Independent audit of the proposed opening facts

Audited 22 October 2026. The original McKinsey article, Datadog report and Datadog launch announcement were opened independently. This audit supplements the agent-economics ledger; it does not reproduce either dataset.

Recommendation

Use the survey and production-telemetry pair in the opening, with explicit scope. They are more relevant to enterprise leadership than a general agent multiplier: one records an existing organisational constraint, while the other records a change in actual production usage. Keep Anthropic’s architecture comparison and the historical price decline in the appendix. None of these sources establishes that Knauf’s bills must rise.

The two slides should take about 90 seconds together. Keep the whole talk at 23–25 minutes by shortening the existing introduction and moving the detailed hosting comparison into Q&A. Preserve the accepted-result denominator and practical case study.

McKinsey: supported as a respondent report

Original: The state of AI in 2026: On the road to ROI, published 25 August 2026, operating-cost section and Exhibit 8.

The article reports approximately one fifth of respondents saying operating costs, including tokens, constrained their organisation’s AI use. The full survey had 1,719 participants across 97 countries, fielded 4 May–8 June 2026 and weighted by national GDP. That is the full-survey sample, not a verified cost-question denominator: the exhibit footnote was not accessible in this audit.

Approved slide wording: “AI running costs already constrain some organisations.” Large figure: “1 in 5”. Caption: “survey respondents report cost-constrained AI use”. Source footnote identifies McKinsey’s 2026 survey. Notes identify the full-survey sample separately.

Do not say one fifth of all companies or one fifth of spending. Self-report and non-random participation limit generalisation. Preserve the counterpoint in speech: the article describes the constraint as emerging rather than widespread, and 60% still expect increased investment. No growth rate in cost constraints was established.

Datadog: supported as an organisation-level usage-intensity trend

Original: State of AI Engineering, Fact 5 and methodology. Its launch announcement is dated 21 April 2026.

Fact 5 reports year-on-year growth exceeding 2× at the median and reaching 4× at the 90th percentile. Method: average tokens per LLM span within each organisation, then cross-organisation summaries. This is not the median individual request, total organisational volume or accepted-outcome cost. “Power users” must not become a claim about individual people.

Approved headline: “Production AI calls are getting larger.” Numerals: “>2× median · 4× 90th percentile”. Caption: “Year-on-year growth in per-organisation average tokens per LLM call, Datadog customers.” Notes retain that the percentiles describe the distribution of organisation-level averages.

The report samples software/cloud-oriented customers; exact Fact 5 sample size, fixed-cohort treatment and precise observation endpoints are undisclosed in prose. It does not clearly specify input-only versus total tokens here. The character-count approximation is described under Fact 4; do not silently assume it applies to Fact 5. Do not equate this trend with bill growth or waste.

Stress test of the proposed argument

The defensible inference is: make cost visibility and runtime control part of scaling AI. It does not follow that every business needs a major optimisation programme now, that agent adoption caused the measured increase, or that lower token prices cannot offset higher usage. Larger requests may buy valuable capability. Different sampling, workload mix, model choice and caching can change the financial result.

Spoken bridge, our synthesis rather than source quotation:

The financial constraint is already real for some organisations. And production calls are becoming more intensive. That does not tell us what Knauf should spend. It tells us what we need to measure: the full workflow cost, the quality we accept, and where another euro of effort pays back.

Avoid an exponential curve, forecast spending line, or causal arrow joining the two studies. A five-mark survey graphic followed by two large growth numerals is enough. Do not add the 4×/15× architecture chart as a third opening statistic. In Q&A, it can explain one mechanism behind workflow intensity, with its original comparator and vendor-specific caveat.

Remaining limits

The McKinsey exhibit-level denominator and Datadog exact chart dates are not independently verified here. The sources support the scoped captions above without those details. A browser attempt to inspect the exhibit returned “Browser is not available: iab”; the text findings themselves were verified through the web research tool. No deck, live site or Git state was changed.