Business & Practice
Where AI actually creates value (and where it doesn't yet)
The technology is rarely the hard part. Capturing real value from AI comes down to choosing the right tasks, redesigning how work flows, and building the habit of using it well.
Most AI disappointment does not come from the AI being weak. It comes from pointing it at the wrong thing, or buying a tool and never really changing how you work.
Where does AI genuinely help today? Tasks that involve drafting, summarising, explaining, brainstorming and analysing text, the everyday knowledge work most of us do. It is brilliant as a first-draft machine and a tireless thinking partner.
Three simple lessons make the difference between hype and results: - **Use it with you, not instead of you. Let it draft; you judge, edit and decide. That partnership beats either working alone. - Pick real problems. Start from a task that actually costs you time, not "we should use AI". - Stay in charge of anything important.** AI makes confident mistakes, so a human checks the things that matter.
And the biggest secret: you get good at AI by using it on your real work, not by reading about it. The value is unlocked by the habit, not the tool.
The evidence and the practitioner wisdom point the same way: value comes from how you apply AI, more than from which model you use.
Start from a real use case
A good use case has real pain or value, plays to AI's strengths (drafting, summarising, analysis), is verifiable, and fits your workflow. The common failure is technology-led adoption, "we need an AI", with no concrete problem, which produces demos that never pay off (ROI).
Augment, don't just automate
There are two ways to apply AI: augmentation (AI helps a person do better) and automation (AI does the whole task). In knowledge work, augmentation often captures more value, human judgement plus AI speed beats either alone. This is the idea of co-intelligence: treat AI as a capable but flawed collaborator, not an oracle or a gimmick.
Mind the "jagged frontier"
AI is unevenly good: strong on some tasks, surprisingly weak on similar-looking ones. Studies show that on tasks within AI's ability, people do better and faster with it, but on tasks beyond it, uncritical reliance makes results worse. So the key skill is judging where AI is trustworthy, which requires AI literacy built through hands-on use.
Keep a human in the loop
For consequential or irreversible decisions, keep a human in the loop to review and approve. This catches the confident errors and keeps responsibility with people.
Adoption beats the model
Two teams with the same tool get very different results depending on adoption, workflow redesign, training, culture. The model is necessary but rarely sufficient; the value comes from changing how work actually happens.
Realising AI value is a socio-technical problem dominated by use-case selection, workflow redesign and human capability, not model choice.
Use-case portfolio, not a single bet
Returns depend on use-case fit and workflow redesign more than raw capability. A disciplined approach runs cheap experiments to discover fit, measures against a baseline, and scales winners, treating AI literacy and change management as prerequisites. The dominant ROI failure is unmeasured, technology-led pilots with underestimated total cost of ownership.
The jagged frontier and reliance dynamics
Capability is uneven and non-obvious across nearby tasks. Controlled studies of knowledge workers found significant gains on tasks inside the frontier and degraded performance when workers deferred to AI on tasks outside it (Dell'Acqua et al. 2023). The practical implication is calibrated task routing, knowing where the model is reliable, which is a learned, verification-backed skill rather than a property of the tool.
Augmentation, automation, and human oversight
The augmentation-vs-automation choice should be made at the sub-task level, weighted by reversibility and stakes: automate bounded, verifiable, high-volume steps behind guardrails; augment judgement-heavy work while keeping a human in the loop. The framing of co-intelligence (Mollick 2024), an active human directing and verifying a capable-but-fallible collaborator, is the operating stance that both captures value and contains hallucination/bias risk.
Adoption as the binding constraint
Because value accrues through changed workflows, adoption, sponsorship, enablement, psychological safety to experiment, process redesign, and trust/governance, is usually the real bottleneck, not model quality. Measure usage and workflow outcomes, not licences purchased. The strategic takeaway: invest in use-case discipline, verification, and people, and treat the model as one (rapidly improving, interchangeable) component of a larger system.