Skip to content
Book a discovery call
Rahey AI

Premium

Beyond the AI Hype: AI Strategy 2027/2028

A strategy paper separating demonstrated AI capability from hype, and technical progress from business value, for leaders deciding where AI belongs in their organisation.

Published 6 October 2026

Executive summary

For fifty years, artificial intelligence has moved through a repeating cycle: promise, excitement, disappointment, renewed investment. Twice that cycle ended in a funding winter. Today the cycle seems to be running again, and the rational executive question is: what, specifically, is different this time?

This study investigates one proposition: the current acceleration is not the product of a single breakthrough, but of five forces converging: data, compute, algorithms and architectures, infrastructure, and capital.

Its findings, in brief:

  1. The capability gains are real and measured at a level no previous AI cycle achieved.
  2. High adoption coexists with widespread failure to show profit-and-loss impact.
  3. The binding constraint is now reliability, workflow design, data quality and organisational absorption.
  4. The strategic question has changed from "which AI tool should we buy?" to "which parts of our work should be redesigned around increasingly capable intelligence?"

Key findings

Breakthroughs are convergences, not single algorithms. In 2012 AlexNet cut the ImageNet top-5 error rate from 26.2% to 15.3% using ideas that were decades old; what changed was data, compute, training practice and storage arriving at once. Ideas can stay dormant until enabling conditions catch up. (page 9)

Source: Krizhevsky, Sutskever and Hinton (2012); CNN lineage from LeCun 1989 to 1998; ImageNet from 2009.

Reliability does not improve automatically with scale. Scale improves benchmark performance and breadth, but hallucination, calibration and consistency do not reliably improve with it; the study calls this the section's most commercially important finding. (page 14)

Source: Persistent hallucination and calibration findings across model generations; Kaplan et al. (2020); Hoffmann et al. (2022).

Adoption is not value. 78% of organisations used AI in at least one function in 2024, up from 55% a year earlier, while roughly 95% of about 300 enterprise AI initiatives showed no measurable profit-and-loss impact. The two figures measure different things. (page 18)

Source: Stanford AI Index 2025; MIT NANDA (2025).

Feeling faster is not being faster. In a 2025 randomised trial, experienced developers were about 19% slower with AI tools on their own mature codebases, while expecting to be 24% faster and believing afterwards they had been 20% faster. (page 18)

Source: METR randomised controlled trial (2025), 16 developers, 246 tasks.

The constraint has moved inside the organisation. Workflow design, process maturity, data quality, integration depth, governance, employee adoption, task-level reliability and measurement discipline now decide whether capability becomes value, and every one sits inside the buyer's organisation, not the vendor's model. (page 19)

Source: Synthesis of the Section 11 evidence (MIT NANDA 2025; METR 2025; Peng et al.; Humlum and Vestergaard 2025).

Table of contents

  1. 1. Executive summary, page 3
  2. 2. 01 The hype cycle and funding winter, page 4
  3. 3. 02 When intelligence was coded, page 6
  4. 4. 03 Teaching machines to learn, page 7
  5. 5. 04 The neural network story, page 8
  6. 6. 05 The breakthrough was not the algorithm, page 9
  7. 7. 06 Recombination: the AlphaGo lesson, page 10
  8. 8. 07 The Transformer inflection, page 12
  9. 9. 08 The scale question, page 14
  10. 10. 09 Reasoning and action, page 16
  11. 11. 10 The AGI question, page 17
  12. 12. 11 Capability vs value: the ROI evidence, page 18
  13. 13. 12 Why this time may actually be different, page 20
  14. 14. The frontier models analysis, page 21
  15. 15. 13 Where machine intelligence comes from, page 22
  16. 16. 14 What this means for AI strategy, page 23
  17. 17. The final argument, page 24
  18. 18. Evidence notes and claims requiring caution, page 25
  19. 19. Bibliography note, page 26

Methodology

The study traces the field's history and current evidence through primary research papers and institutional sources. Conflicting sources are shown in conflict, forecasts are labelled as forecasts, company claims are labelled as company claims, and anything that could not be traced to a source was removed. Each major claim is graded in an evidence table as established, supported or speculative.

Request the study

USD 100Launch offer USD 50

Sources

  • Turing (1950)
  • The Dartmouth Proposal (1955)
  • Rosenblatt (1958)
  • Minsky and Papert (1969)
  • Lighthill (1973)
  • Rumelhart, Hinton and Williams (1986)
  • LeCun et al. (1989, 1998)
  • Hochreiter and Schmidhuber (1997)
  • Krizhevsky, Sutskever and Hinton (2012)
  • Silver et al. on AlphaGo (2016) and AlphaZero (2017)
  • Vaswani et al. (2017)
  • Devlin et al. (2018)
  • Kaplan et al. (2020)
  • Brown et al. (2020)
  • Bommasani et al. (2021)
  • Hoffmann et al. (2022)
  • Wei et al. (2022)
  • Schaeffer et al. (2023)
  • The DeepSeek-R1 report (2025)
  • Stanford AI Index 2025
  • Epoch AI
  • MIT NANDA (2025)
  • METR (2025)
  • Peng et al.
  • Humlum and Vestergaard (2025)
  • Google DeepMind and the 2024 Nobel Committee on AlphaFold
  • KKR, Goldman Sachs and Nordvig, and SEC filings on capital expenditure
Start a conversation