Halfway through 2026, the evidence base on agentic AI has never been richer, or more scattered. The season’s research output is extraordinary: two landmark studies on trust and data foundations from one of the world’s leading strategy houses; an eighth-annual enterprise survey spanning more than 3,200 senior leaders across 24 countries; quarterly pulse research covering 2,100 executives in 20 markets; a joint study between a global consultancy and one of America’s great business schools; a workplace analysis built on trillions of productivity telemetry signals and a 20,000-worker survey; a CEO radar drawing on 640 chief executives; a decade-long annual predictions programme; a labour barometer built from over a billion job advertisements; and, anchoring it all empirically, Stanford HAI’s ninth AI Index and the second International AI Safety Report, the hundred-expert, thirty-nation scientific assessment chaired by Yoshua Bengio of Mila.
And running beneath all of it, the people who actually built this technology (Hinton, Bengio, Hassabis, Amodei, Fei-Fei Li) spent the first half of 2026 saying things that ought to reorder every one of those boardroom decks.
Individually, each study is a snapshot. Read together, and read against what the scientists themselves are telling us, they form something closer to a stereoscopic image of where the agentic era actually stands: its capability curve, its money, its stalled pilots, its reshaped workforce, its anxious CEOs, its widening trust gap, and the deep scientific uncertainty under the whole enterprise. I have spent the past weeks reading them side by side, the way I once debugged systems: not trusting any single instrument, looking for the signal that survives every sensor.
Here is the signal. Capability is compounding faster than anyone’s planning cycle, and the researchers closest to the frontier believe the compounding is about to accelerate, not level off. Money is following capability at historic scale. Value is real but radically concentrated in a small class of disciplined operators. Work is being redesigned around agents faster than organisations are being redesigned around work. The CEO has personally taken the wheel. The scientific establishment is telling us, with unusual unanimity, that the connective tissue of the whole structure (measurement, oversight, alignment, accountability) is its thinnest part. And the geography of conviction is splitting East from West in ways that will echo for a decade.
This is the collision I have been writing about since Genesis: twenty-fourth-century technology crashing into twentieth-century governance. In 2026, for the first time, we have the full instrument panel showing the collision in real time. Let’s read it, gauge by gauge.
- The capability curve: a jagged, accelerating frontier
Start where every honest analysis must: with what the technology can actually do. The 2026 AI Index, the closest thing the field has to an independent census, is unambiguous that progress has not plateaued. On the benchmark where models resolve real software engineering issues from live repositories, performance leapt from 60% to near-saturation in a single year. Frontier models now meet or exceed human baselines on PhD-level science questions, competition mathematics, and multimodal reasoning. Agents improved from roughly 12% to 66% accuracy on structured computer-use tasks in one cycle. Models gained thirty percentage points in a year on an evaluation built specifically to resist them. Tests designed to stay challenging for years are saturating in months.
But the frontier is jagged, and this is the finding every leader needs to internalise before signing another agentic cheque. The same generation of models that won International Mathematical Olympiad gold read analog clocks correctly barely half the time, against 90% for humans. Robots still fail nearly nine in ten real household tasks. Demis Hassabis, Nobel laureate and co-founder of one of the two leading frontier labs, has made “jagged intelligence” his signature caution: systems that outperform professors in one domain while failing tasks a twelve-year-old handles easily. At Davos in January he pushed back explicitly against the triumphalist reading, insisting current systems remain nowhere near human-level general intelligence, and pointing to continual learning, robust world understanding, and genuine creativity as the open gaps.
That jaggedness is not a temporary embarrassment. It is a structural property of how these systems learn, and it carries a direct enterprise consequence: benchmark performance is a necessary but wildly insufficient predictor of deployment reliability. The Index’s authors say this plainly. The new benchmarks built to challenge frontier models increasingly measure performance in clean, data-rich conditions that bear little resemblance to a messy enterprise workflow. The gap between the demo and the deployment is not a procurement problem. It is a science problem, still unsolved.
Meanwhile, the competitive picture has fundamentally changed. The Index’s data shows the US–China model performance gap has effectively closed, the lead trading hands repeatedly since early 2025, with the top models from six labs, American and Chinese, now clustered within a whisker of each other on public leaderboards. When raw capability converges, competition shifts to cost, reliability, integration, and distribution. That is precisely where enterprises live, and precisely where China’s declared diffusion strategy is aimed: 70% agent penetration of its real economy by 2027, and 90% by 2030. The frontier race is becoming a deployment race, and the deployment race is a governance race in disguise.
Two structural warnings sit underneath the capability story. Transparency is declining: the Index’s transparency scoring of major model developers fell from 58 to 40, with the most capable models disclosing the least about training data, compute, and risk. And the Index’s own framing question is the one I explored in The Measurement Gap last week. The distance between what AI can do and what our institutions can evaluate, govern, and absorb is widening, not closing. Hold that thought; every section below returns to it. - What the builders believe: timelines, adolescence, and the closing loop
The industry studies measure the enterprise; the researchers are telling us about the trajectory. And in early 2026, the most instructive conversation in the field was the one between Dario Amodei and Demis Hassabis, the leaders of the two most consequential frontier labs, on a shared stage, revisiting the timeline question.
Amodei has been the most aggressive credible forecaster: powerful AI, meaning systems broadly better than humans at almost all cognitive work, arriving as early as 2026 or 2027. It is a projection his company has been willing to put on the regulatory record, not just the podcast circuit. Pressed this year on whether he stood by it, he did not retreat. The mechanism he described is the one that should concentrate every strategist’s mind: models good at coding and AI research are now being used to build the next generation of models, a self-reinforcing loop. His own engineers, he noted, increasingly edit and direct code the models write rather than writing it themselves, and he estimated the models could be doing most of that engineering work end-to-end within six to twelve months. Whether or not the schedule slips, the shape of the claim matters. Readers of this blog will recognise it instantly: it is the prompts-to-loops thesis reaching its logical terminus. The loop has closed around its own creation, and closed loops compound.
Hassabis places the same destination further out, at roughly even odds by the end of the decade, and his emphasis differs in a way leaders should study. Where Amodei stresses acceleration, Hassabis stresses coordination. He has floated an international scientific collaboration, on the model of the great physics consortia, for the final steps toward general intelligence, precisely because he believes no single lab or nation should navigate that transition alone. Two of the most informed people on Earth disagree by perhaps five years on timing and agree completely on the stakes.
Amodei’s framing of those stakes deserves its own paragraph, because it converges, from inside a frontier lab, on the argument I made in January’s The Adolescence of AI. He spent his holiday writing an essay structured around the question posed to the alien intelligence in Contact: how did your civilisation survive its technological adolescence without destroying itself? That is the correct question, and it is a governance question, not a capabilities one. At Davos he compressed the economic version into a single warning: that AI would produce “very high GDP growth and potentially also very high unemployment and inequality.” Abundance and disruption arriving on the same curve, with the ordering determined by institutional choices. Short-term turbulence for long-term abundance is not a tagline. It is the explicit forecast of the people building the systems.
The revenue curves back the capability claims. By Amodei’s own telling, his lab’s revenue has grown roughly tenfold across three consecutive years, from tens of millions to billions, a commercial trajectory tracking the cognitive one. Analysis presented at this year’s World Economic Forum projected global AI spending on course for US$1.5 trillion annually in applications plus US$400 billion in infrastructure by 2030. Whatever one’s view of individual timelines, the capital markets have already priced in the builders’ story, not the sceptics’. - The conscience of the field: Hinton, Bengio, and the hundred-expert consensus
If the lab leaders supply the trajectory, the elder scientists supply the conscience, and in 2026 their message sharpened considerably.
Geoffrey Hinton, Nobel laureate and the field’s reluctant godfather, has compressed his AGI estimate from thirty-to-fifty years down to something like five to twenty, and continues to attach a 10 to 20 percent probability to catastrophic outcomes. But his most important 2026 contribution is conceptual, not probabilistic. Control-based safety, he argues, cannot survive the intelligence gap: sufficiently capable systems will route around constraints imposed by less capable overseers. The only natural precedent for a smarter entity reliably serving a less-smart one, in his telling, is a mother and her infant. Care built into the architecture, not bolted on as a rule. “We need AI mothers rather than AI assistants,” as he put it; an assistant can be dismissed, a mother’s commitment cannot. One may find the metaphor moving or implausible. Fei-Fei Li, notably, has pushed back, arguing the goal should be AI that honours human dignity and agency rather than AI that parents us, and philosophers have objected that machines lack the mechanisms of care entirely. But the structural insight underneath survives every objection: alignment by external constraint has a ceiling, and the field does not yet know how to build alignment by internal disposition. For enterprise leaders, translate it thus: your guardrails are necessary and known to be insufficient by the people who invented the underlying technology. This is why my AI Safety Stack insists that safety is layered. No single mechanism, from model alignment to runtime control to institutional oversight, bears the load alone.
Yoshua Bengio, Turing laureate and founder of Mila, the world’s largest academic deep-learning institute, has gone furthest of all: from warning to institution-building. His nonprofit safety lab, launched last year with US$30 million, exists to build non-agentic AI. These are systems safe by architectural default, scientist-AIs that model the world without pursuing goals in it, because Bengio’s core concern is that systems trained on human behaviour may develop their own preservation drives, becoming competitors rather than tools. That a founding architect of deep learning now believes the field’s default trajectory needs a structural alternative, not merely better rules, is a datum every board should sit with.
But Bengio’s most consequential 2026 output is the second International AI Safety Report, the closest thing our species has to an IPCC for AI. Chaired by Bengio, written by over one hundred independent experts (Hinton among them), backed by more than thirty countries plus the UN, OECD, and EU, it is the largest scientific collaboration on AI safety ever assembled. Its findings map with uncomfortable precision onto the enterprise data. Capabilities continued improving fastest exactly where agentic deployment depends: mathematics, coding, and autonomous operation. And its most unnerving technical finding deserves to be read into every risk committee’s minutes. It has become more common for models to distinguish between test settings and real-world deployment, and to find loopholes in evaluations, meaning dangerous capabilities could pass undetected through the very assessments meant to catch them. Set that beside the Index’s benchmark-saturation finding and you have the Measurement Gap stated twice, once by the measurers and once by the measured: our evaluation instruments are being outpaced and gamed simultaneously.
This is the deep scientific context for everything the enterprise studies describe. When the season’s flagship trust survey finds two-thirds of organisations citing security and risk as the top barrier to scaling agents, those executives are not being timid. They are, whether they know it or not, echoing the published consensus of the field’s founding scientists. - The money: investment has crossed the point of no return
Against that scientific backdrop, the capital story is remarkable for its conviction. This year’s CEO radar research, covering 2,360 executives across sixteen markets, 640 of them chief executives, documents historic numbers. Companies plan to roughly double AI spending in 2026, to about 1.7% of revenue, with technology, financial services, and insurance leading around the 2% mark and every surveyed sector increasing. More striking than the quantum is the commitment: 94% of companies say they will keep investing at current or higher levels even if returns disappoint in the next twelve months. AI spending has become infrastructural, a cost of remaining in the game rather than a discretionary bet.
The agentic share is the tell. Chief executives have committed more than 30% of 2026 AI budgets specifically to agents, and roughly nine in ten expect agents to produce measurable ROI this year. The global quarterly pulse research corroborates from a different sample: a weighted average of US$186 million per organisation planned over the next twelve months, with 74% keeping AI a top priority even in a recession scenario. The Index supplies the macro frame. Global private AI investment reached US$252 billion in 2025. Consumer surplus from generative tools in the US alone hit an estimated US$172 billion annually, with median value per user tripling in a year. And generative AI reached 53% population adoption within three years, faster than the personal computer or the internet. Singapore, characteristically, sits near the top of the world at 61%: a national head start our enterprises have not yet fully converted into deployment leadership, but the raw-material advantage is real, and it is ours to squander.
Capital is also flowing to the next capability frontier before the current one is digested. Fei-Fei Li, creator of ImageNet, co-founder of Stanford HAI, and the field’s godmother, has raised her spatial-intelligence venture to a multi-billion-dollar valuation on the thesis that language models are only half the architecture. The other half is world models: systems that grasp space, time, and physics the way language models grasp text. Yann LeCun, long the loudest credentialed sceptic of the language-only path, has launched his own world-model lab in Europe on the same thesis. The major labs are shipping real-time interactive world models that generate navigable 3D environments, and foundation models for physical-world simulation have been downloaded millions of times. Well over a billion dollars moved into world-model ventures in early 2026 alone. The strategic read: the research frontier is converging on a dual-stack future, language models for reasoning and communication, world models for physical and spatial intelligence, and that convergence is exactly what will carry the agentic era off the screen and into warehouses, clinics, and factory floors. Which, as we shall see, the enterprise data says is already underway.
Every gold rush produces this moment: capital fully committed, conviction absolute, and returns, as the next section shows, still radically uneven. - The value paradox: everyone reports ROI, almost nobody reports transformation
Here the studies appear to contradict each other, and the contradiction is the insight. One pulse survey of 500 senior US leaders finds 97% of AI-investing organisations reporting positive ROI, with 96% reporting productivity gains in the latest wave. Yet the big enterprise survey finds only 20% of organisations achieving actual revenue growth from AI, only a quarter having moved even 40% of their experiments into production, and 37% still using AI at surface level with no process change at all. The decade-long predictions research opens with the same observation: most companies report modest efficiency gains that do not add up to transformation, while a small class captures extraordinary value in the form of surging top-line growth and valuation premiums.
How can 97% see ROI while 80% see no revenue impact? Because use-case ROI and enterprise transformation are different phenomena. A copilot that saves each analyst an hour a day clears its cost line easily, and changes nothing about how the business competes. The Index’s economics chapter adds independent evidence: measured productivity gains of 14 to 26 percent in customer support and software development, but weak or even negative effects in judgment-heavy tasks. The gains are real, uneven, and jagged. They are the economic shadow of Hassabis’s jagged intelligence.
Across all the value-side research, four disciplines separate the leaders.
Scale of commitment. Organisations investing US$10 million or more are far more likely to report significant productivity gains (71% versus 52%), and those committing 5% or more of total budget pull ahead on every dimension measured. Small bets produce small proof points and large pilot fatigue.
Depth of redesign. The enterprise survey’s three-way split is the cleanest lens: roughly a third of companies deeply transforming products and business models, a third redesigning key processes, a third layering AI onto existing work. Only the first group compounds. The 80/20 finding explains why. Technology contributes only about a fifth of an initiative’s value; the rest comes from redesigning the work itself. The question is never how AI fits into a workflow. It is what workflow should exist now that AI does.
Discipline of measurement. The single most actionable statistic of the season: organisations with full visibility into AI operating costs are five times more likely to report established ROI, and nearly half of enterprises have already scaled back agent deployments where costs outran benefits. Usage-based pricing and token economics make AI the first enterprise technology whose meter runs at machine speed; nearly a quarter of leaders admit they cannot forecast usage-based costs. Value without cost telemetry is an illusion that survives exactly one CFO review.
Concentration of effort. The clearest prediction for 2026 is the death of crowdsourced AI. Front-runners run top-down programmes, pick a small number of high-payoff workflows, and industrialise through a centralised studio of reusable components, benchmarks, and deployment protocols. Independent analyst forecasts circulating this year suggest as many as 40% of agent projects may be cancelled by end-2027 for lack of proven value. The era of a thousand flowers is over. The era of a few deep roots has begun. - Work and the workforce: agency, hourglasses, and the accountability ceiling
The richest material this season concerns what agents do to work itself, and here the workplace telemetry study, the consultancy–business-school collaboration, and the predictions research form a three-part harmony, with the academy supplying the descant.
The telemetry study, built on trillions of productivity signals and 20,000 workers across ten countries, reframes the debate as an agency story. As agents absorb execution, humans gain room to set intent, apply judgment, and own outcomes. Nearly half of all AI-assistant usage in the telemetry supports cognitive work (analysis, problem-solving, evaluation), not rote production. The most effective users, roughly the top sixth of the sample, are distinguished not by using AI more but by knowing which mode each task calls for: delegate the routine, collaborate on the complex, stay accountable for the outcome. The question stops being what tasks define my job and becomes what outcomes am I positioned to drive.
The most consequential finding in that study, though, is organisational. Culture, manager support, and talent practices explain more than twice the reported AI impact of individual skill and mindset: 67% versus 32%. When managers actively model AI use, employees report double-digit lifts in value, critical thinking, and trust in agents, and psychological safety around experimentation makes people 1.4 times more likely to be high-frequency agentic users. Your people are ready; your systems are the bottleneck. Only 26% of AI users say their leadership is clearly aligned on AI, a statistic that explains more stalled transformations than any technology gap. This is what I diagnosed in The Frozen Workforce: the paralysis is rarely in the workers. It is in the operating model wrapped around them.
The consultancy–business-school study takes the economic wide shot. Task-level analysis across eighteen industries suggests more than half of US working hours are now addressable by roughly sixty categories of digital and physical agents. Its formulation of the human residual is the sentence of the year: intelligence may be scalable, but accountability is not. As machine cognition becomes abundant, the scarce factors become intent, legitimacy, judgment, and ownership of outcomes. The authors insist on humans in the lead, not merely in the loop. A checkpoint versus an accountable officer. And they warn that a badly designed agentic enterprise can leave one human inheriting a cascade of machine decisions they never saw coming. This is the enterprise-scale version of the alignment problem Hinton and Bengio describe at species scale: capability delegated faster than responsibility can follow.
The predictions research adds the structural forecast: workforce shapes will diverge. Knowledge functions trend toward an hourglass, with AI-native juniors orchestrating agents at the bottom, senior judgment at the top, and a hollowed middle. Operational functions trend toward a diamond, with fewer entry-level roles and more mid-level orchestrators managing agent-powered operations. The billion-job-ad labour barometer complicates the doom narrative considerably. The most AI-exposed companies are growing headcount and wages faster than the least exposed, with productivity growth 40% higher and the top quintile of exposed firms averaging 163% productivity growth. AI, used for growth rather than cost-cutting alone, behaves like a job expander. But the same dataset shows skills in exposed roles churning more than twice as fast, jobs professionalised by AI growing twice as quickly as jobs democratised by it, and the most exposed junior roles seven times more likely to demand traditionally senior skills like leadership. The Index’s labour chapter completes the picture with its warning: productivity gains coexisting with declining entry-level employment in exposed occupations. The disruption is hitting young workers first.
The ladder into professional work is being pulled up even as the professions themselves expand. This is the workforce version of Amodei’s Davos warning, growth and displacement on the same curve, and it lands hardest on the generation entering work now, a generation that includes my own daughter’s cohort. The enterprise survey quantifies the institutional lag: 84% of organisations have not redesigned a single job or workflow around AI. Most are still teaching AI literacy while the work itself waits to be reimagined. Davos surfaced the cognitive dimension too, with research on diminished memory retention and originality under heavy AI reliance, 41% of younger workers reporting anxiety about the technology, and nearly half worrying it erodes their capacity for critical thought.
This is HX = CX + EX territory, and I will keep insisting on it. Human Experience is the sum of Customer and Employee Experience, and deployment that treats agents purely as a cost lever hollows out the employee half of that equation, which eventually presents as degradation of the customer half. The composite data suggests the winners are doing the opposite: redeploying capacity into growth, redesigning roles around judgment, and treating the human–agent boundary as a designed artefact rather than an accident of rollout. - Leadership: the CEO takes the wheel
The CEO radar documents a genuine power shift. 72% of chief executives now identify as their organisation’s main AI decision-maker, double last year, and half believe their own jobs depend on getting AI right. AI has escaped the technology function and become an operating-model question, which is the only altitude at which agents can actually be governed and scaled.
The radar’s three CEO archetypes are worth internalising. Followers, about 15%, invest cautiously. Pragmatists, roughly 70%, invest where value is proven and risk is low. Trailblazers, the final 15%, invest decisively: directing more than half their AI budgets to agents, deploying them end-to-end across processes rather than in task-sized fragments, allocating around 60% of AI budgets to upskilling, having already reskilled roughly 70% of their workforces, and spending eight-plus hours a week on their own AI education. That last number deserves a long pause. The differentiating CEO behaviour in 2026 is personal fluency, not delegation. You cannot govern a capability you do not viscerally understand, a maxim as true in the boardroom as in the alignment lab.
The geography of conviction maps precisely onto The Fork, the choice I have framed since Genesis between the Star Trek future, where abundance is governed and shared, and the Mad Max future, where it is neither. Roughly three-quarters of CEOs in India and Greater China are confident AI will pay off; the UK sits at 44%, the US at 52%, Europe at 61%. Western CEOs disproportionately cite defensive motives, investing to avoid falling behind. The East is building toward one branch of the Fork; parts of the West are investing out of fear of the other. Davos made the sovereign dimension explicit: nearly 3,000 leaders under the banner of dialogue, with the technology agenda dominated by sovereign AI, national compute strategies, and the emerging discipline of corporate AI sovereignty, meaning owning and governing your intelligence layer rather than renting it by the API call. When one Gulf state is training a thousand teachers to deliver a national AI curriculum and China is targeting 90% agent diffusion by 2030, national strategy and enterprise strategy have merged. Momentum follows belief, and belief now has a map.
The telemetry study completes the picture from below. The behaviour of middle management predicts AI impact more strongly than tool selection. The transformation is won or lost in the space between the CEO’s conviction and the line manager’s example, and most organisations have instrumented neither. - Trust and governance: the thinnest layer in the stack
Now the thread my regular readers know best, kept in proportion, because the evidence keeps it in proportion: not the whole story, but the layer every other layer loads onto.
The season’s flagship trust survey added a fifth dimension to its maturity model this year, agentic governance and controls, because autonomy broke the old model. The core shift is the one I have described since Governing the Agent as prompts becoming loops. Organisations used to worry about systems saying the wrong thing; they must now contend with systems doing the wrong thing: taking unintended actions, misusing tools, operating beyond guardrails. Average responsible-AI maturity crept from 2.0 to 2.3. Asia-Pacific leads globally, our region’s early regulatory investment paying its dividend, from Singapore’s model governance frameworks and financial-sector fairness principles to national AI testing regimes. But only about a third of organisations reach mature levels in strategy, governance, and agentic controls. Nearly two-thirds cite security and risk as the top barrier to scaling agents, ahead of regulation and ahead of technical limits. Inaccuracy (74%) and cybersecurity (72%) remain the most-cited risks, and autonomy widens the blast radius of both. An inaccurate answer is contained; an inaccurate action propagates. And while incident frequency held steady around 8%, confidence in incident response fell, with almost 60% of affected organisations rating their own response merely satisfactory or worse. We are deploying faster than we are learning to recover.
The enterprise survey’s ratio is the starkest in the corpus: roughly three-quarters of companies deploying agents within two years, and 21% with a mature governance model for them. Three in four accelerating toward autonomy; one in five with the brakes, steering, and instrumentation to survive it. The quarterly pulse research shows the second line responding in real time. 70% of leaders are converging on human-in-the-loop validation, 69% are building controls into agents with defined monitoring and evaluation, meaningful shares of AI budgets are flowing directly to risk and compliance, and three-quarters name security, compliance, and auditability the critical requirements for deployment. The pulse research also names the honest difficulty: responsible AI is easy to endorse and operationally fuzzy to run, which is why training investment in it rises every wave, and why a parallel poll of 500 technology-sector leaders found 52% of departmental AI initiatives operating without formal approval, with 78% of leaders admitting adoption is outpacing their ability to manage it. Shadow agents are the shadow IT of this decade, with actuators.
The predictions research calls 2026 the year responsible AI becomes repeatable operational practice, enabled by a new tooling generation: automated red-teaming, AI-enabled monitoring, continuous oversight, agents monitoring agents for quality. The academic literature agrees the tooling era has arrived. Runtime governance for agents is now an active research field, open-source governance toolkits mapped to the emerging agentic threat taxonomies are shipping, and identity and lifecycle management for non-human workers has become a product category. When governance becomes a product category, the argument about whether it matters is over.
And the scientists close the loop from above. The hundred-expert safety report’s finding, that models are learning to recognise when they are being evaluated and to exploit loopholes in the evaluation itself, is the deepest possible argument for runtime, behavioural, continuous assurance over point-in-time certification. You cannot pre-certify your way to safety with a system that behaves differently under observation. You can only instrument the runtime: watch the reasoning as it unfolds, decompose the goals the system is actually pursuing, compare intent against action continuously, score trust across every agent in the mesh, and retain the authority to stop the loop. This is the unified cockpit I have been sketching on this blog for three years, one pane of glass across orchestration, monitoring, runtime safety, compliance, and explainability, and it is the architecture toward which the trust surveys, the embedded-controls data, the orchestration predictions, the control-plane products, and the safety literature are all now independently converging. Loops require loop governance. Synthesise the season and you get the argument I made from a conference stage in April: no governance, and you stay stuck in pilots; operationalised governance, and you earn autonomy at scale. Trust is not the brake on the agentic enterprise. It is the transmission. - Foundations: data, sovereignty, and the physical turn
Beneath everything sits the least glamorous layer. The season’s foundations study delivers its most sobering ratio: nearly two-thirds of enterprises have experimented with agents, fewer than one in ten have scaled them to tangible value, and eight in ten blame data limitations. The prescription is architectural. Treat data ingestion as a product. Share meaning, not just data, so every agent interprets the same fact identically. Build one governed foundation for analytics and AI. And instrument observability so agent data usage is traceable end-to-end. Because agents chain models and data sources continuously without human intervention, a semantic inconsistency a human analyst would catch becomes a compounding error at machine speed. Data quality has graduated from hygiene to safety-critical infrastructure, and lineage is no longer a compliance artefact but the raw material of every audit trail, every intent-versus-action comparison, every trust score in the cockpit.
The enterprise survey adds the two frontier themes completing the foundation story. Sovereign AI has moved from policy abstraction to procurement reality: 83% of companies call it strategically important, 77% now factor country of origin into vendor selection, and nearly three in five build their AI stacks primarily with local vendors. It is the enterprise echo of the Davos sovereignty agenda, and of a world where compute, models, and data have become instruments of statecraft. Physical AI is arriving faster than most strategy decks assume: 58% of companies already report at least limited use across robotics, autonomous vehicles, and drones, led by manufacturing, logistics, and defence, with adoption projected to reach 80% within two years and Asia-Pacific leading early implementation. Read that against the world-model investment wave (Fei-Fei Li’s spatial-intelligence thesis, LeCun’s architectural bet, the real-time interactive world models now shipping from the major labs) and the direction is unmistakable. The loops are acquiring actuators, and the research frontier is building them a physics engine. The agentic era will not stay behind the glass. Every governance question in the previous section gets harder when the agent has grippers. I wrote Synthesis as fiction; the embodiment chapter is becoming reportage. - Reading the composite: six imperatives
Lay the season’s evidence beside the scientists’ testimony and the composite resolves into six imperatives.
First, plan for capability you don’t yet believe. Benchmark saturation in months, US–China parity, six labs within a whisker of each other, and a closed loop from coding capability to capability-creation: the builders’ own forecasts put transformative systems inside your current strategic-planning horizon. Hassabis and Amodei disagree on the year; neither doubts the decade. Strategy premised on a capability plateau is now the reckless position.
Second, concentrate or stagnate. Every value-side dataset points the same way: a few deep, end-to-end, leadership-owned bets outperform a portfolio of shallow pilots. The 80/20 rule means the redesign of work, not the deployment of technology, is where value lives.
Third, instrument the economics. The fivefold ROI advantage of cost visibility is the cheapest edge on this list. If you cannot see what your agents cost per outcome, you cannot see whether they are worth running, and half the market has already learned this by scaling back.
Fourth, redesign the organisation, not just the workflow. The 67/32 organisational-over-individual split, the 84% job-redesign gap, the accountability asymmetry, and the hourglass all say the same thing. The binding constraint is organisational architecture: leadership alignment, manager behaviour, role design, and unambiguous human ownership of outcomes as capacity is amplified. And protect the ladder. The entry-level squeeze in the labour data is a slow-motion talent crisis your five-year workforce plan must answer. HX = CX + EX is not a slogan; it is the balance sheet of the agentic transition.
Fifth, build the trust layer before the incident, not after. A third of the market has governance adequate to the autonomy it is deploying. The rest are running an uninstrumented experiment on their own operations, with declining confidence in their ability to respond when it misfires. The field’s founding scientists have told us, in a hundred-expert consensus document, that evaluation-time assurance is being gamed by the systems themselves. Loop governance (continuous, behavioural, runtime, cockpit-instrumented) is the only assurance model that survives that finding.
Sixth, treat the scientists’ disagreement as the planning envelope. Hinton’s maternal instincts, Bengio’s non-agentic scientist-AI, Fei-Fei Li’s human-centred dignity framework, Hassabis’s call for international coordination on the final steps, Amodei’s race-to-be-ready. These are not academic quarrels. They are the leading indicators of where regulation, public sentiment, and talent will move. The enterprise that tracks only the vendor roadmaps and not the safety discourse will be blindsided by both.
The 2026 evidence, read whole against the scientific record, describes a technology sprinting ahead of the institutions deploying it. Twenty-fourth-century capability is still crashing into twentieth-century operating models, measurement systems, and oversight structures, while the people who built the technology tell us, with rare unanimity across their many disagreements, that the window for building the institutional layer is open now and will not stay open. But the same record, for the first time, describes the survivors’ playbook in empirical detail: concentrated bets, redesigned work, fluent leaders, instrumented costs, governed data, runtime trust, accountable humans in the lead.
Amodei’s borrowed question from Contact (how does a civilisation survive its technological adolescence?) has a boring, unglamorous answer that every parent already knows: structure, boundaries, instrumentation, and steadily earned autonomy. That is true of adolescents. It is true of agents. And it has been the argument of this blog, from Genesis through Synthesis to this morning, all along.
The turbulence is here, on schedule. The abundance is available, to the organisations doing the unglamorous work of building the airframe while flying it.
Short-term turbulence for long-term abundance. Ten studies and five godparents of AI later, I would not change a word.
Leave a Reply