A recent Harvard Business Review article has given a name to something many organisations have been feeling but struggling to put into words: the “GenAI wall.” The argument is deceptively simple. Generative AI now lets almost anyone produce work that looks expert — a credible strategy memo, a passable contract clause, a plausible first diagnosis. But the moment that work has to be executed, judged and refined under real conditions, the distance between an expert and a non-expert quietly reopens. There is still a wall. There is still a delta. There is still an expertise gap.
We think the article is right. We also think it stops one or two steps short of the conclusion that matters most to anyone actually redesigning how their organisation works.
What the GenAI Wall Actually Tells Us
The HBR piece corrects a familiar fantasy — that generative AI quietly dissolves the difference between people who know what they are doing and people who don’t. It doesn’t. AI lowers the cost of entering the conversation. It does not hand you the domain expertise required to finish the work, defend it, or hold your own in the human discussions that surround it.
But “experts versus non-experts” is too blunt a framing to design around. In practice the wall runs along several dimensions at once:
- Experts, adjacent experts and distant non-experts — proximity to the domain changes everything.
- Narrow, verifiable tasks versus open-ended expert work — the wall is low for the first and high for the second.
- Execution versus supervision and review — doing the work and checking the work are different jobs.
- Specification versus deep execution — describing what “good” looks like is not the same as producing it.
- Individual tasks versus end-to-end workflows — a fast task does not make a fast process.
- Quality per output versus total throughput — more output is not automatically more value.
- Automation versus organisational redesign — the largest gains live in the second, not the first.
Generative AI can almost close the gap in conceptual work. In deep execution, the gap holds. That single distinction is where most of the practical consequences live.
The First Dimension the Article Misses: Quantity
The first thing we would add is some uncomfortable arithmetic. When experts use AI well, they can become five or ten times more productive on specific tasks. Encouraging. But when non-experts use the same tools, they tend to produce five or ten times more work that is almost good — and someone still has to tell the difference.
That someone is usually the expert. So the productivity does not disappear; it moves. The bottleneck shifts from production to supervision, and from supervision to quality control. An organisation that celebrates a tenfold jump in output without noticing the tenfold jump in review has not become faster. It has simply relocated the queue.
The Second Dimension: Which Tasks Hit the Wall
The second addition is that the wall is not the same height everywhere. For narrow, structured and verifiable tasks — classification, triage, routing, document extraction, matching — it is remarkably low. These are precisely the tasks where expertise can be externalised into rules and evaluation criteria rather than held in someone’s head.
This is also where agentic systems earn their keep. A well-designed system decomposes work into smaller steps, automates the parts that are narrow and verifiable, and escalates the parts that genuinely require expert judgement. The point is not to replace the expert, but to stop spending the expert on work that never needed them.
From Tasks to Operating Models
This is where the conversation stops being about tools and starts being about the organisation. Because once you redesign a process around AI, you are quietly redesigning the organisation around it too.
Tasks become workflows. Workflows become processes. Processes define roles. Roles define the organisation. The real productivity gain does not come from bolting AI onto yesterday’s operating model — it comes from designing an AI-native one, in which you have deliberately answered:
- what can be safely automated,
- what requires human review,
- which review can be handled by juniors or non-experts,
- where adjacent expertise is sufficient,
- and where true domain expertise is still non-negotiable.
Seen this way, the GenAI wall is not an obstruction. It is a design constraint — a map of where to redesign workflows, where to keep humans firmly in control, and where expertise still has to sit.
Conclusion
Generative AI cannot manufacture taste, judgement or human sophistication. What it can manufacture, in enormous quantities, is leverage — and for the overwhelming majority of business work, leverage is the thing that actually matters.
The organisations that pull ahead will not be the ones that hit the wall and conclude the technology was overhyped. They will be the ones that read the wall as a blueprint: redesigning work around it, deciding where machines lead and where people must, and building operating models that know the difference.
The original article
The Harvard Business Review article that prompted this is well worth reading in full — we would simply add quantity and task design to its picture. Read the original article
If you are redesigning how your organisation works with AI, this is exactly the kind of problem we like to put a few complementary experts around a table to solve.
.avif)
