Sam Altman says he misjudged how quickly artificial intelligence would disrupt the economy. After GPT-4 arrived in 2023, the OpenAI chief executive expected established software businesses to become vulnerable much sooner. Instead, people continued buying from familiar suppliers and doing familiar work in familiar ways.
His explanation was economic inertia. In an interview released by David Senra on 23 August, Altman said the AI industry had been too ambitious about adoption timelines and that changing behaviour was harder than technologists realised. He regards the delay as partly positive because it may make the transition smoother.
That is already being retold as evidence that “AI is slowing down”. It is a misleading compression. Altman did not say model capability had stalled. The episode summary states the opposite: he expects capability to advance faster than society and the economy can absorb it.
The more interesting conclusion is that most organisations have not learned to keep pace. Some humans can. The difference is not superhuman typing speed or indiscriminate automation. It is a willingness to redesign work around what AI can now do, while preserving human judgement where consequence demands it.
The bottleneck is encoded in the working day
Altman offered an unusually revealing example. Despite running one of the world’s leading AI companies, he still handles email and moves information between applications in ways that resemble his working habits of the past 20 years. Better tools are available, but the old activity still feels like work.
That pattern extends beyond individual preference. Organisations have procurement cycles, data boundaries, software contracts, reporting lines, professional identities and informal conventions. A technically superior route may be available while the surrounding institution continues rewarding the old one.
This explains why buying an AI licence rarely produces transformation by itself. The new tool is inserted into the existing process, usually near the beginning as a drafting assistant or near the end as a summariser. The underlying decisions, handovers and duplicated checks remain intact. People then conclude that AI has delivered a modest productivity improvement when they have tested only a faster version of yesterday’s workflow.
AI is moving at three different speeds
It helps to separate three kinds of progress that are often collapsed into one story.
Capability speed is the rate at which models gain the ability to reason, generate, analyse, use tools and complete longer sequences of work. That frontier is still moving quickly, even if individual benchmarks or releases do not improve evenly.
Product speed is the rate at which those capabilities become reliable and comprehensible tools. Altman said AI has not yet had its “iPhone moment”: the technological components exist, but the interface and product idea have not fully changed how people interact with computing. He called that largely a product failure.
Organisational speed is the rate at which institutions change processes, permissions, responsibilities and habits. It is normally the slowest layer. It also determines whether capability creates dependable value or merely produces more content inside an unchanged system.
This three-speed problem makes sweeping claims about an AI slowdown unhelpful. A model may improve while adoption stalls. A useful product may spread while governance remains immature. A laboratory may deliberately pace a particular training run for security reasons while the wider capability frontier continues moving.
Keeping pace means changing the unit of work
A human does not keep pace with AI by attempting to think or type at machine speed. The practical shift is to stop performing every intermediate action personally.
Start with the outcome. Let AI research, compare, inspect, draft, test or monitor bounded parts of the work. Run several strands concurrently where their evidence and authority can remain separate. Keep the consequential decisions human: what is true enough to publish, what risk is acceptable, which source is authoritative, what should be changed and when the work is complete.
This is more demanding than asking a chatbot occasional questions. It requires clear instructions, usable context, visible working state and disciplined review. It also requires people to abandon the reassuring fiction that touching every step themselves is the same as controlling the outcome.
Control comes from boundaries, evidence and intervention. A capable operator should be able to see what the AI used, distinguish observation from inference, challenge the result and stop or redirect the work. Our analysis of why human oversight must be a workflow explains the same principle: a person needs timely evidence and real authority, not a ceremonial place at the end.
Inertia can help, but it is not governance
Altman is right that slower adoption can soften disruption. Time can allow organisations to evaluate systems, prepare staff, protect sensitive information and establish accountability. The danger is assuming that delay automatically performs those functions.
Many institutions are not deliberately pacing adoption while building controls. They are running disconnected pilots, preserving old processes and postponing decisions about ownership. When adoption eventually accelerates, they may discover that the time created by inertia was never used to prepare.
A responsible organisation should convert experimentation into an operating decision. That means identifying the purpose, owner, evidence, data boundary, human checkpoints, monitoring signals and events that force reassessment. The handover from AI pilot to production is where promising capability becomes either dependable infrastructure or an unmanaged dependency.
It also needs an operational memory. As capabilities, suppliers and workflows change, leaders should be able to reconstruct what was approved and why. That is the case for AI governance decision logs: policy sets a boundary, while a decision record shows how it was applied to real work.
A practical test for keeping pace
Leaders should stop measuring AI readiness through licences purchased, prompts submitted or pilots announced. Ask instead:
- Which recurring task has been redesigned around current capability rather than merely accelerated?
- What work is still manual only because the old sequence feels familiar?
- Which decisions must remain human, and what evidence does that person need?
- Can several AI-assisted activities proceed safely without one person supervising every keystroke?
- What has been learned and converted into a repeatable operating method?
The organisations that answer those questions well will move differently. They will not automate everything, and they will not wait for a perfect product to remove every uncertainty. They will develop the ability to redistribute work as capability changes, while making accountability more visible rather than less.
What to watch next
The immediate question is whether Altman’s admission changes how AI providers design products. Better models alone will not resolve organisational inertia. Useful systems will need to fit real decisions, expose evidence, carry context across tasks and give people intelligible control over longer-running work.
For everyone else, the test is more personal. The humans who keep pace with AI will not be those who try to work as quickly as a machine. They will be those willing to reconsider, repeatedly, what their own work is for.
Sources
- David Senra: Sam Altman interview and episode summary
- The Next Web: Sam Altman says AI still has not had its iPhone moment
- CTech: Sam Altman says he underestimated how slowly the economy would embrace AI
- TechCrunch: Sam Altman and AI’s separate development-pacing debate
How we work: Artificially Confident articles are source-led, AI-assisted and editorially reviewed.