Key takeaways

  • At one technical engineering client measured by SkuzaAI, an AI agent answers 98% of employee questions without error at $0.04 per question across 9,000 documents, yet that number represents only one of four cost lines a board needs to see.
  • The IBM Institute for Business Value’s 2025 CEO Study reports that only 25% of AI initiatives have delivered the expected return on investment, and only 16% have scaled enterprise wide.
  • McKinsey reports that 40% of large companies are scaling AI agents, but only 6% of respondents qualify as AI high performers by EBIT impact, which means most organizations are spending before they are earning.
  • Gartner predicts that over 40% of agentic AI projects will be canceled by end-2027, a figure that reflects missing cost discipline far more than missing capability.
  • Four lines cover the full picture: build, run, maintain, and adoption. Every board approval that skips one of them is making a partial decision.

The number that relaxes the room

$0.04 is a genuinely good result. At one technical engineering client, an AI agent searches across 9,000 documents, answers 98% of employee questions without error, and returns each answer in about 15 seconds, all at $0.04 per question, measured by SkuzaAI at acceptance. When a board sees that number, the room relaxes. That is exactly the moment to press harder, because $0.04 is the run cost per transaction and nothing else. It is the end of the story, not the beginning.

The IBM Institute for Business Value’s 2025 CEO Study reports that only 25% of AI initiatives have delivered the expected return on investment over the last few years, and only 16% have scaled enterprise wide. The gap between expectation and outcome is rarely explained by the per-query price. It is explained by what was left off the page when the decision was made.

Build: the cost that only appears once

Build is the one-time investment required to design the agent, connect it to the data and systems it needs, and test it until people can rely on it. In at least one real deployment studied by McKinsey researchers, over 80% of total project effort went to implementation work rather than model work. That ratio should appear in every board slide that cites a per-query price.

Agentic workflows also consume materially more tokens than simple chat interfaces. Published estimates from the Stanford Digital Economy Lab range from 4x to 30x more tokens, with some coding-task analyses reaching far higher. Token volume is not an abstract technical metric. It directly determines infrastructure cost, testing time, and the number of edge cases that have to be resolved before the agent can be trusted at scale. Build cost is not a sunk cost to be forgotten after launch. It is the foundation that determines whether run cost stays at $0.04 or climbs.

Run: the number almost nobody asks for

Run cost is what the organization pays every time the agent completes a task. It is also the number almost no one asks for in a board meeting, perhaps because per-query figures like $0.04 look reassuring and therefore final. They are neither. Volume changes them. As an agent moves from pilot to enterprise deployment, query volume can grow by orders of magnitude, and pricing models vary significantly across platforms. Microsoft Copilot Studio, Salesforce Agentforce, Google Cloud Vertex AI, and Anthropic’s Claude all publish pricing, and the structures differ in ways that matter once usage scales.

Inference prices are falling, and that is a real and material fact. Epoch AI’s analysis of LLM inference price trends shows costs have dropped sharply across recent model generations. Falling prices favor adoption, but they do not eliminate the need to model run cost at scale before the board approves headcount reductions or process retirements that depend on the agent working. A projection built on today’s per-query price, multiplied by enterprise volume, belongs on the same slide as the $0.04 figure.

Maintain: an ownership question, not a technical detail

An agent trained on documents from one quarter can produce wrong answers in the next if those documents change and the agent does not. Products, prices, regulations, and internal processes change continuously, and so do the underlying models. Maintenance is not a background IT task. It is an ownership question that should have a named person accountable for it before the board approves deployment.

Gartner predicts that over 40% of agentic AI projects will be canceled by end-2027. The projects most at risk are the ones where build and run were costed but maintain was treated as an operational afterthought. The IBM Cost of a Data Breach Report 2026 puts the global average breach cost at $4.99 million. An unmaintained agent operating on stale access permissions or outdated compliance rules is a specific and foreseeable path toward that figure. Maintenance cost is not optional. It is the price of continued accuracy.

Adoption: the investment that determines whether the other three matter

Adoption spending covers the training, internal champions, and follow-up that make employees trust the agent enough to use it consistently. Without it, a technically accurate agent sits underused, and the organization pays for build, run, and maintain without collecting the return. MIT Sloan’s analysis of the heavy lifts in deploying AI agents identifies change management as among the most demanding, precisely because it cannot be automated.

McKinsey’s State of AI analysis finds that 40% of large companies are scaling AI agents, while only 6% qualify as AI high performers by EBIT impact. The gap between scaling activity and financial performance is at least partly an adoption problem. A four-cent answer that employees do not trust, do not ask, or do not know exists costs the same to run as one they use constantly. The difference in return is not four cents. It is the entire business case.

What the balanced picture shows

The case for AI agents is real. Verizon has reported $200 million in energy savings tied to AI-driven decisions. Klarna’s SEC filings show revenue per employee at $1.24 million, up 3.6x since 2022, with headcount down 49% and operating expenses down 8% since Q4 2022. Nature Cancer published a clinical agent result of $0.02 per note in a narrowly defined context. These are on-record results.

They are also not universal. Klarna has since publicly shifted toward ensuring customers can reach real people, a move that MIT Sloan commentary noted as a correction to earlier positioning. Organizations that have reported strong AI results have also grown more cautious about assigning clean dollar values to specific AI initiatives over time. The honest reading is that AI agents can create genuine value and that board-level cost discipline is not made optional by the existence of success stories. A four-cent run cost at a client that managed all four lines well is evidence of what disciplined deployment can achieve. It is not evidence that the other three lines can be skipped.

FAQ

What are the four cost lines a board should see before approving an AI agent?

The four lines are build, run, maintain, and adoption. Build covers design, integration, and testing. Run covers the per-task cost at scale. Maintain covers the ongoing work of keeping the agent accurate as data and models change. Adoption covers the human effort required to make the agent trusted and used.

Why is a low per-query cost not enough to justify an AI agent deployment?

Per-query cost measures only the run line. The IBM Institute for Business Value’s 2025 CEO Study found that only 25% of AI initiatives have delivered the expected return on investment, which suggests that total cost of ownership, not unit cost, is the more reliable predictor of whether a project delivers. Build, maintain, and adoption costs can easily exceed run costs in year one.

What happens to AI agent projects that skip the maintain line?

Gartner predicts that over 40% of agentic AI projects will be canceled by end-2027, and unmaintained agents are a specific risk factor. When underlying documents, pricing, regulations, or models change without a corresponding agent update, accuracy degrades, and trust erodes. Without a named owner accountable for maintenance before launch, the project has no one to catch that degradation early.

How should a board think about adoption costs?

Adoption costs are the training, champion investment, and follow-up that determine whether employees actually use the agent at the volume the business case assumed. An agent that runs at $0.04 per query but handles half the projected queries delivers half the projected return, while still carrying the full build and maintain cost.

Is the declining cost of AI inference a reason to delay the cost conversation?

No. Falling inference prices, well documented by Epoch AI, improve the economics of deployment over time, but they do not change the structure of the decision. A board still needs all four lines present to make a sound approval. Waiting for prices to fall further is not a substitute for knowing what build, maintain, and adoption will require from the organization today.

Sources