In Part 1, we argued that a cloud invoice lags. A useful cloud spend practice moves the decision upstream: baseline what happened, attribute ownership, forecast demand, act, and verify the result. AI makes that operating discipline harder to avoid.

A conventional service can become more expensive because it runs longer or serves more traffic. An AI feature can become more expensive while its successful user interactions barely change. One request grows a much longer context. An agent makes five tool calls instead of two. Retries repeat expensive work. Model routing shifts traffic to a higher-priced tier. Caches stop hitting. An enthusiastic rollout drives volume before the finance plan catches up.

The question is not Can we see AI spend? It is Can a named team explain the cost per useful outcome, model the tail risk, and intervene before a surprise becomes a monthly operating pattern?

The FinOps Foundation's 2026 State of FinOps survey included 1,192 practitioners. Of respondents, 98% said they manage AI spend, compared with 31% in 2024. That shows how broadly AI has entered the surveyed FinOps remit. It does not show that 98% of companies have mature AI cost governance. [E01]

A separate Harness 2026 State of AI in FinOps survey asked 700 engineering leaders and practitioners across five countries. In that respondent group, 72% reported a surprise AI bill in the prior year, 52% said there was no clear AI cost owner, and 56% described AI budgeting as guesswork. Harness is a commercial vendor, and the research should be read with that sponsorship and sample in mind. These are self-reported findings, not audited industry loss rates. [E02]

These results are not contradictory. A team can be responsible for AI spend yet lack an accurate cost attribution method, timely operational telemetry, or authority to change a product behavior. Being assigned a problem is not the same as being equipped to govern it.

Market forecasts add a different signal. Gartner projected worldwide AI-optimized IaaS spending of about $42.3 billion in 2026 and $66.1 billion in 2027, up from $21.5 billion in 2025. Gartner also forecasts inference workloads to account for 55% of AI-optimized IaaS spend in 2026 and 59% in 2027. These are market projections for infrastructure, not forecasts for API tokens, all enterprise AI costs, or any reader's cloud bill. [E03]

A bigger market creates urgency, not a universal prescription. The technical risk sits in each product's demand shape, call behavior, model mix and ability to connect AI cost to an outcome.

The simplest AI budget multiplies forecast requests by a blended price per request. That is a useful first estimate and a weak control if its inputs are treated as fixed.

Successful outcomes are the customer or business result worth funding. Billable model calls are what the system actually executes to deliver them. Those counts diverge whenever retries, agent steps, evaluations or failed paths consume tokens. Meanwhile, cost per billable call moves with the token mix and selected model.

A practical cost model is:

Monthly modeled cost = successful outcomes × billable model calls per outcome × [(input tokens per call × effective input rate) + (output tokens per call × effective output rate)] ÷ 1,000,000 + identifiable fixed AI platform cost.

Use rates per million tokens in a consistent currency. Split cached reads, cache writes, multimodal input, hosted capacity, tool/API charges, and vendor-specific billing classes into their own terms whenever material. Do not pretend a single token rate captures every purchase model. Vendor documentation for Amazon Bedrock, for example, distinguishes cached-input rates from normal input rates and notes that cache hits are not guaranteed. [E04]

The more practical denominator is cost per successful business outcome, not simply cost per token. Cost per token helps engineers locate price and workload changes. Outcome cost shows whether the product economics improved. Track both, plus latency, accuracy, safety, and service reliability so that cost control does not reward unusable results.

Consider a hypothetical AI-assisted workflow. The assumptions below are made up for teaching; they are not current vendor prices, survey benchmarks, operational measurements, or recommended budget amounts. A successful outcome means one completed business workflow. The model assumes every workflow ultimately succeeds; real operators should track failures and abandoned runs separately.

 The high-exposure model is about 3.2 times the base-case monthly bill. That change is not driven by request volume alone. It compounds changes in context length, response size, attempts per outcome, rates, and fixed expenses. The output illustrates which inputs must be owned. It proves nothing about any real company.

Suppose the product team has a reasonable base-case demand estimate but cannot confidently bound context growth. A two-dimensional sensitivity view is more useful than an increasingly precise single number. In the following chart we hold base rates, call amplification and fixed costs constant, and vary only outcome volume and tokens per model call.

This sensitivity view is not a probability distribution. The darker or brighter cells do not mean scenarios are more or less likely. Assign probabilities only when evidence supports them. A responsible budget includes explicit downside or tail scenarios, escalation thresholds, and owners for revised assumptions.

Part 1 introduced the Cloud Spend Decision Loop. The five stages remain the same. What changes is the granularity of the evidence and the speed of the feedback cycle. This is CloudLoom's framework, not a claim of external standardization.

Baseline. Reconcile billed AI charges with model usage and application telemetry. Record prompt and completion tokens, cached tokens where reported, billable calls, provider, model, environment, and billing adjustments. Measure freshness and coverage. Flag any gap between usage telemetry and invoice totals before declaring precision.

Attribute. Assign a cost owner per AI product or workflow. Attach usage to an application, team, feature, or business capability using permitted metadata, without logging sensitive prompt content by default. Make shared and unallocated costs visible. Finance should know who approves budgets; engineering should know who can change traffic routing or limits.

Forecast. Estimate outcome volume, token depth, model mix, retries, and effective rates separately. Model a normal range and a stress case. Review assumptions when features, prompts, context windows, agent loops or provider pricing change. An AI forecast is a maintained engineering model, not a financial spreadsheet that is touched once a quarter.

Act. Prioritize low-risk controls: budget envelopes by product and environment, model-selection policies, bounded agent steps, retries with limits, cache evaluation, backpressure, rate caps and clear pause/rollback procedures. Test quality and reliability before tightening controls. A cost-saving change that breaks the workflow is not verified value.

Verify. Compare expected and observed billed cost per successful outcome and material non-token charges. Check quality, latency, failed runs, and customer impact. Separate modeled avoided exposure from confirmed savings. If the decision did not work, adjust the model and the control, not just the dashboard.

The first control is not a cost alert. It is a product owner who can take action. Give that owner a weekly decision-grade view: actual spend versus plan, cost per successful outcome, call amplification, output/input token mix, and the largest unresolved charge categories.

Add budget envelopes and intervention triggers before usage scales. For example, the team might agree to investigate a 15% week-over-week unit-cost increase, slow optional batch work at a forecast-to-budget threshold, or require approval before opening a higher-cost model route. These are illustrative governance triggers, not industry-standard thresholds. Validate any threshold against your product's volatility and customer commitments.

Build safe interruption paths. Rate limiting, circuit breakers, step ceilings and feature flags can prevent runaway work; incorrect settings can also degrade availability. Make the response graduated: investigate, reduce noncritical workload, require explicit owner approval, and only then pause where the business impact is understood. Every automated action needs observability, ownership and rollback.

Finally, tie invoices back to outcomes. Track tokens and GPU utilization where relevant, but report an accessible business measure: cost per resolved support case, per valid document processed, per accepted code review, or per completed research workflow. If the outcome metric can be gamed or quality drops, the economics are incomplete. The FinOps Foundation's unit-economics capability explicitly links AI cost to initiative value and measurable outputs. [E05]

The following statements are CloudLoom analytical predictions, not Gartner forecasts, survey measurements, or claims from an implemented lab.

Prediction 1: Unit economics will become a release criterion. More AI product reviews will require projected cost per successful outcome and an escalation envelope before a feature scales. This follows from the technical instability of token and agent workload shapes, but will not happen uniformly across organizations.

Prediction 2: Execution amplification will become a leading operational KPI. The ratio of billable calls to successful outcomes will become more actionable than raw requests alone, especially for agents that invoke multiple tools or retry pathways. Teams should measure it before assuming savings.

Prediction 3: Spend protection will sit closer to the runtime. Monthly budgets will remain essential for finance, while engineering teams increasingly pair them with near-real-time usage limits, model-routing rules, and product-specific quality constraints. Controls will require cautious rollouts to avoid reliability regressions.

Prediction 4: Value will beat visibility as the test of maturity. Teams that reconcile the bill but cannot name an outcome, an owner, and a decision will struggle to defend continued AI investment. A cheaper model is not a better decision if it produces more failed work.

Pick one production or pilot AI workflow. Write down its owner, last seven days of successful outcomes, billable model calls, tokens by billing class, effective cost, and one intervention threshold. Identify unknowns instead of inventing values. Then ask whether engineering, product, and finance agree on what a successful outcome means.

If they cannot answer, that is the first useful finding. Build the baseline before automating enforcement.

Which path describes your AI spend journey today: chasing surprise bills, building forecasts, or managing cost against delivered value?

Sources. [E01] FinOps Foundation, State of FinOps 2026; survey of 1,192 practitioners. AI scope percentage refers to participating respondents, not the global enterprise population. [E02] Harness, State of AI in FinOps 2026; vendor-sponsored survey of 700 engineering leaders and practitioners across five countries. Percentages are self-reported respondent answers. Cost or waste claims in that research are not independently audited outcomes for this article. [E03] Gartner, Worldwide AI-Optimized IaaS Spending Forecast, August 10, 2026. Segment is infrastructure spending; 2026–2027 figures are published forecasts. [E04] AWS, Amazon Bedrock Prompt Caching documentation. Rates, support and charging classes vary by model and provider; verify against current contracts. [E05] FinOps Foundation, Capability: Unit Economics; conceptual guidance, not empirical savings proof.