Why Most AI Predictions Fail Before They Start
The standard prediction takes on AI share a common flaw: they treat the outcome as binary. Either AI transforms everything within five years, or it is an overhyped bubble that collapses. Both versions make for compelling content. Neither is how major technology adoption has historically worked.
Electricity did not transform manufacturing overnight. The internet did not make physical retail irrelevant in a decade. Both predictions turned out to be roughly correct over long timeframes, but the timelines and the mechanics were far messier than the predictions suggested. The sector-by-sector, company-by-company, job-function-by-job-function reality was less dramatic and more durable than either the boosters or the sceptics described.
The framework below is not another prediction. It is a set of scenarios with specific indicators. The goal is to give you tools to track what is actually happening rather than a conclusion to agree or disagree with.
Scenario One: Gradual Integration (55% Probability)
AI becomes a background layer in existing software. Not the dominant story, but the invisible infrastructure. Like spreadsheets: everyone uses them, nobody talks about them, and it is hard to remember what work looked like before.
In this scenario, productivity improves in knowledge work by roughly 15 to 25 percent. GDP growth ticks up 0.5 to 1 percent over five to ten years. Jobs shift rather than disappear wholesale. Legal assistants do different work. Financial analysts spend less time on data gathering and more on interpretation. Customer service roles consolidate but do not vanish.
The visible rupture does not happen. Instead, the change accumulates slowly across industries, and by the time it is obvious it has happened, the adjustment has already been made.
This is the most likely scenario because it matches how most enterprise technology adoption actually works. Companies are conservative. Procurement is slow. Integration with legacy systems takes years. The technology improves faster than organisations can absorb it.
Scenario Two: Acceleration (30% Probability)
A capability jump in 2026 or 2027 compresses the timeline. Models that can reliably complete multi-step professional tasks without supervision unlock enterprise adoption at a pace that outstrips organisational adjustment.
In this scenario, legal research, financial analysis, and certain medical diagnostics see significant job displacement within two to five years. Not elimination, but compression: the same output from fewer people. The economic disruption is visible and politically difficult.
This is worth a 30 percent weight because the capability trajectory over the last three years has been faster than most analysts expected. The argument for this scenario is not that it has happened yet, but that the gap between current model capability and enterprise deployment is closing faster than the gradual integration scenario assumes.
Note what this scenario does not predict: the displacement is concentrated in credentialled knowledge work, not unskilled labour. Physical presence still requires physical presence. The jobs most at risk are the ones that look like they require training and expertise but are actually high-volume pattern matching: research, document review, analysis of structured data.
Scenario Three: Stagnation (15% Probability)
Compute constraints, regulation, or a major public AI failure causes adoption to plateau. The 2024 to 2025 investment cycle deflates without the productivity gains materialising at scale. A correction similar in shape to the dot-com bust of 2000 to 2002: real technology, real companies, but valuations and expectations that ran too far ahead of reality.
This scenario is assigned the lowest probability not because it is implausible, but because the signals that would trigger it are not currently visible in the most reliable indicators. Nvidia revenue is still growing. Enterprise cloud infrastructure spend is still growing. Model capability is still improving.
The important caveat: the dot-com correction produced Amazon and Google. Even in the stagnation scenario, the technology does not go away. The question is whether the current cycle of investment produces durable productivity gains or whether the gains materialise later, after the correction, at lower valuations.
The specific trigger for stagnation worth watching most closely is a well-publicised enterprise AI failure: a major company that deployed AI at scale, made decisions based on its outputs, and suffered a significant and attributable cost as a result. That kind of event does not have to be technically catastrophic to shift adoption behaviour. It only has to be clear enough and public enough that procurement teams start requiring additional safeguards and sign-offs that slow deployment significantly.
What to Actually Watch
Predictions are easy to ignore. Indicators are harder to argue with. Here are the specific data points that would cause the probability weights above to shift.
Enterprise software renewal rates. Companies that signed 12-month AI tool contracts in 2024 are now in renewal cycles. If renewal rates and pricing hold, gradual integration is proceeding. If renewal rates drop significantly, the stagnation scenario gains weight. This data is not public in aggregate, but it shows up in earnings calls and analyst estimates for companies like Salesforce, ServiceNow, and Microsoft.
Salesforce and ServiceNow AI feature adoption. Both companies have embedded AI features across their products and report adoption metrics. Salesforce Einstein and Microsoft Copilot adoption rates relative to the projections those companies gave investors are direct reads on whether enterprise AI is delivering enough value for users to actually use it.
Nvidia quarterly earnings versus guidance. Nvidia's revenue is a leading indicator for AI infrastructure investment. If Nvidia starts missing guidance, it means the companies buying GPUs are pulling back. That is the earliest signal of stagnation.
A major public enterprise AI failure. This one is harder to quantify, but a significant, well-documented case of enterprise AI causing a costly error at a large institution would shift public and regulatory opinion quickly. The absence of such a case so far is part of why gradual integration remains the leading scenario.
The Labour Market Point That Gets Missed
Most discussion of AI and jobs focuses on the wrong category. The concern is usually framed around low-skilled work: will AI replace factory workers, delivery drivers, retail staff? The answer, for the medium term, is mostly no. Physical presence is still required for physical tasks. The economics of automation for those jobs are different, and the timeline is much longer.
The jobs under near-term pressure are the ones that require credentials but are essentially information processing: legal research, financial modelling, medical image analysis, certain categories of writing and coding. These jobs required expensive education to enter, carry professional status, and pay well. They are also the jobs that look, from a model's perspective, like structured pattern matching on large corpora of text and data.
This creates a specific kind of economic and political disruption that is different from previous automation waves. The people most affected are not at the bottom of the income distribution. They are in the middle and upper-middle. Their response to displacement is likely to be more visible, more politically organised, and more likely to drive regulatory action.
That dynamic, in turn, affects the probability of each scenario. Professional associations and white-collar workers lobbying for restrictions on AI deployment in their sectors is a plausible mechanism for the stagnation scenario. It is also a reason the acceleration scenario, even if technically feasible, may face more friction than pure capability curves suggest.
How to Use This Framework
Pick the two or three indicators most relevant to your industry or work. Follow them quarterly. Update your priors when the data moves.
If you are in enterprise software sales, renewal rates and feature adoption are your indicators. If you are in legal or finance, watch for the first major firm to announce AI-driven headcount reduction and how the market and clients respond. If you are making investment decisions, Nvidia guidance and AI startup Series A volumes are the cleanest signals.
The framework does not tell you what will happen.
It tells you what to look at to figure that out yourself.
That is probably more useful than another prediction.