically in cost and quality. One successful run tells you very little about what the next hundred will cost—or how many will require a person to intervene.
That is why the right measure is cost per successful outcome.
To calculate it, the organization first has to define “successful.” The work may need to be accurate, complete, on time, compliant, documented, and accepted without material correction. Then the organization adds the full cost of producing those successful outcomes: model usage, software, integrations, monitoring, retries, corrections, professional review, training, and support.
The resulting measure is simple:
Total operating cost ÷ successful outcomes = cost per successful outcome.
That denominator matters. Dividing by every task the agent attempted can make weak performance look efficient. A failed task is not a unit of value. Neither is an answer that saves five minutes of drafting but creates fifteen minutes of checking and repair. The software-development evidence makes the point. CIO Dive’s coverage of McKinsey’s 2026 technology report says only one-quarter of companies reported meaningful acceleration in product-development cycles from agentic tools, while 30% reported lower productivity after adoption. At the same time, equity investment in the agentic-software-development category is projected to grow more than twelvefold from 2025 to 2026.
That 12-times figure describes investment flowing into the technology category—not enterprise spending on coding agents. The mismatch still matters: money is moving faster than dependable operating results. CIO Dive
The market is beginning to respond. Ema, an AI-agent vendor that raised $77 million this week, told TechCrunch it prices around completed tasks and business outcomes rather than seats or tokens. Its growth, retention, and margin figures are company-reported vendor results—not independent evidence of customer ROI.
But the pricing model points in the right direction: charge for work completed, then judge whether that work creates more value than it costs. TechCrunch
The EMDUIT take — for the professional putting AI to work
Consider a real estate brokerage using an AI transaction-coordination apprentice.
The easy metrics are impressive: tasks created, emails drafted, documents reviewed, reminders sent, and agent runs completed.
None of those proves the transaction was supported successfully. The brokerage needs to define the unit of value more carefully. A successful outcome might be one transaction milestone completed accurately, on time, documented in the approved system, and escalated to a qualified person whenever judgment or compliance review is required.
Now the economics can be measured.
How many milestones were completed successfully? How many required correction? How much coordinator and agent time was still needed? Did the workflow reduce missed deadlines, rework, follow-up time, or outside support? Could the same team safely support more files without weakening the client experience?
The full cost must include the platform, the AI workforce, integrations, supervision, training, maintenance, and the time people spend reviewing exceptions. That total is divided only by the milestones that met the success standard.
The brokerage can then compare the new workflow with its baseline. If cost per successful milestone falls while accuracy, timeliness, and service remain strong, the AI is creating an economic advantage. If token prices fall but retries and review time rise, the apparent savings are not reaching the business.
This approach also protects the human role. Review time is not automatically waste. In a licensed profession, human judgment may be the control that makes the outcome trustworthy. The goal is to measure that time honestly, place it where it matters, and remove avoidable correction—not pretend oversight has no cost.
Cheap intelligence is useful. Dependable completed work is valuable.
Do not ask what one AI run costs. Ask what one successful business outcome costs. What outcome would give your organization the clearest unit for measuring AI value?
If you are a licensed professional or business leader, share the task you would measure by successful completion rather than usage, seats, or tokens.