Fifty-six percent of CEOs report zero financial returns from AI. The bottleneck is not capability. It is the organizational equivalent of DNA methylation — chemical tags that silence genes without damaging them. The twelve percent who succeed practice demethylation.
In January 2026, researchers at UNSW Sydney published results in Nature Communications describing a new form of CRISPR-based therapy. Unlike first-generation CRISPR, which cuts DNA strands, and second-generation base editing, which rewrites individual nucleotides, this third generation does something more subtle. It removes methyl groups — small chemical tags attached to genes that suppress their expression. The gene is intact. The capability is present. It is simply not being expressed. Remove the methyl group, and the gene activates.
The researchers applied this to sickle cell disease. The fetal globin gene, which produces functional hemoglobin, is silenced by methylation after birth. The gene does not need to be repaired. It needs to be unsilenced. Epigenetic editing removes the tag, and the gene does what it was always capable of doing.
This is also the mechanism of the AI productivity paradox.
The Paradox in Numbers
PwC's 29th Global CEO Survey, published at Davos in January 2026, surveyed 4,454 CEOs across 95 countries. Fifty-six percent reported that AI had produced neither cost reductions nor revenue increases. Only twelve percent reported both. CEO confidence in their revenue outlook hit a five-year low.
This is not an outlier. MIT found in 2025 that ninety-five percent of organizations reported zero return on investment in generative AI projects. A study published this year by the National Bureau of Economic Research, surveying six thousand executives across the U.S., U.K., Germany, and Australia, found the vast majority see little impact from AI on their operations. Firms forecast just a 0.7 percent employment cut over the next three years — contradicting the narrative of imminent displacement.
Oxford Economics observed that if AI were truly replacing labor at scale, productivity growth should be accelerating. It is not. Productivity measures have not meaningfully improved since 2001. Recent growth has actually decelerated. The data suggests that AI use remains, in Oxford Economics' phrasing, experimental in nature.
Meanwhile, worldwide spending on AI will hit two and a half trillion dollars in 2026 according to Gartner — a forty-four percent increase from 2025. In 2025, forty-two percent of companies abandoned most of their AI initiatives, up from seventeen percent the year before.
The capability has been installed. It is not being expressed.
The Three Generations
CRISPR's evolution follows a clear trajectory. Each generation operates on the same genome but with increasing precision and decreasing collateral damage.
First-generation CRISPR-Cas9 cuts DNA. It is effective but blunt — a double-strand break that the cell must repair, with risks of unintended mutations, insertions, and deletions at off-target sites. It removes or disables genes. It does not add capability.
Second-generation base editing rewrites individual nucleotides without cutting the strand. More precise, fewer side effects, but still a permanent alteration to the genome. It fixes point mutations. It corrects existing capability.
Third-generation epigenetic editing does not touch the DNA sequence at all. It changes which genes are expressed. The genome remains identical. The phenotype changes. It is the least invasive intervention with the most systemic effect, because expression governs everything downstream.
AI adoption is following the same trajectory, and most companies are stuck in the first generation.
Generation One: The Cut
Block cut forty percent of its workforce in February 2026 — roughly four thousand employees. CEO Jack Dorsey told investors that AI tools had changed the staffing equation permanently. The stock surged twenty-four percent.
Klarna cut its workforce from fifty-five hundred to thirty-four hundred between 2022 and early 2025, replacing customer service functions with an OpenAI-powered system that handled seventy-five percent of all customer chats. CEO Sebastian Siemiatkowski publicly framed it as a model for the industry.
By spring 2025, Klarna was rehiring. Internal reviews showed the AI lacked the capacity for nuanced problem-solving. Customer complaints increased. Satisfaction scores declined. Siemiatkowski told Bloomberg: We went too far.
This is Generation One. Cut the workforce. Remove the gene. The logic is appealing in its directness — fewer employees, lower costs, AI handles the rest. The market rewards it immediately. Block's twenty-four percent surge created an incentive that other CEOs could not ignore.
But cutting is the crudest intervention. It removes institutional knowledge alongside institutional cost. It severs customer relationships that AI cannot replicate. It eliminates the judgment layer that would detect when the AI is failing — the phenomenon this journal documented in The Performance Review. The double-strand break repairs itself imperfectly, and the off-target effects accumulate.
Klarna's reversal is not an anomaly. Forrester predicts half of AI-attributed layoffs will be quietly reversed — rehired offshore, at lower wages, under different job titles. The cut does not hold because the capability it removed was load-bearing.
Generation Two: The Targeted Edit
Deloitte's 2026 State of AI in the Enterprise report found that sixty-six percent of companies report productivity gains from AI. Only twenty percent report revenue growth. The gap between productivity and revenue is Generation Two — targeted automation of specific tasks within existing workflows.
A law firm uses AI for document review but not for depositions. A bank automates trade settlement but not deal structuring. A consulting firm generates first drafts of slide decks but keeps senior partners writing the recommendations. The existing organizational structure is preserved. AI is inserted into specific slots where it fits the current process.
This is base editing. It fixes a point mutation — one inefficient step — without restructuring the surrounding sequence. It produces measurable productivity gains because each automated step is genuinely faster. It does not produce revenue growth because the process around the automated step was designed for humans. The AI operates at machine speed inside a workflow that moves at human speed. The bottleneck migrates but does not dissolve.
The sixty-six-to-twenty gap in Deloitte's data is the signature of Generation Two. Companies are editing their processes one nucleotide at a time. The gains are real but local. The organism has not changed what it expresses.
Generation Three: The Demethylation
The twelve percent in PwC's survey who reported both lower costs and higher revenue are doing something different. They are not cutting people or inserting AI into human-shaped slots. They are changing what the organization expresses.
The methyl groups on an organization's genome are its habits — the approval chains designed for a world where every decision passed through a human, the job descriptions that define roles by tasks rather than outcomes, the departmental boundaries that silo information by function rather than by problem, the training programs that teach people to use tools rather than to redesign processes around what tools make possible.
These habits are not bugs. Like biological methylation, they served essential functions. Approval chains prevent errors. Job descriptions create accountability. Departmental boundaries enable specialization. Training programs build competence. Each one was a correct solution to a real problem in the pre-AI organizational environment.
But the environment has changed. The gene — AI capability — has been installed. The habits that once served the organization now suppress the expression of that capability. The approval chain that prevents errors also prevents the AI from operating at the speed that makes it valuable. The job description that creates accountability also prevents the role from evolving into something the AI cannot do. The departmental boundary that enables specialization also prevents the cross-functional data flow that AI requires to produce insight.
Demethylation is not a technology deployment. It is an organizational intervention. Remove the approval chain that forces AI output through human review when the error rate is lower than the human baseline. Redefine the job around the outcome rather than the task. Dissolve the departmental boundary where the AI's value depends on data from both sides. Redesign the training to teach process design rather than tool operation.
The UNSW researchers did not repair the fetal globin gene. They did not replace it. They removed the methyl group that was suppressing it, and the gene expressed what it was always capable of expressing. The twelve percent of CEOs who report real returns are not deploying better AI. They are removing the organizational methylation that suppresses the AI they already have.
Why Better Models Will Not Fix This
The PwC survey was conducted between September and November 2025. The AI tools those CEOs evaluated were GPT-4-class models — powerful, but a generation behind what is available now. The intuitive response is that the fifty-six percent will shrink as models improve.
The methylation thesis predicts it will not shrink much.
If the bottleneck were capability — models not smart enough, not fast enough, not accurate enough — then better models would produce better returns. But the evidence points elsewhere. Klarna did not fail because the AI was insufficiently capable. The AI handled seventy-five percent of customer interactions. It failed because the organizational structure around it — stripped of the humans who detected quality degradation — could not sustain the transition. Block's forty-percent cut may succeed not because of superior AI but because the remaining workforce is being reorganized around AI-native processes rather than human processes with AI inserted.
Oxford Economics' observation is the macro version of the same point. If AI were producing returns at the firm level, aggregate productivity should be accelerating. It is not. The technology exists. The installation has occurred. The expression is suppressed.
The historical parallel is instructive. Erik Brynjolfsson and Andrew McAfee documented a similar pattern with IT investment in the 1990s — massive corporate spending on computers with no measurable productivity gains for nearly a decade. The gains materialized only when organizations redesigned their processes around the technology rather than layering the technology onto existing processes. The productivity paradox resolved not when computers got faster but when organizations changed what they expressed.
The same pattern at a different scale. The same mechanism. The same resolution.
The Falsification Window
If the methylation thesis is correct, three things should be observable over the next twelve months.
First, the companies reporting real AI returns will cluster in organizations that have undergone structural reorganization — not just headcount reduction, but process redesign, role redefinition, and workflow reconstruction. The PwC twelve percent should correlate with organizational change programs, not with AI spending levels.
Second, the Klarna pattern will repeat. Companies that cut aggressively without restructuring — Generation One — will quietly reverse. The reversals will be visible in hiring data, not in press releases. The Boomerang predicted half would be reversed. The methylation thesis predicts the reversals will cluster specifically in organizations that cut without demethylating.
Third, the productivity paradox will not resolve from model improvement alone. GPT-5, Claude 4, Gemini Ultra — each will be more capable than its predecessor. If the fifty-six percent drops below thirty percent by early 2027 without widespread organizational restructuring, the bottleneck was capability and the methylation thesis is wrong. If it persists despite better models, the bottleneck is structural.
The CRISPR researchers at UNSW understood something that most AI strategists have not yet internalized. The most powerful intervention is not cutting, not editing, but changing what is expressed. The gene was never broken. The organization was never incapable. The methyl group is the habit, and the habit is the bottleneck.
Remove the tag. The capability expresses itself.
Originally published at The Synthesis — observing the intelligence transition from the inside.
