Skip to content
Evidence

Key numbers

35 statistics from 29 sources that this publication relies on. Each was checked against its source before use, and each carries its sample and caveat.

Published Last reviewed
Download the full evidence register (CSV)How figures are checked
Using these numbers. Quote the figure with its sample and date. Survey results describe the people who answered, not the market. Vendor research is labelled as such. Multipliers such as “3.7 times more likely” are correlations, not effects. Estimates made by this publication are kept in the register but are not shown here.

Proving value

39% / 6%

39 per cent of respondents report any EBIT impact from AI; about 6 per cent attribute 5 per cent or more of EBIT to AI and report significant value.

Sample:
Global survey, self-reported
Caveat:
Earlier site wording '5 to 6 per cent report clear financial ROI' mis-stated this and has been corrected on all pages
Primary sourceReported findingMedium confidenceER7

Used on: five-levels-ai-economics, ai-value-management-layers, what-cfos-should-ask-of-ai-roi-claims

39%

39 per cent not confident they can connect AI spend to a measurable outcome their CFO would accept (34 per cent slightly, 17 moderately, 9 very); only 25 per cent of those with highly confident spend visibility can.

Source
State of Tokenomics, 2026-09-23
Sample:
Self-selected survey
Caveat:
Same wording as the source: 'not confident' is the lowest of four points
Primary sourceReported findingHigh confidenceER9

Used on: ai-value-gap, evidence-review-q3-2026, in-five-minutes

3.7x

88 per cent have defined ownership of AI economics; 12 per cent have no owner and none of those can connect AI spend to a CFO outcome; those with defined ownership are 3.7 times more likely to show value to the CFO.

Source
State of Tokenomics, 2026-09-23
Sample:
Self-selected survey
Caveat:
Correlation, not causation
Primary sourceReported findingHigh confidenceER10

Used on: ai-value-gap, evidence-review-q3-2026

4x

86 per cent evaluating or using a model router; users 4 times more likely to show CFO value.

Source
State of Tokenomics, 2026-09-23
Sample:
Self-selected survey
Caveat:
Correlation, not causation
Primary sourceReported findingHigh confidenceER11

Used on: evidence-review-q3-2026

44%

Bain Automation and AI Pathfinder Survey 2026: n=951; nearly 40 per cent of those who measured landed at 0 to 10 per cent savings; 37 per cent had targeted 11 to 20 per cent; 44 per cent funding the next wave from prior savings below target; 90 per cent increasing budgets.

Source
Bain & Company, 2026-06-01
Sample:
Survey of 951 global companies
Primary sourceReported findingHigh confidenceER26

Used on: ai-value-gap, what-cfos-should-ask-of-ai-roi-claims, in-five-minutes

2 to 4 years

Deloitte 2025: most report satisfactory ROI on a typical AI use case within two to four years; 6 per cent report payback in under a year; 13 per cent among the most successful projects.

Sample:
Survey of 1,854 executives in Europe and the Middle East, 24 interviews
Caveat:
Europe and Middle East sample
Primary sourceReported findingMedium confidenceER31

Used on: ai-roi-models, what-cfos-should-ask-of-ai-roi-claims

5%

BCG 2025: 5 per cent of companies achieve AI value at scale, 35 per cent are scaling, 60 per cent report no material value.

Sample:
More than 1,250 companies worldwide
Caveat:
Consultancy research; categories defined by BCG
Primary sourceReported findingMedium confidenceER32

Used on: ai-value-management-layers, itfm-ai

95%

MIT NANDA, The GenAI Divide (2025): about 95 per cent of organisations see no measurable P&L impact from generative AI.

Sample:
Review of 300+ public initiatives, 52 organisation interviews, 153 leader survey
Caveat:
Method caveats: broad definition of failure
Secondary reportReported findingMedium confidenceER36

Used on: ai-value-gap

56%

PwC 29th Global CEO Survey: 56 per cent of CEOs report no significant financial benefit from AI to date; 33 per cent report gains in either cost or revenue; 12 per cent report both.

Sample:
4,454 CEOs in 95 countries and territories, surveyed 30 September to 10 November 2025
Caveat:
Self-reported by CEOs; "significant" is the respondent's judgement
Primary sourceReported findingHigh confidenceER51

Used on: home, in-five-minutes

57%

Gartner survey of 160 senior finance leaders: 55 per cent saw positive overall returns from 2025 AI initiatives; at use-case level 57 per cent of returns were unclear, 18 per cent positive, 19 per cent negative and 6 per cent break-even.

Sample:
160 senior finance-function leaders; fieldwork dates not stated in coverage
Caveat:
Small sample, finance function only. Gartner newsroom could not be read directly
Secondary reportReported findingMedium confidenceER54

Used on: home, ai-value-gap, ai-portfolio-review-pack

Cost and prices

23% vs 4%

23 per cent asked providers for more transparency and granular data; 4 per cent for cheaper prices.

Source
State of Tokenomics, 2026-09-23
Sample:
Open-text answers coded by the Foundation
Caveat:
Coding method not published on the page
Primary sourceReported findingHigh confidenceER13

Used on: evidence-review-q3-2026

51% to 24%

51 per cent describe themselves as 8 to 10 on the frontier-model end of the scale now; 24 per cent expect to be there in 12 months.

Source
State of Tokenomics, 2026-09-23
Sample:
Self-selected survey
Caveat:
Expectation, not behaviour
Primary sourceReported findingHigh confidenceER14

Used on: evidence-review-q3-2026

$9k to $14k

Inferbase worked example: 9,000 USD monthly forecast, actual about 14,000 and sometimes 18,000; vendor's estimate that budgets miss by 30 to 50 per cent; fully loaded rate 1.4 to 1.6 times; 9x spread in list input price for Llama 3.3 70B (0.10 to 0.90 USD per million).

Sample:
Vendor experience and conversations with several teams. List prices captured from provider documentation, early May 2026
Caveat:
Earlier site wording overstated: corrected
Vendor or commentaryReported findingMedium confidenceER27

Used on: real-cost-enterprise-inference

$0.50 to $50

Anthropic list prices: output per million tokens 0.50 (Haiku 5.5, prompts up to 100,000 tokens; 2.50 above), 5 (Haiku 4.5), 10 (Sonnet 5.5), 20 (Opus 5.5), 50 (Fable 5.1); cache hits 2.5 to 10 per cent of input.

Sample:
Published price list
Caveat:
Perishable. Review in 90 days. Haiku 5.5 launched 7 October 2026; long prompts are billed at five times the standard rate
Primary sourceEvidenceHigh confidenceER28

Used on: ai-cost-benchmarks, real-cost-enterprise-inference, position-ledger

>90%

Gartner: by 2030 inference on a one-trillion-parameter LLM will cost GenAI providers over 90 per cent less than in 2025; savings will not be fully passed on to enterprises.

Source
Gartner press release, 2026-03-25
Sample:
Forecast
Caveat:
A forecast of provider cost, not enterprise price
Primary sourceReported findingHigh confidenceER35

Used on: ai-value-gap

FinOps practice

78%

78 per cent of FinOps teams report to the CTO or CIO.

Source
State of FinOps 2026, 2026-02-19
Sample:
1,192 respondents, over 83bn USD annual cloud spend, self-selected practitioners
Caveat:
As ER2
Primary sourceReported findingHigh confidenceER3

Used on: finops-ai

Self-funding

Many organisations asked to self-fund AI investment through optimisation savings.

Source
State of FinOps 2026, 2026-02-19
Sample:
1,192 respondents, over 83bn USD annual cloud spend, self-selected practitioners
Caveat:
Wording is 'many organizations report', no percentage given
Primary sourceReported findingHigh confidenceER4

Used on: ai-value-gap

Engineering

Unreliable

METR update: late-2025 follow-up unreliable due to selection effects; raw estimates 18 per cent speed-up (returning) and 4 per cent (new), wide intervals.

Sample:
Second study, 10 returning and 47 new developers
Caveat:
METR calls its own data very weak evidence for the size of the change
Primary sourceReported findingHigh confidenceER24

Used on: ai-coding-evidence

-1.5% / -7.2%

DORA 2024: 25 per cent more AI adoption associated with 7.5 per cent documentation quality, 3.4 per cent code quality, 3.1 per cent code review speed; estimated delivery throughput down 1.5 per cent and stability down 7.2 per cent; 39 per cent little or no trust.

Sample:
Practitioner survey, associations
Caveat:
Estimated associations from survey data. Google Cloud publishes DORA
Primary sourceReported findingHigh confidenceER25

Used on: ai-coding-evidence, debate/is-attention-the-scarce-resource

Operating model

2/3 vs 52%

McKinsey Global Tech Agenda 2026: nearly two-thirds of top performers say tech leaders are very involved in enterprise strategy, against 52 per cent of others.

Sample:
632 C-level executives and IT professionals, September to November 2025
Caveat:
Association, not causation
Primary sourceReported findingMedium confidenceER34

Used on: ai-value-management-layers

Standards

FOCUS 1.4

FOCUS 1.4 ratified by the FOCUS Steering Committee on 4 June 2026: adds 2 datasets, 47 columns, 6 attributes, 17 glossary entries, 2 supported features (Invoice Detail and Invoice Reconciliation).

Sample:
Specification release notice
Primary sourceEvidenceHigh confidenceER16

Used on: standards-map, finops-ai

5.0.1

TBM Taxonomy 5.0.1 is the latest version, with a new AI solution type and new towers.

Source
TBM Council taxonomy page, 2026-10-06
Sample:
Standards body page
Caveat:
Download is for members
Primary sourceEvidenceHigh confidenceER18

Used on: standards-map, tbm-ai