AI Answer Summary
- 93% of enterprise AI teams exceeded their AI budgets in 2026, according to McKinsey's QuantumBlack FinOps survey (n=75, May 2026)
- 60% of all agentic AI costs are consumed by response refinement — the loop where agents check, revise, and regenerate their own outputs
- Token prices have fallen.
AI Summary
- 93% of enterprise AI teams exceeded their AI budgets in 2026, according to McKinsey's QuantumBlack FinOps survey (n=75, May 2026)
- 60% of all agentic AI costs are consumed by response refinement — the loop where agents check, revise, and regenerate their own outputs
- Token prices have fallen 99%+ since 2022, yet enterprise LLM spending tripled in the 12 months to end-2025
- 1 in 5 organisations has already constrained AI use specifically because of operating costs (McKinsey State of AI 2026, n=1,719)
- The correct unit of measurement is cost per correctly completed and accepted business outcome, not cost per million tokens
Why Cheaper Tokens Are Not Producing Cheaper Work
Token prices have fallen more than 99% in roughly two years. Enterprise AI bills have tripled anyway.
That contradiction sits at the centre of a July 2026 McKinsey QuantumBlack report — Is That AI Agent Worth It? Agentic Economics and the Modern Operating Model — which maps, in unusually precise terms, where agentic AI money actually goes and why most organisations cannot yet account for it.
The numbers are stark. According to McKinsey's Enterprise AI FinOps Survey, conducted in May 2026 across 75 qualified respondents spanning five major industries, 93% of enterprise participants report exceeding their AI budgets. Separately, data from Menlo Ventures cited in the report showed that enterprise large language model spending tripled over a 12-month period by the end of 2025. Meanwhile, Stanford's HAI 2025 AI Index documented that inference cost for GPT-3.5-level capability collapsed from $20 per million tokens to $0.07 through 2024 — a drop of more than 99%.
The divergence between falling unit costs and rising total bills is not a paradox. It is a volume and architecture problem.
The Hidden Cost Structure of Agentic Workflows
When a business user submits a single request to an agentic AI system, the visible interaction masks a chain of operations that each generate token consumption and cost:
- Orchestration — the agent interprets the request and plans its approach
- Retrieval — relevant context, documents, or data are fetched from memory or external sources
- Tool use — the agent calls APIs, databases, or external services
- Reasoning loops — the agent works through intermediate steps
- Response refinement — the agent checks, revises, and regenerates its own output
- Validation — the output is evaluated against quality or safety criteria
- Exception handling — failures trigger additional attempts
- Human review — some outputs require human correction before acceptance
Each of these phases generates token consumption. A model with a lower token price can therefore produce a more expensive completed workflow when it requires more iterations, corrections, or supervision.
The 60% Problem: Response Refinement
The single largest cost centre in an agentic AI deployment is response refinement — the iterative loop in which an agent checks, revises, and regenerates its own outputs before returning a final answer. According to McKinsey, that process consumes 60% of total agentic AI costs.
For enterprise teams that assumed their spend was split evenly across retrieval, reasoning, and generation, that concentration is a significant recalibration.
Three structural forces compound the problem, according to the McKinsey report:
- Scale acceleration — enterprises are expanding AI workloads rapidly, so more tasks are generating token consumption
- Consumption pricing shift — LLM providers have broadly moved from flat subscription to consumption-based models, creating an incentive for longer, more elaborate outputs
- Model-task mismatch — expensive frontier models are frequently deployed for routine tasks that cheaper, smaller models could handle
The self-reflection loop is frequently presented as evidence of agent intelligence. Operationally, it can also represent expensive indecision. More reasoning is valuable only when it produces a measurable improvement in accuracy, completion rate, risk reduction, or acceptance without human correction.
The Uber Signal
The cost pressure is not theoretical. Uber reportedly consumed its entire 2026 AI budget by April — four months into the year — driven primarily by agentic coding sessions. CockroachLabs confirmed a similar pattern. In KPMG's May 2026 survey of 2,145 business leaders, nearly half had already scaled back AI agent deployments due to cost.
McKinsey's forthcoming 2026 State of AI global survey, fielded between May and June 2026 with 1,719 participants, found that one in five organisations has already constrained AI use specifically because of AI-related operating costs. That is a material brake on adoption, one that shows up in deployment decisions, model selection, and the scope of workflows assigned to agents.
Rethinking the Unit of Measurement
David Tepper, CEO of Pay-i, cited in the McKinsey report, frames the core problem precisely: "Tokens are not value, tokens are the bill."
That reframe matters operationally. Cost-reduction conversations anchored to per-token pricing miss the actual lever, which is the ratio of agent output quality and business impact to total operational cost.
For a VP of Operations or a CIO evaluating agentic AI portfolios, the relevant metric is not cost per million tokens. It is cost per completed, accurate, human-review-free task.
An agent that runs more refinement loops but delivers output that never requires correction may be cheaper in practice than a cheaper-per-token agent whose outputs routinely need human repair. Marketing operations teams will need to attribute agent costs to campaigns, customers, and workflow outcomes rather than treating AI expenditure as a single software line item.
The Agentic AI Unit Economics Framework
Enterprises need AI unit economics, not only AI adoption metrics. A practical framework connects every agent workflow to four measurable dimensions:
| Dimension | What to Measure |
|---|---|
| Accuracy | % of outputs accepted without human correction |
| Completion rate | % of tasks completed end-to-end without failure |
| Human intervention | Average corrections or reviews per completed task |
| Total cost per outcome | Full operational cost ÷ accepted business outputs |
Without this instrumentation, organisations cannot determine whether any given agent deployment produces value exceeding its full operational cost — which is precisely the question McKinsey identifies as the defining operational challenge for enterprise AI in this phase of adoption.
What B2B Leaders Should Do Now
1. Audit where your token spend concentrates
If you cannot attribute costs by workload phase — retrieval, reasoning, and refinement — you cannot target the 60% that McKinsey identifies as the dominant cost driver. Most enterprise AI dashboards show total token consumption. That is insufficient for cost management.
2. Evaluate model-task fit across your agent stack
Deploying frontier models on routine tasks is a documented cost driver. Map each agent use case to the minimum capable model tier and quantify the savings before your next budget cycle. Tasks like classification, summarisation, and formatting rarely require the most expensive models.
3. Build output-quality tracking alongside cost tracking
Measuring tokens spent without measuring accuracy, human correction rates, and task completion rates makes cost-value comparison impossible. The two data streams must be connected before any meaningful optimisation can occur.
4. Set refinement budgets for each agent workflow
Rather than allowing agents to refine outputs indefinitely, set explicit refinement limits per workflow. A maximum of two or three refinement passes, combined with clear acceptance criteria, can significantly reduce the 60% cost concentration without materially reducing output quality for most business tasks.
5. Require cost-value documentation for every active agent deployment
Use the McKinsey framing as an executive forcing function: for each active agent deployment, require a documented answer to whether the value of completed work exceeds the full operational cost of generating it. If the answer cannot be provided, the deployment is not ready for production scale.
The Commercial Implication for Marketing Operations
For B2B marketing and GTM teams, the agentic cost problem has a specific shape. Marketing agents — those handling content generation, campaign optimisation, lead scoring, and personalisation — are among the highest-frequency users of refinement loops, because marketing output quality is inherently subjective and often requires multiple revision passes before acceptance.
This means marketing operations leaders face a disproportionate share of the 60% refinement cost concentration. The solution is not to avoid agentic AI in marketing. It is to design marketing agent workflows with explicit acceptance criteria, quality gates, and human review checkpoints that reduce unnecessary refinement cycles.
Our Agentic AI service helps B2B marketing teams design agent workflows with built-in cost controls, quality gates, and attribution frameworks that connect every agent action to commercial outcomes. If your AI bills are rising faster than your results, speak with Modi about building the right unit economics from the start.
Frequently Asked Questions
Why are enterprise AI costs rising when model prices are falling? Enterprise AI costs are rising because agentic systems multiply the number of model calls, reasoning loops, retrieval operations, and validation steps required to complete a single business task. According to McKinsey's July 2026 FinOps survey, 93% of enterprise AI teams exceeded their budgets despite token prices falling more than 99% since 2022. The cost driver is workflow architecture and volume, not the per-token rate itself.
What is response refinement in agentic AI and why is it so expensive? Response refinement is the iterative loop in which an AI agent checks, revises, and regenerates its own output before returning a final answer. McKinsey's July 2026 report found this single process consumes 60% of total agentic AI costs. Each refinement cycle triggers additional model calls, reasoning, and token consumption. Agents that require many refinement passes before producing acceptable output are significantly more expensive than their per-token rate implies.
What is the right metric for measuring agentic AI cost efficiency? The right metric is cost per correctly completed and accepted business outcome — not cost per million tokens. David Tepper, CEO of Pay-i, cited in the McKinsey report, frames it as: tokens are the bill, not the value. An agent that runs more refinement loops but delivers output requiring no human correction may be cheaper in practice than a lower-cost-per-token agent whose outputs routinely need repair or supervision.
Sources: McKinsey QuantumBlack, "Is That AI Agent Worth It? Agentic Economics and the Modern Operating Model," July 2026; MarketScale reporting on McKinsey Enterprise AI FinOps Survey, July 19, 2026; Stanford HAI 2025 AI Index; Menlo Ventures enterprise LLM spending data; KPMG May 2026 business leader survey.
About the Author
Modi Elnadi is the founder of Integrated.Social, a London-based AI growth marketing agency specialising in agentic AI deployment, AEO, GEO, and B2B performance marketing. With hands-on experience designing and deploying AI agent workflows for enterprise marketing and GTM teams, Modi writes at the intersection of AI economics, commercial strategy, and operational accountability. His work focuses on connecting AI investment to measurable pipeline, revenue, and ROAS outcomes for B2B technology, fintech, and professional services clients.
