Institutions are handing authority to AI systems one rational step at a time, and each step raises the price of taking it back. That is the verdict. No model needs to want power. No executive needs to make a catastrophic mistake. The strongest pathway to functional loss of control runs through procurement, not through scheming. It is an economics of domination without a villain, and it can be measured in tokens, contracts, and audit trails.
The claim has a shape. Delegated authority is a function of time, and the function is a curve with a bend in it. Early steps are cheap to reverse. Later steps are not, because the assets that make reversal possible, skills, records, alternative suppliers, atrophy with use of the replacement. Somewhere along the curve the cost of leaving crosses the value of everything saved by staying. Past that crossing, the institution formally remains in charge. Functionally, it cannot say no. This article defines the curve, prices its current position from dated public records, names the mechanisms that bend it, and states what observation would break it.
Method. Every number carries its denominator, its date, and its source, woven into the prose. Every number is followed by what it does not prove. The historical comparison class is drawn from outsourcing, cloud computing, and the ratings agencies before 2008, three earlier delegation curves whose reversals are already priced. The evidence record runs to late August 2026.
Kill condition for the entire article, stated up front. Four observations would break the curve: enterprise adoption flattens, switching costs measurably fall, audit trails become portable across vendors, or oversight budgets grow in step with delegated authority. Any two of the four break the master chain. None is currently reproduced at scale. The rest of this article is the case for that sentence.
The curve, defined
Delegation begins as arithmetic any manager can defend. A task costs fifty dollars of human time and two dollars of machine time. The machine's output passes review. The task moves. Nothing about the decision is wrong. The error lives one level up, in the shape of the sequence.
Each delegated task drags three things with it. The record of how the task is done moves into the vendor's system. The person who used to do the task stops practicing it. The alternative supplier of the task, another team, another vendor, an internal capability, loses volume and eventually loses existence. All three effects are silent. None appears on the invoice.
the record of the the skill follows the a task is delegated work migrates record the alternative the exit price supplier dies resets
The bend in the curve comes from interconnection. Savings scale roughly linearly: each delegated workflow saves its own hours. Reversal cost does not. One hundred delegated workflows have 100 × 99 / 2 = 4,950 pairwise interfaces, and an exit requires revalidating essentially all of them. At two hundred workflows the interfaces number 19,900. The savings doubled. The revalidation burden quadrupled. That quadratic gap is the curve. It is not a metaphor. It is the arithmetic of integration.
This gives a precise meaning to irreversibility. A delegation is reversible when the institution retains three assets: people who can still do the work, records sufficient to reconstruct how decisions were made, and a credible alternative supplier. Irreversibility is the state where all three are gone. Note what the definition excludes. It excludes any property of the model. The curve is an institutional object. The model is just the cheapest counterparty.
Two falsifiers bound the definition. First, if switching cost grows linearly with adoption rather than superlinearly, there is no bend and no point of no return, only a price. Second, if institutions demonstrably re-skill, re-record, and re-source while delegating, the three assets never deplete and the curve never closes. Both conditions are observable. Neither is currently reproduced, and the next three sections explain why.
The measured present
The curve is not hypothetical. Four dated measurements fix its current position: a depth gradient inside one customer base, a delegation share inside one provider's output, a duration distribution of delegated tasks, and a contractual layer of forward commitments. Each measurement comes with the limit of what it shows.
First, the depth gradient. OpenAI's August 2026 enterprise report defined frontier firms as the top ten percent of customers by monthly output tokens per active user, and typical firms as the 45th-to-55th percentile band. In June 2026 the frontier cohort produced 8.3 times the typical cohort's intensity, up from 2.6 times in January, in an evidence base of more than ten million messages. The ratio widened by a factor of about 3.2 in five months. What it does not prove: business value. Output-token intensity can reflect longer responses, product routing, or task mix. It is a measure of depth of use, not of profit. What it establishes is narrower and sufficient: within one customer base holding similar tools, adoption depth is splitting into two populations. The curve's slope is not uniform. It is steep exactly where institutions have gone furthest.
Second, the delegation share. OpenAI reported that Codex, its agentic coding product, generated 64 percent of combined Codex and ChatGPT output tokens among enterprise customers by June 2026. The provider itself flagged the mechanical inflation: agentic workflows generate more tokens because they run longer, multi-step tasks. What it does not prove: that 64 percent of work is delegated, or that delegated tokens carry delegated authority. Token share is not authority share. What it establishes is the direction of traffic. The majority of one major provider's enterprise output now comes from the product category that acts, not the category that answers.
Third, the duration of delegated tasks. In a May 2026 Codex sample, OpenAI estimated that 80.6 percent of individual users delegated at least one task exceeding thirty minutes of human work, 70.2 percent delegated one exceeding an hour, and 25.6 percent delegated one exceeding eight hours. By June, the 99th-percentile daily users regularly generated more than sixty hours of agent turns per day across parallel agents. What it does not prove: successful completion. The durations are model-derived estimates of human-equivalent work, not wall-clock runtimes, and agent-turn hours include waiting and retries. What it establishes: users are handing over work whose human performance would span days, and parallelism has decoupled delegated machine time from the human day. Oversight capacity has not decoupled from anything.
Fourth, the contractual layer. Oracle closed fiscal 2026 with 638 billion dollars of remaining performance obligations, contracted future revenue not yet delivered, up 363 percent year over year, attributing most of the recent growth to large AI contracts. Of the related hardware, 75 billion dollars was prepaid or supplied by the customers themselves. CoreWeave reported 60.7 billion dollars of such obligations against 5.1 billion dollars of revenue, a ratio near twelve to one. What these do not prove: revenue, margin, or delivery. Obligations span years, carry cancellation and availability terms, and can evaporate with a counterparty. What they establish: customers are signing multi-year, balance-sheet-scale commitments to specific AI capacity before the capacity exists. The curve is being financed forward. A contract denominated in 2030 delivery is a delegation that cannot be reconsidered quarterly.
The seat counts complete the picture at the shallow end. Microsoft reported more than 30 million paid Microsoft 365 Copilot seats by July 29, 2026, with customers above 50,000 seats growing sevenfold year over year. Alphabet reported more than eight million paid Gemini Enterprise seats across more than 2,800 companies within four months of launch. Paid seats are not active use, and active use is not dependence. But distribution is the wide mouth of the curve. Default placement in the office suite is where delegation starts being free at the margin.
The pattern across the four measurements: depth is concentrating, the acting share is rising, delegated duration is lengthening, and the commitment horizon is extending past the tenure of the executives signing it. None of the four proves dependence. Together they fix the coordinates of a system moving up the curve's early, still-reversible stretch. The mechanisms that make the stretch temporary are next.
Why the reversal cost compounds
Four mechanisms bend the curve: specificity, the migrating audit trail, the junior pipeline, and default placement. Each is independently documented in the current record. Together they are the engine, and none of them requires a villain.
Mechanism one: specificity. Switching costs compound because the investment is not portable. Prompts, fine-tunes, memory stores, permission structures, governance workflows, and staff expertise are all built to one vendor's interfaces. The failure-front ledger's diagnosis is blunt: a migration requires revalidation of everything, so the threat of exit is weak. A switching cost that grows with every month of use is not a price. It is an annuity paid to the incumbent.
Mechanism two: the audit trail migrates. This is the deepest lock of the four, because it attacks the institution's ability to even conduct an exit, and it deserves its own chain.
decisions move into the vendor's systemthe logs of those decisions live there tooleaving requires reconstructing what was decidedthe evidence needed to leave is held by the party being left
An institution that cannot reconstruct its own past decisions cannot defend a switch to its board, its regulator, or a court. The audit trail is the evidence base for exit itself. When it lives in the vendor's system, in the vendor's format, under the vendor's retention policy, the option to leave becomes theoretical before it becomes expensive. This is the point of no return of the corporate pathway, and it arrives quietly, as a feature called compliance tooling.
Mechanism three: the junior pipeline. The failure-front ledger calls it the slowest invisible pathway, and its slowness is exactly its strength, because a mechanism that produces no incident never triggers a response.
automate the junior tasksjuniors stop learning the fundamentalsthe senior pipeline thinsoutput holds while competence drainsa failure arrives that nobody left can read
The best dated evidence sits in a peer-reviewed study of a staggered deployment across 5,172 customer-support agents. AI assistance raised issues resolved per hour by 15 percent on average and about 30 percent for less-skilled workers, and produced no measurable gain for the most experienced. Read it as a productivity result and it is a success. Read it as a training-system result and it is a warning: the tool absorbed the learning loop. Junior tasks are not just output. They are the curriculum from which seniors are made. An institution that automates the curriculum keeps this year's output and mortgages the decade's competence. The reversal cost here is the highest of all, because an institution cannot rebuild expertise faster than a crisis. Skills regrow on a timescale of years. Crises arrive on a timescale of hours.
Mechanism four: default placement beats marginal quality. Distribution decides which system gets delegated to before quality is ever compared. Eight million Gemini Enterprise seats did not win eight million bake-offs; they arrived inside a surface the customer already paid for. The ledger's formulation: default placement in operating systems, office suites, clouds, and developer platforms lowers acquisition cost and raises switching friction, and hundreds of millions of default users outweigh benchmark gaps measured in single points. The vendor that owns the surface owns the first step of the curve. The first step is the one every customer takes.
The four mechanisms share a property. None requires a vendor to act badly. Each is the ordinary consequence of ordinary enterprise software economics, applied to a product that does work instead of storing it. That is what makes the curve load-bearing: dismantling it requires not better villains but different incentives.
The delegation discount
Markets misprice the curve, and the mispricing has a structure. The benefits of delegation are immediate, legible, and attributable. The reversal costs are deferred, illegible, and unattributable. Every accounting system on earth prices the first kind and ignores the second.
Run the arithmetic the way a finance department runs it. A delegation decision saves two million dollars this year. The reversal cost, if the dependency sours, lands in year six and might run to fifty million in revalidation, retraining, data extraction, and lost negotiating position. At a twelve percent discount rate, that year-six cost is worth 50 × 0.507, about 25 million in present value. Then apply the tenure correction: the executive who books the saving holds the role for roughly three years. Inside the decision horizon that actually governs the choice, the reversal cost is not 25 million. It is zero. There is no line item, in any enterprise system, for "the ability to say no in five years." An asset with no line item has no defender in a budget meeting.
The asymmetry has a second layer. The benefits of delegation accrue to the delegating unit, the team that ships faster. The reversal costs accrue to the institution as a whole, later, under a different management, in a crisis. This is cost displacement across time rather than across parties, and it is the same shape as every underpriced tail risk in financial history. The seller of the option collects the premium now. The buyer of the option, the institution's future self, is not in the room.
the saving posts the reversal cost the discount rate this quarter lands in year six halves it executive tenure the curve is priced erases it at zero
The delegation discount explains why the strongest players move fastest. The frontier cohort running at 8.3 times typical intensity is not a population of reckless firms. It is a population of firms whose internal accounting correctly reports the savings and correctly reports nothing else. Under the prices they see, full delegation is the rational optimum. The thesis of this article is not that institutions are stupid. It is that the price system they trust has no quote for exit, and that a curve priced at zero will be climbed to the top.
What would falsify the discount claim: observable enterprise budgeting that carries reversal cost as a line item, exit-drill budgets, portability reserves, oversight headcount scaled to delegated authority. The final section prices exactly that signal. It is currently absent at scale.
What the last three curves cost
This delegation curve has been climbed before, three times, by three industries. The reversals are all priced. The comparison class is the strongest evidence that the curve is real, because in each earlier case the shape was identical: individually rational steps, silently compounding exit costs, and a reversal bill that arrived all at once.
Outsourcing. From the 1990s onward, moving production to cheaper suppliers was rational factory by factory and contract by contract. Economists Autor, Dorn, and Hanson estimate that Chinese import competition accounts for roughly one million of the United States manufacturing jobs lost between 1999 and 2011, within a total decline of several million. The number worth carrying is not the job count. It is the discovery, made two decades later during attempts to reshore, that what had been lost was not capacity but ecosystem: suppliers, tooling, process knowledge, and the workers who hold them. Each sourcing decision looked reversible. The system-level reversal turned out to require rebuilding a cohort that no longer existed. That is mechanism three, run on a national scale, with a twenty-year lag.
Cloud computing. Here the reversal has been attempted and priced, which makes it the most instructive case. In 2022 the software firm 37signals left Amazon and Google's clouds. It cut a 3.2-million-dollar annual bill to about 1.3 million, saving by its CTO's accounting on the order of ten million dollars over five years. The roughly 700,000 dollars of replacement hardware was recouped inside the first year. The exception proves the rule in both directions. 37signals could leave because it had deliberately avoided proprietary managed services, because it retained the skills to run its own machines, and because its workload was stable, precisely the three assets the delegation curve consumes. Even so, one remnant stayed locked: its object storage sat in a four-year contract that did not expire until 2025. For the median enterprise the exit tax is heavier. Egress fees, the charge for moving your own data out, run roughly 0.09 dollars per gigabyte at the largest provider. Gartner estimates them at 10 to 15 percent of total cloud spend. The industry learned the word exit only after the meter was running.
Ratings agencies. This is the closest structural match, because the delegation was of judgment itself. Before 2008, the financial system outsourced credit assessment to three agencies, paid by the issuers they graded, with the grades embedded into law: pension rules, capital requirements, and fund mandates were written in the agencies' vocabulary. Each institution's reliance was rational, the ratings were cheap, standardized, and legally sufficient. When the curve broke, the reversal was not a migration. It was a cascade. One academic tabulation of structured-finance ratings counts roughly 8,000 downgrades in 2007 against roughly 37,000 in 2008, a four-and-a-half-fold surge after the underlying paper had already failed. Funds mandated to hold investment-grade paper were forced to sell into a market with no buyers. Standard and Poor's eventually paid 1.375 billion dollars to settle the United States government's fraud case. The reversal cost of that delegation curve is measured not in settlements but in a global financial crisis. And the dependence survived its own discrediting: the issuer-pays model and the regulatory embedding remain substantially intact, because no institution could rebuild the judgment layer the agencies had absorbed.
The three cases rhyme exactly. Entry was rational at every step in all three. Exit costs compounded silently in all three, in skills, in architecture, and in judgment respectively. And in all three, the reversal, when it came, was paid in a lump: two decades of industrial erosion, a years-long migration under contract lock, a cascade measured in trillions. The AI delegation curve is steeper than any of the three, because it consumes all three asset classes at once: the skill, the record, and the alternative.
The strongest case against
The case against this thesis is strong and deserves its best form rather than its weakest. It rests on three pillars, each with real evidence behind it, and each fails on the concentration numbers already cited.
First, competition disciplines vendors. Foundation-model capability remains contested, leadership rotates, and a vendor that exploits dependence invites substitution. Second, multi-vendor strategy is now standard doctrine: enterprises route across providers, and the June 2026 national-security AI directive explicitly mandated multi-vendor model onboarding for the government itself. Third, exit rights are contractual. Enterprise agreements carry termination clauses, data-export terms, and portability commitments that did not exist in the outsourcing era. On this reading, the curve is a line, exit is a purchase order, and the point of no return is a rhetorical device.
Now the answer, taken in the same order. Competition disciplines only through credible exit, and credibility is exactly what the four mechanisms erode. A customer who can leave in ninety days disciplines a vendor. A customer whose prompts, memory, permissions, governance workflows, and audit trail are vendor-specific, and whose migration requires revalidating 4,950 interfaces, holds a ticket, not a weapon. The ledger's switching-cost hypothesis names the test: a multi-model migration drill reveals the real cost. The drills are not being run.
Multi-vendor strategy, second, operates at the wrong layer. Routing inference across three model providers diversifies the model layer. It does not diversify the identity layer, the permission layer, the logging layer, or the compliance layer, the surfaces where dependence actually lives. The flow of the failure:
two vendors at the model layerone vendor at the identity and audit layerthe records stay putthe option to leave expires unexercised
Contractual exit rights, third, confuse a legal right with an operational capacity. 37signals held every exit right Amazon offered. The exit still required retained skills, portable architecture, a stable workload, and a four-year wait on one locked component. A right you cannot afford to exercise is a clause, not a control. And the concentration numbers say the market itself does not believe the substitution story: customers are prepaying 75 billion dollars of hardware into one vendor's AI contracts and signing twelve-to-one obligation-to-revenue structures with another. Institutions do not prepay suppliers they expect to leave.
The honest residue of the counterargument survives, and it should be stated. Capability leadership is genuinely contested, and that contest is the strongest force currently bending the curve's early stretch. If three vendors remain at rough capability parity with falling switching costs, exit stays cheap and the curve stays a line. The thesis does not require monopoly. It requires only that exit costs outrun contestability. Today the evidence, the 8.3-times depth split, the migrating audit trail, the forward-financed contracts, says they are outrunning it. The measurement that would settle the question is named below.
What would break the curve
The verdict is dated late August 2026 and it carries its own kill conditions. Four measurements would break the delegation curve, any two of which suffice. None is currently reproduced at scale.
One: adoption flattens. The frontier-versus-typical intensity ratio, 8.3 times in June, 2.6 times in January, same customer base, stops widening and begins to converge. Convergence would mean depth is saturating rather than concentrating, and the two-population split is an artifact of early adoption, not a slope. Watch the ratio quarterly, on the same cohort definitions.
Two: switching costs fall measurably. The test is not contracts but drills. Critical institutions run timed, priced migration exercises, moving a live workload, with its prompts, memory, permissions, and logs, to an alternative vendor within a declared recovery objective, and publish the cost. The failure-front ledger's own viability condition is stated exactly this way: critical users can migrate within tested recovery objectives, and no single failure domain exceeds a systemic loss budget. A world where exit is rehearsed and priced is a world where the curve stays a line.
Three: auditability becomes portable. Decision records, permission histories, and action logs export in open formats, revalidatable by a successor vendor or a regulator without the incumbent's cooperation. The tooling category already exists in embryo, policy-as-code systems for agent permissions, approvals, credentials, audit trails, and rollback, but it is being built largely by the vendors themselves, inside their own stacks. Portability built by the incumbent is a feature. Portability enforced from outside is the break.
Four: oversight budgets grow with delegated authority. The quantity to track is a ratio: human oversight capacity, reviewers, auditors, unaided-competence headcount, divided by delegated machine output. Today the numerator is flat and the denominator compounding; one provider's top enterprise users already generate sixty-plus hours of agent turns per human day, a ratio no review function staffed on human time can hold. A flattening of that divergence, or longitudinal evidence that institutions retain unaided competence in the tasks they delegate, would cut the third mechanism at its root.
Until two of these four observations exist, the verdict stands. The curve is real, it is measured, it is being climbed by the strongest institutions first, and it is priced at zero by every accounting system currently in use. The cinematic story, a model that seizes authority, remains SPECULATIVE, and this article has needed none of it. The actual story is an invoice that never arrives, a skill that quietly stops being taught, and a contract that outlives the signature. Domination without a villain does not look like a coup. It looks like a renewal.
