newest wave of articles, opinion items, and forecasts from consultants and AI distributors, all confidently predicting the approaching takeover of human work by agentic AI, I just lately discovered myself occupied with my first yr at Imperial School London, again in 2009. Our lecturer for the introductory programming course would typically inform us to sketch out the options to coursework assignments with pen and paper earlier than coding them up on a pc, as a result of that slower, extra deliberate technique of writing and drawing by hand might help set up a deeper understanding of the subject material and result in higher options.
As we speak that recommendation appears extra pertinent than ever. The seductive promise of AI has drawn in people, organizations, and governments alike. A pupil submits an AI-generated essay, bypassing the cognitive wrestle by which writing turns into understanding. An organization replaces human staff with AI brokers, shedding the tacit, institutional data wanted to guage the AI output. A courtroom dietary supplements sentencing choices with opaque and doubtlessly biased algorithmic danger scores. Regardless of the motivation in every case, ceding cognition, judgment, and accountability to AI could effectively grow to be the default over time, till our dependency on AI — and the gamers in its broader ecosystem — will get so deeply embedded and normalized that reversing it appears neither mandatory nor possible. This forfeiture of human company is regarding sufficient in itself. What makes it extra so is that the polished-looking AI output being deferred to is, on nearer inspection, typically of solely generic worth, and generally merely unsuitable.
There may be an illuminating parallel right here with the administration consulting trade, which performs a strikingly related confidence trick, or “con”. Consultants routinely bundle generic analyses in polished shows and mission far higher certainty of their suggestions than the underlying proof would warrant. Organizations that defer to exterior consultants wholesale steadily shed the capability to suppose for themselves. Consultants will be unsuitable, each within the high quality of their insights and of their moral conduct, and AI methods are, at their core, probabilistic sample matchers with no intrinsic values and no accountability for the outputs they produce. Each derive their authority much less from demonstrated correctness than from the asymmetry between the arrogance with which their suggestions are delivered and the shopper’s diminishing means to problem them.
Taking the consulting trade’s structural dynamic as an instructive analogy, this text synthesizes what we find out about our rising dependence on agentic AI, traces the dangers of overdependence on the particular person, organizational, and societal ranges, and proposes methods to reclaim company at every of these ranges earlier than our capability to take action irretrievably slips away.
The Authentic Con
In 2023, the economists Mariana Mazzucato and Rosie Collington revealed The Massive Con, arguing {that a} confluence of incentive buildings, data asymmetries, and institutional pressures had allowed the administration consulting trade to extract returns in extra of the worth it creates. Absolutely, refined purchasers would merely cease hiring costly corporations that didn’t ship? Not fairly, since a shopper who outsources a strategic perform over a few years could now not possess the interior experience to guage whether or not the recommendation it receives is sound, which is a key dynamic that perpetuates the engagement. The parallel with at the moment’s AI distributors is noteworthy: deep integration, excessive prices, and a buyer base that progressively grows much less able to evaluating what it’s being offered.
Regardless of its conspiratorial connotation, the central downside The Massive Con highlights is structural quite than ethical. Many consultants in Mazzucato and Collington’s account had been genuinely attempting to have a optimistic influence, and lots of the officers who employed them had been doing their finest to behave responsibly below price range and staffing pressures. The association required neither cynicism from the consultants nor negligence from the purchasers, however merely that repeated delegation be allowed to run its pure course. What The Massive Con finally provides is much less a critique of any trade or particular person than a warning about how markets reward short-termism. Outsourcing as a “rationalization” measure can paint a falsely favorable monetary image within the close to time period, protect determination makers by externalizing accountability, and hole out the aptitude to strategize and execute. It’s the company equal of managing health by costly private trainers quite than growing the behavior of train your self: little question, trainers can ship actual worth, however they’ll additionally construct (over-)dependence quite than self-sufficiency.
The mechanism at work is one thing akin to the darkish facet of what administration students name “unlearning by not doing.” The much less a corporation performs a perform internally, the much less it is aware of the way to do it; the much less it is aware of, the extra it wants exterior assist; and the extra it pays for out of doors assist, the much less it builds the data that might ultimately make exterior assist pointless. A reform motion that swept by public administration from the Nineteen Eighties onward, within the U.S. and past, popularized precisely this logic below the slogan “steer extra, row much less.” However Mazzucato and Collington present why that strategy backfired for features the place doing and directing can’t be cleanly separated. As they put it, “The much less [an organization] rows, the much less it learns, the much less productive it turns into: the much less it may possibly steer.” The chance is particularly acute for strategically central features, as a result of as soon as these features migrate exterior, rebuilding them internally could take years.
The broader sample is rooted within the idea of comparative benefit. In a posh, specialised world, it may be rational to focus on what one does finest and pay others to deal with every little thing else. People rent plumbers and electricians. Firms herald specialists for tax and authorized work. Governments contract out the place personal suppliers have deeper experience. Every determination is defensible in isolation. It begins to grow to be an issue, nonetheless, when the logic of comparative benefit carves out not simply peripheral duties however the important, strategic work that defines what a person, group, or establishment truly is. A pupil who outsources her pondering, an organization that outsources its technique, a authorities that outsources its policymaking: in every case the delegating social gathering cedes the very capability that shapes its identification. With agentic AI, this forfeiture goes additional nonetheless, as company passes to a system with no actual accountability to the entity utilizing it.
The Subsequent Massive Con
The AI story of the 2020s bears intriguing parallels to the consulting story of previous many years: promising early wins, rational incremental delegation, and a structural dependency that solely turns into seen as soon as reversing it’s too expensive. Early AI use instances, from drafting textual content and writing code to summarizing paperwork and automating routine processes, are intuitive and ship outcomes with little friction. The barrier to preliminary adoption feels decrease than for nearly any prior expertise. The mainstreaming of generative AI for the reason that launch of ChatGPT in late 2022 compressed what may need been a decade or extra of gradual institutional adoption into only a few years.
AI distributors are determined to quickly seize market share and present large-scale adoption to justify growth-stock valuations, so AI providers are presently priced effectively under the true value of supply. The short-term economics are thus artificially compelling, a lot as early cloud and SaaS choices used backed pricing to construct the combination depth that later made switching painful. As such, agentic AI is poised to speed up the displacement of labor at a tempo and scale that consulting by no means approached: the place enterprise software program as soon as offered instruments for expert professionals to function, AI-driven platforms are more and more closing the books, drafting contracts, and producing technique paperwork immediately, eradicating people from the loop, and with them, the organizational data that justified preserving the perform in-house.
Earlier than analyzing the dangers, it’s value acknowledging the numerous advantages that the considered use of AI can ship. Automating repetitive, rules-based, high-volume duties frees human consideration for work that’s extra intellectually stimulating and requires cautious judgment. Customized AI tutoring instruments present early promise — early research recommend measurable positive factors in pupil engagement and comprehension. A well-used AI sparring associate can sharpen pondering quite than changing it; e.g., one can ask it to argue the other facet, discover the weakest hyperlink in a single’s reasoning, or floor arguments that haven’t but been thought of. That is what Cory Doctorow, in his current ebook The Reverse Centaur’s Information to Life After AI, calls being a “centaur”: an individual who chooses how and when to make use of AI as a device, retaining company over the work. Used properly, agentic AI could make folks extra succesful, not much less, so it’s apparent to see the attraction.
Nonetheless, the profit curve bends towards diminishing returns when the delegation to AI turns into so pervasive and inconsiderate that the human can now not critically consider the AI output and train significant judgment to override it. Past that time, the human isn’t augmented however changed, and with them goes the organizational data wanted to correctly handle the AI — each by way of methodology and output. And the place people stay, they’re more and more required to serve the AI (generally to the purpose of exploitation), in impact turning into what Doctorow calls “reverse centaurs.” A person who stops wrestling with issues and easily accepts the solutions, a corporation that invests in AI-led course of automation with out growing staff who can perceive and oversee it, a authorities company that stops growing regulatory experience and merely ratifies what the AI recommends: in every case using AI results in atrophy, and the descent into overdependence will be gradual sufficient that the shift goes unnoticed till reversing it’s tough.
In the meantime, the business narrative driving AI adoption is present process a major shift. Distributors and buyers are more and more positioning AI not solely as a option to increase human work however as a option to change it. Software program budgets are sometimes smaller and fewer strategic than payroll budgets, so reframing AI by way of labor substitute implies a far higher addressable market. The pure beachhead is outsourced labor: actions with well-defined scope, separable budgets, and organizations already accustomed to exterior options. Extra ambiguous and strategically central in-house roles could comply with, as positions are decomposed into duties amenable to AI automation; see Sangeet Paul Choudary’s ebook Reshuffle for a deeper dialogue. If not managed fastidiously, this course of dangers automating the duties the place human oversight and judgment are most consequential.
There may be additionally a possible battle of curiosity in who’s advising the transition. The consultancies that drove the labor-outsourcing wave are actually accumulating profitable charges as advisors on AI adoption methods. McKinsey has projected $2.6 to $4.4 trillion in annual worth from generative AI, a forecast that conveniently underpins the case for large-scale advisory engagements. The ethical hazard is structural: any exterior advisor paid to help with AI transformation has a transparent disincentive to assist purchasers construct the interior functionality that might ultimately make follow-on engagements pointless. No particular person guide has to behave in dangerous religion for this dynamic to function. It’s the logic of markets, not the malice of corporations, and it’s the identical logic that sustained the unique con of consulting elaborated by Mazzucato and Collington.
Allow us to now look at, in flip, the price of this rising overdependence on AI on the particular person, organizational, and societal stage.
The Particular person Value
The person value of cognitive delegation is insidious exactly as a result of AI output can more and more go for the conclusions of human evaluation. George Orwell recognized the important thing sample in his 1946 essay Confessions of a E book Reviewer, which lamented how industrial ebook reviewing had hollowed the craft to a simulacrum of engagement. The reviewer was “consistently inventing reactions in the direction of books about which [he] has no spontaneous emotions no matter,” pouring “his immortal spirit down the drain, half a pint at a time.” AI has the capability to show all of us into Orwell’s manufacturing unit reviewer. This phenomenon isn’t new — folks upvote social media posts they haven’t learn and signal petitions they don’t perceive — however AI industrializes it, making it simple to connect one’s title to consequential work regardless of minimal involvement. The worth of cognitive work lies not solely in what’s produced however in what the particular person turns into by the act of manufacturing it. In offloading that wrestle to agentic AI, we danger forfeiting not simply the output however the capability for unique thought.
In Might 2026, MIT professor Micah Nathan wrote in The Guardian a few pupil who submitted an AI-generated story to a inventive writing workshop. What struck him was the character of the coed’s mental capitulation. She started by feeding her story to the AI for a grammar verify. It instructed line edits and he or she accepted them. Then structural edits. Then a full rewrite. It was, Prof. Nathan wrote, “a sequence just like an addict’s descent,” every step feeling small, each making the subsequent extra seemingly. His pupil had not got down to give up authorship so fully, however at every level the place she may need reclaimed it, concern of judgment by her friends and her trainer proved stronger, and he or she let the AI take over as an alternative. As Prof. Nathan noticed, “Writing isn’t simply the manufacturing of sentences — it’s the coaching of endurance by means of sustained consideration. It’s a means of studying what one thinks by trying to say it.” Extended use of AI help weakens the capability to kind and articulate ideas independently, till at some point a clean web page appears to be like not merely uncomfortable however daunting, and what was as soon as a useful gizmo has grow to be a crutch one can not think about doing with out.
Educational research are starting to supply empirical grounding for this concern. Lately revealed findings from large-scale randomized managed trials present that even temporary AI help can impair subsequent impartial efficiency. Examine members who used AI not solely carried out worse unaided but additionally stopped attempting sooner. The hypothesized mechanism is hedonic adaptation, whereby as soon as ready-made solutions grow to be the norm, working by an issue independently begins to really feel disproportionately expensive. One other research of almost two thousand skilled adults discovered that whereas the bulk agreed AI “did many of the pondering” throughout government perform duties, additionally they reported that the ensuing concepts didn’t really feel totally their very own. A 3rd research discovered measurably decrease neural connectivity amongst members who used AI to write down essays. The researchers coined the time period cognitive debt to explain “a situation wherein repeated reliance on [AI] replaces the effortful cognitive processes required for impartial pondering.” Whereas these research measure particular task-level efficiency, the broader declare that atrophy compounds throughout advanced judgment-intensive work is an extrapolation according to hedonic adaptation and deserves additional analysis. Taken collectively, the proof means that sustained use of agentic AI could erode each the cognitive ability and the urge for food for the wrestle that might rebuild it.
The Organizational Value
The organizational dangers prolong past particular person deskilling. When many individuals delegate cognitive work over time, distributed institutional reminiscence thins out, governance buildings grow to be hole with out the interior judgment they depend upon, and strategic autonomy disappears as soon as nobody contained in the group can consider or problem the seller. If the consulting analogy holds, a troubling descent may unfold. Agentic AI initially augments staff, then hiring slows down because the AI covers sufficient floor that headcount progress appears onerous to justify, and ultimately — within the absence of deliberate governance — roles are eradicated and workflows redesigned round AI brokers. The important institutional data that the departing workers carry leaves with them and isn’t rebuilt. Given how far more quickly AI is being embedded into organizations than consulting relationships ever had been, this endpoint may arrive ahead of many count on. Furthermore, AI providers are presently backed far under their true value to seize market share. When the main focus ultimately shifts to profitability, distributors could increase costs (e.g., by switching from subscriptions to usage-based billing, as we’re already seeing with GitHub Copilot). Organizations that considerably — and maybe indiscriminately — slash headcount in the course of the subsidy interval will discover the ensuing dependence on AI distributors expensive to unwind.
The short-term economics are already shocking early adopters. Microsoft (satirically, additionally a significant purveyor of AI hype and an AI vendor in its personal proper) reduce its inside AI coding licenses in 2026, roughly six months after encouraging company-wide adoption, citing spiraling prices. Uber exhausted its annual AI coding price range in 4 months. Salesforce expects to pay Anthropic roughly $300 million in 2026 alone. An Nvidia vp acknowledged publicly that “the price of compute is way past the prices of the workers.” Token costs are falling, however agentic AI consumes tokens at a price that outpaces the worth decline. Goldman Sachs forecasts a 24x enhance in consumption by 2030, and at some corporations AI token expenditures already equal 10% of complete labor prices. A current survey of almost 2,500 corporations discovered that for each greenback spent on AI tokens, a pitiful 18 cents generated user-facing worth, whereas 44 cents went towards fixing bugs the AI methods themselves launched. Worse, when organizations observe adoption and productiveness primarily based on token spend, staff have interaction in “tokenmaxxing” (inflating utilization metrics by intentionally utilizing AI inefficiently). So not solely does token spend fall wanting delivering commensurate productiveness positive factors, however sources channeled into AI infrastructure should not being invested in individuals who would accomplish the identical work extra cost-effectively whereas constructing the institutional data the group will ultimately want.
There may be additionally what practitioners name the oversight tax. In high-stakes contexts (e.g., authorized, monetary, medical), organizations sometimes preserve human reviewers within the loop to fulfill legal responsibility necessities. The group pays for each the system and the reviewer, whereas the reviewer’s impartial capability steadily weakens within the presence of AI. Managed research of AI-assisted medical screening recommend that the obvious ability atrophy could also be resulting from a “safety-net impact,” whereby people put in much less effort after they know the AI is there to catch their errors. Fred Brooks, in The Legendary Man-Month, argued that high quality is dependent upon conceptual integrity: the design should be held within the thoughts of somebody who understands the entire. A human who evaluations AI outputs they didn’t produce and can’t totally interrogate has surrendered that integrity. The oversight tax subsequently signifies that productiveness positive factors are overstated, and the accountability these positive factors are purported to justify is hollower than it seems.
The Societal Value
When the group in query is a faculty, a courtroom, or a authorities ministry, the stakes are excessive. Normalizing using agentic AI as an alternative choice to crucial pondering in schooling dangers producing a technology of graduates that undermines the financial system and tradition, and is extra susceptible to political manipulation. A decide who can not interrogate the algorithmic danger rating informing a sentencing determination could inadvertently (but systematically) drawback defendants of a sure race or ethnicity. A legislature that cedes undue affect over the drafting of payments to AI brokers could find yourself governing in opposition to the desire of the folks it represents. The time period “human within the loop” (HITL) was coined to explain people who’re in management, can problem the AI, and are accountable for the outcome. In observe, nonetheless, this typically means blindly approving AI output below time stress, and serving as what analyst Dan Davies calls an accountability sink — there not primarily to forestall errors however to soak up blame for them. The issue is that the individuals who care most in regards to the societal final result sometimes don’t have any entry to the AI agent’s reasoning, whereas the distributors don’t have any institutional incentive to problem their very own methods. Accountability requires somebody each motivated and outfitted to behave. HITL at the moment typically produces neither.
There may be additionally a geopolitical dimension to all of this. The United States and China collectively management roughly 90% of worldwide compute capability and 70-80% of worldwide AI funding. Anton Leicht has just lately described the closing of what he referred to as the “Andy Warhol period of AI entry,” the temporary interval when frontier capabilities had been obtainable to wealthy and poor customers alike. The rising construction is decidedly hierarchical: new frontier capabilities circulation first to nationwide safety institutions, then to giant trusted enterprises, then to chose worldwide companions, and solely then to everybody else. European nations that depend upon American AI infrastructure, and Asian economies counting on both American or Chinese language platforms, are constructing crucial public capability on a basis that may be conditioned, priced past attain, or withdrawn completely as per the supplier’s strategic pursuits. Singapore parliamentarian Kenneth Tiong astutely noticed of his nation’s AI technique that “we’re constructing an AI hub on an assumption we don’t management.”
That summary danger grew to become concrete in June 2026, when the U.S. authorities ordered Anthropic to droop entry to its most superior fashions, Fable 5 and Mythos 5, for all overseas nationals, citing nationwide safety considerations. Anthropic, which had publicly referred to as for higher authorities oversight of AI, discovered itself unable to defend tons of of tens of millions of customers in opposition to a directive it described as poorly justified and primarily based on “verbal proof of a slim, non-universal” vulnerability. The corporate was given no selection and needed to abruptly disable the mentioned fashions for all clients worldwide to make sure compliance. The episode illustrates what dependency on frontier AI from a single supplier means in observe. The U.S. authorities has the authorized authority to order the suspension of business mannequin entry globally, a incontrovertible fact that holds no matter whether or not this particular order was justified. And the pricing, entry, and deployment choices of a small variety of AI distributors now have the facility to constrain the strategic choices of whole organizations and governments in a single day, with out warning, and for causes these organizations can not contest.
Statistician and economist Ernst Friedrich Schumacher examined a structurally related sample in 1973, analyzing what occurred when capital-intensive Western expertise was transferred to growing economies. It elevated combination output whereas making a “twin society”, with the advantages distributed erratically throughout social lessons, and constructing dependency on overseas experience that native populations may neither keep nor replicate. Within the age of agentic AI, open-weight fashions partially tackle this danger by permitting native deployment and adaptation, however the hole between open and frontier functionality stays giant and is actively maintained by the main suppliers. The acceptable expertise motion that grew from Schumacher’s work requested: Does the brand new expertise enhance native company, or does it focus profit elsewhere? Can it’s operated with out everlasting exterior dependency? These questions should be requested of AI deployment at the moment, at each scale from the person to the societal.
Two Useless Ends and a Third Path
Two responses to the dangers described above are tempting however almost definitely unsuitable. The primary is banning agentic AI outright, as some jurisdictions have beforehand contemplated with facial recognition or autonomous weapons. This might cede the sphere to the actors least inclined to train restraint whereas forfeiting potential advantages, together with advances in medication, scientific analysis, and schooling. The second is uncritical adoption, letting market incentives and vendor roadmaps decide how deeply agentic AI penetrates cognitive and democratic life. As Reich, Sahami, and Weinstein warn of their ebook System Error, this isn’t a impartial selection. Permitting the optimization logic of markets — which is detached to cognitive possession, institutional reminiscence, and democratic accountability — to make choices that have an effect on folks is itself a call, made by default. Even well-intentioned innovators can optimize the unsuitable issues at scale and find yourself imposing sure values and selections on the remainder of us whether or not we prefer it or not.
The third path lies someplace between prohibition and sleepwalking and requires lively selections at every of the three ranges of study coated within the previous sections.
On the particular person stage, agentic AI is well-suited to automating grunt work and serving as a sparring associate. Judgment, nonetheless, ought to stay with the human consumer who writes the synthesis, makes the decision, and indicators off the end result. The productive wrestle is value sustaining intentionally. Analysis on skilled adults means that staff who modified AI outputs extra typically confirmed considerably increased confidence of their impartial reasoning, making override frequency a sensible proxy for retained cognitive company. For every job, we must always ask whether or not AI help is constructing capability or substituting for it. The place it’s the former, use it totally. The place it’s the latter, deliberate resistance could also be mandatory.
On the organizational stage, institutional reminiscence should be handled as a strategic asset that requires lively funding. Construct documentation practices, mentorship buildings, and rotation of workers by features that AI assists however doesn’t personal. Diversify AI suppliers intentionally: no single-vendor dependency for any crucial cognitive perform, and hybrid architectures that mix small, locally-operable fashions for routine duties with bigger fashions for high-stakes reasoning to protect strategic choices. Be certain that “people within the loop” do in truth have the data and entry to problem the AI output, the time to take action, and organizational accountability for the choice. And above all, make structural sovereignty (retaining significant management over the data, compute, information, and fashions that underpin crucial operations) a board-level concern.
On the societal stage, public coverage can draw on the suitable expertise standards Schumacher articulated: Does this deployment enhance native company? Can it perform with out everlasting exterior dependency? Governments investing in AI for public providers also needs to put money into open-weight fashions, interoperability requirements, and shared compute infrastructure to cut back geopolitical vulnerability. HITL in consequential public choices (e.g., judicial, welfare, healthcare, immigration) ought to carry a authorized customary: the human reviewer should have the ability to clarify and defend the output they ratify, not merely attest that they had been current when it was generated. The sector of AI ethics deserves extra recognition and should be empowered to make sure rigorous governance of agentic AI alongside the strains that bioethics established for medication. Particular person and organizational self-discipline are mandatory however not ample. Structural change requires regulation, procurement requirements, and worldwide coordination along with particular person resolve. That shift is not going to occur by market forces alone, which have to this point pointed in the other way — residents and establishments should demand it.
What all three ranges share is identical underlying precept: the instruments ought to serve the individuals who use them, which requires a deliberate option to preserve it that means. Overdependence isn’t inevitable. And past a sure adoption threshold, the proof factors to important dangers of hurt, even when the general financial influence stays contested. We are able to acknowledge this with out being anti-technology. It’s the identical maturity that medication developed when it established contraindications, dosage limits, and knowledgeable consent. It’s the recognition that deliberate governance ought to complement the passion of a rollout.
The Wrap
The consulting and AI waves share an identical structural logic. Seemingly rational short-term delegation accumulates right into a dependency that erodes the capability to steer, with out the necessity for dangerous actors, and with out the results turning into totally seen till reversing course has grow to be too tough. The AI model of this course of is quicker, extra deeply embedded, and operates concurrently on the particular person, organizational, and societal stage.
Mockingly, the analytical work, code technology, and software program supply that consulting corporations have lengthy offered as high-value providers are the very classes that AI distributors are actually automating and bundling immediately into their platforms. The identical disintermediation logic the consulting trade utilized to company features is now being utilized to the consulting trade itself, as indicated by the sharp declines in share costs of enormous, people-heavy consulting corporations (Accenture, Capgemini, and so on.) over the previous a number of months. Whether or not that proves a corrective is uncertain: substituting a relationship-based dependency for a platform-based one, the place the seller’s leverage is structural quite than private, may very well result in deeper lock-in. What appears seemingly, although, is that the logic of the con is not going to spare the conmen.
In the end, it’s value remembering that we are those constructing AI instruments, and we will determine the way to use them. The recommendation that stayed with me from that first-year course on programming was not about avoiding computer systems — it was about doing the pondering first. That precept scales to the design of corporations and public establishments. The query is whether or not we’ll apply it intentionally, or let the present of short-term comfort carry us someplace none of us intend to go.
