The strategy paper reached Tsitsikamma's executive committee with a familiar chart on page three: a curve rising to a peak, falling into a trough and climbing slowly to a plateau, with technologies placed along it like beads on a string. AI agents sat near the top. The paper's recommendation followed from the picture. Agents were at the peak of expectations, so a fall was coming; the sensible course was to wait about two years for the plateau and adopt then, when the technology had matured and prices had fallen.
The head of the trade desk read the paper with growing frustration. Her team spent much of each week turning builders' bills of quantities into quotes, line by line, matching descriptions to stock codes. She had seen a narrow, supervised agent do a credible first pass in a demonstration. She was not asking for a transformation programme. She wanted to know whether the thing worked on Tsitsikamma's own quotes, and the paper had answered a different question.
Tsitsikamma Hardware Group is fictional, and its systems, claims and figures are illustrative. It is the building-supplies retailer headquartered in George that has carried this series, with branches across the Garden Route and Eastern Cape and a trade desk that quotes for builders and contractors.
What a hype cycle is
The best-known version of the chart is Gartner's Hype Cycle. Gartner's methodology page describes five phases: the Innovation Trigger, the Peak of Inflated Expectations, the Trough of Disillusionment, the Slope of Enlightenment and the Plateau of Productivity. The vertical axis is expectations; the horizontal axis is time. Two Gartner analysts, Jackie Fenn and Mark Raskino, set out how organisations should use it in Mastering the Hype Cycle (Harvard Business Press, 2008): the aim is to avoid adopting too early or too late, and giving up too soon or hanging on too long.
That is a reasonable purpose. Three features of the chart make it a poor substitute for an adoption plan.
It describes the market's expectations, not a technology's fitness for your work. A placement on the curve is an analyst's judgement of how a broad market feels about a broad category. It says nothing about whether a specific capability works on your data, in your process, at your cost.
Its categories are wide. "AI agents" covers everything from a supervised assistant that fills in one form to an autonomous system that runs a supply chain. Parts of a category can be useful today while other parts are years away. One position on a curve averages them all.
Its path is a model, not a law. Researchers who reviewed the hype cycle model, Dedehayir and Steinert, writing in Technological Forecasting and Social Change in 2016, compared Gartner's reports for three energy technologies with empirical data and found incongruences. The curve is a useful reminder that excitement and value move at different speeds. It is not a timetable.
Two different questions
The strategy paper treated one question as if it answered another.
The market question is: how inflated are general expectations of this technology right now? A hype cycle is one attempt to answer it, and the answer is useful context. High expectations mean vendor claims will be bold, prices may be high, talent will be expensive and products will change quickly.
The adoption question is: what evidence do we have that a specific use of this technology improves a specific piece of our work? No external chart can answer that. Only your own evidence can.
The two combine into four situations. High market expectations with weak evidence of your own call for a bounded experiment, not a programme and not a ban. High expectations with strong evidence call for scaling with care, watching for vendor lock-in and price changes. Low expectations with strong evidence are often the best investments, because nobody else is paying attention. Low expectations with weak evidence call for watching, and nothing more.
Tsitsikamma's paper had looked only at the market-expectations axis. The trade desk's question sat in the first box.
From forecast to adoption plan
A forecast becomes useful when it is converted into a sequence of decisions that each rest on evidence. Tsitsikamma now uses five steps whenever a forecast, analyst chart or trend report shapes an AI proposal.
1. Record the forecast as a claim. "AI agents are at the peak of inflated expectations" goes into the AI claims register from Part 4, worded as stated, with its source and date. At best it is graded Published. It can inform a decision about how cautiously to proceed. It cannot, on its own, support a decision to invest or to refuse.
2. Name the use case and the decision it would change. Not "adopt agents", but "use a supervised agent to draft trade quotes from bills of quantities, so that quoting staff review drafts instead of keying lines". This is the narrow, embedded kind of use that Part 2 argued is where returns appear. It is also timely: with the credit-approval constraint from Part 1 relieved by the summaries described in Part 2, quoting time had become the step that limited completed trade orders.
3. Set the evidence thresholds before the experiment. Decide in advance what result would justify the next step and what result would stop the work. Writing them afterwards invites the thresholds to move to wherever the result lands.
4. Run a bounded experiment. A fixed scope, a fixed budget, a fixed duration, a named owner and a baseline measured before it starts. Part 1 showed what happens when time savings are claimed without a baseline.
5. Decide in stages. Each stage ends with one of three decisions: stop, extend with a changed question, or move to the next stage with a larger scope. The forecast is reviewed at each stage too, because the market will have moved.
The trade desk experiment
The trade desk's experiment was small by design.
- Scope: 120 recent bills of quantities from residential builders, already quoted by staff, so the correct answer was known.
- Baseline: staff time per quote and the share of lines needing correction after the quote was sent, measured over the previous quarter.
- Thresholds, set in advance:
- Stop if fewer than 70% of lines were matched to the correct stock code, or if any draft priced a line at the wrong unit of measure without flagging it.
- Extend if line accuracy was between 70% and 90%, to test whether a better product catalogue fixed the gap.
- Move to a live pilot if accuracy exceeded 90%, every uncertain line was flagged, and review time per quote fell by at least a third in a timed trial.
- Budget and time: six weeks, a fixed model spend and one analyst.
- Owner: the head of the trade desk, with the finance team's AI value lead checking the measures.
The result landed in the middle band. Line accuracy was high for standard items and poor for descriptions that used builders' shorthand, which matched what the workflow-first approach would predict: one uncertain step, not an uncertain path. The decision was to extend for four weeks with a catalogue of common shorthand. The second round crossed the pilot threshold.
The strategy paper's two-year wait would have produced nothing in that time except an older chart. A full programme launched on enthusiasm would have spent more and learnt less, because nobody would have agreed in advance what failure looked like.
When the forecast should change the plan
Market forecasts are not useless. They should shape how an organisation adopts, even when they should not decide whether it does.
- High expectations call for short contracts, exit clauses and suspicion of vendor statistics. Prices and products will change.
- A move into disillusionment is a reason to check whether your own evidence still holds, not a reason to stop work that is producing measured results.
- A move towards the plateau usually means standards, cheaper options and more experienced people. It may be the moment to replace an early custom build with a standard product.
- No position at all, because the use is too narrow or too dull to chart, is often a good sign.
A one-page test for any forecast-driven proposal
Before a proposal citing a hype cycle, trend report or analyst forecast goes to a decision, ask:
- Is the forecast recorded as a claim, with its source, date and grade?
- Which specific use case is proposed, and which decision or piece of work would it change?
- What is the baseline, and who measured it?
- What results would stop, extend or advance the work, and were they written before the experiment?
- What is the scope, budget, duration and owner of the first stage?
- How does the market forecast change the commercial terms, as opposed to the decision itself?
- When will the forecast and the evidence next be reviewed?
Key points
- A hype cycle describes market expectations of a broad category; it does not measure whether a specific use works for you.
- Its path is a model, not a timetable, and a published review found incongruences between Gartner's reports and empirical data.
- Separate the market question from the adoption question, and treat high expectations with weak evidence as a case for a bounded experiment.
- Convert any forecast into a recorded claim, a named use case, thresholds set in advance, a bounded experiment and staged decisions.
- Use the forecast to shape commercial caution and timing of standard products, not to decide whether to learn.
This is the final part of After the Hype. Start again from Part 1, Faster emails are not a business case.
How CloudNala can help
CloudNala helps executive teams turn AI forecasts and trend reports into decisions they can defend. We record the claims behind a proposal, narrow it to a use case with a baseline, design a bounded experiment with stop, extend and advance thresholds agreed in advance, and review the evidence with you at each stage, so that adoption follows your own results rather than the shape of a curve.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an AI Value Review or write to us at consult@cloudnala.co.za