The AI hype cycle: pick use cases by phase, measure real metrics

In the World Economic Forum's Future of Jobs Report 2025, 86 per cent of employers surveyed expect AI and information processing technologies to transform their business by 2030 (World Economic Forum, 2025, p. 62). In the Forum's executive opinion survey, 50 per cent name a lack of skills and 43 per cent a lack of vision among managers and leaders as barriers to AI adoption (World Economic Forum, 2025, p. 63). A CFO handed an AI spending request next quarter is standing between those two numbers. The hype cycle offers a way to frame the questions. This post uses it that way, then joins it to OpenAI's method for finding use cases and a table of metrics by phase.

What is the hype cycle, and what can it claim?

It is a model in which expectations of a new technology spike, collapse, then settle with real productivity; the library records the origin as Fenn and Raskino (2008). Its question is the budget holder's: is this technology at the peak of inflated expectations, in the trough of disillusionment, or on the way to productivity?

The site's framework gives five phases: technology trigger, peak of inflated expectations, trough of disillusionment, slope of enlightenment and plateau of productivity.

Everything I say about the model comes from this summary. I could not read the book, so I do not know how its authors place a technology on the curve or how long each phase lasts, and the library holds no source that tests whether technologies really follow it. I therefore treat it as a model, not a law. The framework itself warns that placing generative AI in the trough usually comes without a source. This post places no AI technology on the curve; it uses the five phase names as five questions about evidence, set out as a table at the end.

Do the 2025 AI numbers measure expectation or results?

Mostly expectation and planned spending. The WEF report says investment flows into AI have grown nearly eightfold since ChatGPT's release in November 2022, while more general adoption stays comparatively low, with only a small fraction of firms using it in 2023, and "the full extent of long-term productivity gains from the technology remains uncertain" (World Economic Forum, 2025, p. 11). OpenAI's guide quotes McKinsey figures: 92 per cent of companies plan to raise AI investment, yet only 1 per cent believe their investments have reached full maturity (OpenAI, 2025, p. 3). Money runs ahead of evidence.

But a series like that does not reveal the phase: the same numbers fit a peak or a slope. The report also notes uneven adoption, with information technology leading and construction lagging (World Economic Forum, 2025, p. 11). So the answer to "where is AI on the curve?" may differ by sector and by task. That is an inference; the report does not say it. What is worth measuring sits in-house: each use case, in your own company.

Where should you look for use cases before talking about investment?

Look at three everyday situations: repetitive manual tasks, skill bottlenecks and moments of ambiguity when it is hard to start (OpenAI, 2025, p. 6). OpenAI (2025), which sells ChatGPT Enterprise, advises each team to list work that fits those three, and to use the lists to open a workshop or hackathon (OpenAI, 2025, p. 9).

The next step is the six types. OpenAI analysed over 600 customer use cases and says most fall into one of six (OpenAI, 2025, p. 10).

Two points are worth keeping for anyone standing at the peak of expectations. OpenAI writes that complex use cases can feel impressive but often slow you down, and that letting employees find the use cases that suit them is often faster (OpenAI, 2025, p. 5). And Estée Lauder's process for building a GPT opens with a two-page brief written by the business user; its director of ChatGPT Enterprise adoption asks three questions: why build it, what problem does it solve, what impact will it have (OpenAI, 2025, p. 24).

How do you prioritise, and why re-score every quarter?

OpenAI uses an impact/effort matrix with four boxes: high-ROI focus to do first, self-service, high-value but high-effort, and high-effort with low impact to set aside (OpenAI, 2025, p. 26). Indeed sits in the high-value, high-effort box: it built an automated way to explain to jobseekers why an opening was recommended, which took months of testing and iteration, and job applications started rose 20 per cent (OpenAI, 2025, p. 27).

The line that matters most for this post is on page 28: re-evaluate the scoring each quarter, because the high-effort use cases of today may become low-effort as AI capabilities advance (OpenAI, 2025, p. 28). I read it alongside the hype cycle. "Effort" and "impact" are both estimates, and estimates made at the peak of expectations tend to be more optimistic than those made after a trial run, so quarterly re-scoring forces estimates to meet data. That pairing is mine; the guide does not make it.

The guide advises starting with quick wins to build momentum (OpenAI, 2025, p. 26). That suits you if the team is new to AI and has no baseline yet. It does not suit you if every quick win is personal convenience and none touches the company's costs or revenue.

Which metric counts as real?

A "real" metric here is one with a baseline: you measure the task before using AI, then measure again in the same unit. The library's concept of the ROI measurement gap describes the opposite: using AI without showing benefit through metrics that have a baseline. That definition belongs to the site and carries no source.

Two examples in OpenAI's guide show the distance. Promega saved 135 hours in its first six months using ChatGPT Enterprise to draft campaign emails (OpenAI, 2025, p. 13). The text does not say how this was measured or what the baseline was, and it is an input metric (time). Indeed's 20 per cent is an output metric (applications started), also without a measurement period or baseline (OpenAI, 2025, p. 27). Both give a sense of scale, but neither is enough to defend a budget line.

A hypothetical example shows what a baseline looks like. Company A, a food distributor, has six sales staff answering price-quote requests; the company and every number below are invented by me. In the two weeks before the trial they logged 120 quotes, an average of 40 minutes each, and 9 quotes with a wrong unit price or term (7.5 per cent). In a two-week trial with AI drafting and a person reviewing: 130 quotes, 25 minutes each including review, and 6 wrong quotes (4.6 per cent).

Saving 15 minutes across 130 quotes is 32.5 hours in two weeks, and you can trust it because both ends were measured. "Fewer errors" is another matter: nine against six in two small samples proves nothing. Company A should run a few more weeks, and subtract licence costs from the saving before presenting it.

What should you ask and measure at each phase?

Choose the row by the evidence you hold, not by where the press says the technology sits: with no baseline, you are on row two. The table is an inference from the five phase names and the figures above.

Two limits apply. The table needs you to pick a row, by the technology's position or by internal evidence; the original model does not promise you can locate that position. And a phase belongs to a technology while a decision belongs to a use case. Suppose one tool: drafting emails already has a baseline and repeats across several teams (row four), while revenue forecasting has never been measured (row two). For the narrower question of whether one specific task should go to AI at all, the post which tasks to hand to AI has a matrix of its own.

Going deeper: what did I read, and what did I not?

OpenAI's guide (2025): I read the full 34-page PDF through automated text conversion. The customer examples and figures are OpenAI's own and I did not check them one by one. The BCG and McKinsey figures in the foreword are quoted second-hand and I did not open the originals.

The WEF report (2025): I read pages 11, 62 and 63, which the post cites, and scanned the full text for mentions of AI; I have not read all 290 pages. Figure 4.13 on page 63 is labelled "share of employers" but its source is the Executive Opinion Survey, so I wrote "executives".

Fenn and Raskino (2008): I did not read the book. The five phases and the warning about generative AI come from the library's framework summary and definition; I do not know how the authors define or measure each phase, or what empirical evidence they offer. The remark that this is a model and not a tested law is a limit I state because the library has no source that tests it, not a conclusion from a critique I have read.