# Enterprise evidence for AI cost control

Research date: 22 October 2026. Scope: public primary research, with a preference for recent measured use and disclosed survey methods. This is evidence for the Knauf leadership talk; none of these observations describes Knauf.

## Recommendation

Use two complementary observations early in the talk:

1. **About one in five AI-using survey respondents reports that operating costs constrained their organization's AI use.** McKinsey's August 2026 survey supplies direct evidence of a current financial constraint.
2. **The median customer's average tokens per LLM call more than doubled year over year in Datadog's production telemetry.** This supplies measured evidence of changing usage intensity.

The defensible message is that cheaper models alone do not make workflow economics predictable. Cost control helps preserve room to scale useful work. Neither observation establishes that every company has a cost problem, that agents caused the increase, or that cost control itself increases ROI.

For the sparse deck, use a short assertion title such as **“Cost control protects the ability to scale.”** Pair a five-mark graphic with one accented mark and the large **“1 in 5”** figure, followed by a separate **“>2× tokens per call”** callout. Give each figure its own caption and citation; the populations are different. Avoid placing the two observations on a shared axis or multiplying them into an invented composite metric.

Suggested spoken bridge: “This is already constraining some deployments. And production requests are getting larger. We need a way to see what a completed piece of work costs, keep it within a useful range, and know when an optimisation pays back.” This is our interpretation, not a research finding.

## Main claim E01 — a current constraint

- **Approved wording:** “About 1 in 5 respondents at AI-using organisations say operating costs constrained AI use.”
- **Source:** [McKinsey, The state of AI in 2026: On the road to ROI](https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai), published 25 August 2026. [Official PDF, Exhibit 8, printed page 15](https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/the-state-of-ai-in-2026-on-the-road-to-roi-aug-2026.pdf#page=15).
- **Measurement:** Self-reported constraint from AI-related operating costs, including tokens. Not a measured bill increase or an independently audited budget overrun.
- **Population:** Survey total 1,719 respondents in 97 countries, fielded 4 May–8 June 2026, weighted by national contribution to global GDP. This item was asked of respondents whose organisations regularly use AI in at least one function. The survey total is not the item's denominator. Exhibit 8 provides industry-specific item counts but no single printed total item count.
- **Interpretation:** Supports introducing cost as an operating design concern. The reported constraint affects a minority; the paper explicitly presents it as emerging, not universal.
- **Graphic:** Five equal marks; one accented. Caption retains “respondents” and “AI-using organisations”. Put the eligibility detail in notes.

## Main claim E02 — usage intensity is changing

- **Approved wording:** “Median customer's average tokens per LLM call: more than 2× year over year.”
- **Source:** [Datadog, State of AI Engineering 2026, Fact 5 and Methodology](https://www.datadoghq.com/state-of-ai-engineering/). The [publisher's launch announcement](https://investors.datadoghq.com/node/16711/pdf) is dated 21 April 2026. The live report is not separately publication-dated and can be updated.
- **Measurement:** The methodology describes average tokens per LLM span for each organisation, followed by the median across organisations over a year. The same report gives a fourfold increase at the 90th percentile. This is per model call, not per end-to-end agent task, accepted result, user or euro spent.
- **Population:** More than 1,000 Datadog customers sending LLM telemetry; the broad methodology says thousands of customer companies. Exact Fact 5 sample size, matched-cohort membership and chart endpoint months were not disclosed in the readable methods. Do not fabricate these details.
- **Limitations:** Vendor customer selection; software/cloud-oriented population; no causal attribution. Changing models, customers and workloads may influence the trend. Falling prices, caching and changing input/output mix can offset token growth. Neither total cost growth nor a future growth rate follows from this observation.
- **Graphic:** A large “>2×” callout with the exact unit. Do not draw an apparently exact 2.0× bar or extrapolate the historical trend into a forecast. Use original graphics, not a recoloured copy of Datadog's licensed image.

## Four backup figures

| ID | Approved figure and scope | Source and denominator | Use and limitation |
| --- | --- | --- | --- |
| E03 | **27% → 40%**: respondents at organisations with annual revenue **≥$1bn** reporting agents at least scaling in one or more functions, 2025–2026. **+13 percentage points.** | McKinsey 2026, Exhibit 2, PDF p. 6. 2026 large-organisation group n=595; item asked of regular AI users, results rebased to the full group. | Strong backup for growth in deployment breadth. Repeated cross-sectional surveys, not necessarily the same companies. Does not measure spend, autonomous production volume or enterprise-wide agent rollout. |
| E04 | **31% / 63% / 98%**: surveyed FinOps practitioners managing some AI spend, reported in 2024 / 2025 / 2026. | [FinOps Foundation, State of FinOps 2026](https://data.finops.org/), dedicated AI section n=693; 2025 comparison also confirmed in the [2025 report](https://data.finops.org/2025-report/). Published 19 February 2026, confirmed by the [Linux Foundation release](https://www.linuxfoundation.org/press/state-of-finops-survey-ai-value-and-skills-top-priorities-as-finops-matures-across-technology-value-98-manage-ai-90-saas-64-licensing-48-data-center-1). | Good evidence that AI has entered FinOps' remit. Community sample, not all businesses. Being responsible for AI spend is not evidence of successful control, visibility, savings or maturity. Show discrete years; no projection. |
| E05 | **58%**: surveyed practitioners seeking to add AI cost-management skills through people or tooling in the next 12 months. | [FinOps Foundation's own keynote, PDF p. 9](https://www.flexera.com/sites/default/files/Flexera-FinOps-Forward-2026-FinOps-Foundation-State-of-FinOps-Keynote.pdf#page=9), multi-select n=695. The file is hosted by Flexera but the author and copyright holder are FinOps Foundation. | Supports a skills/capability gap and the practical session. It is a stated future staffing/tooling preference, not the percentage lacking skills or a measured failure rate. |
| E06 | **28%**: respondents reporting that AI technologies consume **more than 10% of enterprise-wide ICT budget**. | McKinsey 2026, body text/PDF p. 16. Survey n=1,719; this statement's item count is not separately printed. | Shows AI can be a material budget category. Not the average budget share, and not 28% of the budget. The item combines AI technologies; do not relabel it token/API/inference spending. |

The JSON ledger provides machine-readable wording, source locators, chart restrictions and research status. Exact short excerpts are stored there once per publication, within 25 quoted words per source.

## Evidence quality and rejected shortcuts

- **Rank E01 highest for the leadership argument.** A current, precisely scoped cost constraint is closer to the decision than an enormous market forecast. It still measures respondents' reports, not audited financial outcomes.
- **Rank E02 highest for a concrete operating mechanism.** Telemetry complements survey perceptions. It does not measure business value, so tie it to the accepted-result denominator already in the deck.
- **Keep FinOps' broad sample numbers out of the slide footer.** The report advertises 1,192 respondents, its keynote labels 966 deduplicated/cleaned responses, and the dedicated AI item is n=693. Use the item-specific count when needed. Methods and raw microdata are insufficient to reconstruct every denominator.
- **Do not use the FinOps executive-influence ratios.** The report's executive summary gives comparator percentages of 12% / 8% / 6%, while its main body and publisher release give 24% / 16% / 12%. An unresolved internal discrepancy makes this a poor deck claim, independent of the further correlation-versus-causation issue.
- **Do not use the technology-scope table to imply all 98% manage AI today.** Its question combines current and next-12-month scope. If E04 is used, cite the dedicated AI section and original publisher's explicit current-management statement, and document the discrepancy.
- **Do not present planned investment increases as realised spending.** McKinsey's 60% investment-increase expectation and FinOps' future priorities are intentions, not forecasts of measured expenditure.
- **Avoid an AI-ROI failure narrative.** McKinsey's 37% reporting enterprise EBIT impact does not show that the remainder lost money, had no local value, or lacked cost controls. It also overlaps the preceding event session more than these cost-specific observations do.
- **No generic cloud-waste percentage.** A cloud-wide waste estimate cannot identify AI waste or Knauf's optimisation potential.

## Reuse and source handling

All final claims should cite the original publication adjacent to the slide figure and retain a longer methods note in speaker notes/download references. The McKinsey PDF is the strongest stable locator because the generic article URL is updated across survey editions. FinOps data are published with an attribution licence; Datadog's report/images list CC BY-ND 4.0. Link the publications and redraw only the underlying factual measurements in an original design; do not copy or recolour publisher artwork.

Sources were inspected via public web retrieval on 22 October 2026. No survey microdata were available or analysed. No private company information was accessed. These are observational indicators for presentation context, not an estimate of Knauf's spend or savings.