Human Futures Engine · T-001 · created 2026-09-25 → horizon 2031-12-31 · last update 2026-10-01

Trusted Delegation: When Consumer AI Value Migrates from Generation to Action and Protection (2026–2031)openMedium

[THESIS] By 31 December 2031, the dominant source of realized consumer value from AI shifts from content/information generation to delegated action and protection — systems that transact, schedule, cancel, dispute, monitor and defend on the user's behalf — with the binding constraint being verified trust (liability, reversibility, auditability, authentication) rather than model capability. Operationalized as: (a) among adults in high-income markets who use AI at least monthly, a majority name an action/protection use case as their most valuable AI use; and (b) agent-initiated commerce and agent-managed account operations reach material, separately-measurable scale.

Static snapshot: runs happen on the scheduled morning job

Output

Probability computed in log-odds from the reference class and the weighted evidence · never stated by the model

Thesis probability	61.7%from a 77.3% prior · 23 active evidence items
Reasonable range	45% – 78% tick = prior · marker = posterior
Historical analogy strength	69/ 100 · 20 analogues in the reference class
Current evidence strength	99/ 100 · 23 active items
Evidence quality	64/ 100
Contradictory evidence	High
Confidence in probability	Medium
Next major confirmation signal	Unchanged and now more precisely dated: a top-10 US issuer or top-3 card network publishing an explicit consumer liability and chargeback-parity rule for agent-initiated transactions, with a stated spend ceiling and revocation standard. Because network rulebook changes ship on April/October cycles, the realistic first window is the April 2027 release — absence then would be a genuine negative update, not noise. Secondary tells over the next 90 days: (a) whether the voluntary federal AI accord's working groups produce an agent-transaction liability annex or explicitly decline to; (b) an insurer or reinsurer publicly quoting a priced agent-error product for consumer transactions; (c) Apple/Google shipping an OS agent with stored payment authority on by default; (d) Adobe Analytics or Shopify disclosing agent-originated GMV as a separate line.

prior log-odds +1.224 + evidence -0.746 = posterior +0.478 → 61.7%

Posterior log-odds = prior log-odds (reference-class base rate) + sum of quality x independence weighted log likelihood ratios. Research-informed estimate until enough theses resolve for Brier calibration.

Formalized thesis

Measurable restatement of the raw thesis

Statement[THESIS] By 31 December 2031, the dominant source of realized consumer value from AI shifts from content/information generation to delegated action and protection — systems that transact, schedule, cancel, dispute, monitor and defend on the user's behalf — with the binding constraint being verified trust (liability, reversibility, auditability, authentication) rather than model capability. Operationalized as: (a) among adults in high-income markets who use AI at least monthly, a majority name an action/protection use case as their most valuable AI use; and (b) agent-initiated commerce and agent-managed account operations reach material, separately-measurable scale.

PopulationAdults 18+ in high-income consumer markets (US, EU, UK, Japan, Korea, Canada, Australia), with US as the primary measured cohort; secondary read on bank/telecom/retail customer bases and smartphone OS install base.

BehaviourGranting a standing, revocable authorization to an AI system to take consequential actions — authenticate, pay, book, cancel, negotiate, dispute, file, monitor, block, or control a device/vehicle — without step-by-step human approval of each action, and paying for that capability (directly via subscription or indirectly via bundled device/bank/carrier fees).

Horizon year2031

Observable outcomeAgentic and protective AI becomes the default packaging of consumer AI: OS-level and bank/carrier-level agents with payment credentials, delegated-authority standards (scoped tokens, spend limits, revocation, chargeback parity) are live and insured; standalone 'ask-it-a-question' chat subscriptions become a commoditized, bundled feature rather than the primary paid product.

Measurable metricComposite, any 3 of 5 by 2031-12-31: (1) ≥50% of monthly AI users in a nationally representative US survey (Pew/Gallup/Ofcom-class, non-vendor) name an action/protection use as most valuable, vs ≤30% baseline 2026; (2) agent-initiated checkout ≥10% of US e-commerce GMV (measured by Visa/Mastercard/Stripe/Shopify agentic-transaction disclosures or Adobe Analytics agent-traffic tagging); (3) ≥40% of US smartphone users have enabled an OS/bank agent with stored payment or account-control authority; (4) consumer spend on AI-enabled protection (fraud/scam/identity/account defense, incl. bundled bank and carrier lines) exceeds consumer spend on generative content subscriptions; (5) ≥3 of the top-10 US banks and ≥2 of the top-3 card networks publish an explicit liability/chargeback framework for agent-initiated transactions. [EPISTEMIC] (1)–(3) are cleanly falsifiable; (4) requires an allocation convention for bundled revenue and is the weakest leg.]

Raw statement

By 2031, mainstream consumers will derive more practical value from AI systems that act, automate and protect on their behalf (trusted delegation: scheduling, purchasing, cancelling, protecting accounts, detecting scams, managing bills, travel, health information and vehicles) than from standalone AI systems primarily used to generate information or content. Convenience and security become inseparable: the limiting factor shifts from 'is the AI intelligent enough' to 'do I trust this system enough to let it act'.

Source observations

5 long-term human-behaviour bullets from 5 morning briefs · the observations this thesis consolidates

2026-09-13Time and productivity get bought through software agents rather than headcount; enterprises will pay for agentic tooling that shows measurable labour substitution, and punish software priced on seats (Adobe's soft guidance, Wix's cut).
2026-09-14Knowledge and productivity are being bought as per-seat agents rather than software licences (Copilot at 30M seats), which structurally favours incumbents with distribution over standalone model vendors.
2026-09-17Delegation of communication: the Edge-AI query set ('ai phone call', 'ai phone answering service', 'ai phone assistant') shows consumers ready to outsource the lowest-status use of their time \u2014 talking to strangers on the phone. Time is the scarce good; AI is being bought as time, not intelligence.
2026-09-24Delegation replaces search: Meta's Muse agent, the human-concierge test, and the $1,299 glasses point to consumers paying for an intermediary that does things rather than shows things — the largest re-plumbing of advertising economics since mobile.
2026-09-25Convenience is being bought as autonomy: 'ai phone assistant', 'ai phone answering service' and 'ai agent builder' are the fastest-rising consumer queries — people want an agent that handles their obligations, and they will pay for reclaimed time before they pay for entertainment.

Causal mechanism

Why humans would do this · the nine parts the analogues are matched on

Human motivation

[INFERENCE] Time and cognitive load are the scarce consumer goods, not information. Administrative burden (bills, subscriptions, claims, appointments, disputes, travel rebooking, vehicle and health admin) is low-status, high-friction labour people already pay to avoid. Simultaneously, AI-enabled fraud raises the baseline anxiety cost of holding digital accounts, so 'someone watching my back' is a second, non-discretionary motivation. Evidence surface in the dossier: fastest-rising consumer queries are 'ai phone assistant', 'ai phone answering service', 'ai agent builder' — outsourcing obligations, not curiosity.

Friction removed

The per-task human step: authentication, form-filling, phone queues, comparison, cancellation retention flows, monitoring for anomalies. Delegation collapses a multi-step, multi-vendor workflow into a single standing instruction. Also removes the vigilance tax: continuous scam/fraud monitoring a human cannot perform.

Enabling technology

Tool-use/function-calling models with reliable long-horizon planning; scoped credential standards (passkeys, delegated OAuth, network tokens, agentic payment protocols); on-device inference for privacy/latency/cost (Apple silicon cost-per-token claims, Qualcomm/edge NPUs); verifiable audit logs and action-level provenance; device-bound identity and behavioural biometrics for agent-vs-attacker discrimination.

Economic incentive

[INFERENCE] For consumers: recovered hours plus avoided fraud losses and avoided subscription waste — a measurable dollar ROI unlike content generation. For platforms: agents capture the transaction, not just the query, converting a $20/month subscription into a take-rate on commerce, payments, insurance and switching. For banks/carriers/insurers: agentic protection reduces loss ratios and churn, so they subsidize distribution. For incumbents with workflow data and distribution (Microsoft, Intuit, Alphabet, Apple, Amazon, Visa/Mastercard, Palantir-class enterprise), delegation is a margin expansion; for standalone model vendors it is a margin compression.

Trust requirement

The binding constraint. Required: (i) bounded liability — a clear 'who pays when the agent is wrong or hijacked' rule analogous to the 1974 Fair Credit Billing Act $50 cap; (ii) reversibility — one-click undo, chargeback parity, spend ceilings; (iii) auditability — human-readable action logs; (iv) authentication that survives prompt injection and social engineering; (v) a credible privacy story for the data the agent must see (email, calendar, bank, health). [INFERENCE] Historically, consumer delegation scales only after liability is socialized onto the institution, not the user.

Distribution mechanism

Pre-installed OS assistants (iOS/Android), bank and card-network apps, carrier bundles, browser/checkout integrations, and employer/insurer benefit bundles. Very little standalone-app adoption: delegation rides on entities that already hold credentials and liability relationships.

Network effects

Two-sided: merchants, airlines, utilities and banks must expose agent-readable interfaces and accept agent-initiated actions; agents become more valuable as more counterparties honour them (a 'do-not-call-the-call-centre' network). Also data network effects — fraud and scam detection improves with cross-customer signal, favouring the largest networks. Weak direct user-to-user effects.

Switching costs

Rising and sticky: an agent accumulates preferences, payment credentials, contact graph, historical disputes, health and vehicle context. Once the agent holds the credential set and the audit history, migration means re-granting dozens of authorizations. [INFERENCE] This makes early credential capture the strategic prize, mirroring default-payment-instrument lock-in.

Regulatory constraints

Payments liability (Reg E/Reg Z in the US, PSD2/PSD3 SCA in the EU), EU AI Act transparency and high-risk duties, FTC negative-option/click-to-cancel rules (which cut in the thesis's favour), state privacy law, HIPAA for health actions, and unresolved questions on agent legal agency, consent to contract, and prompt-injection liability. [SPECULATION] The decisive regulatory event is a chargeback/liability safe harbour for agent-initiated transactions; without it adoption stalls at read-only assistance.

Historical analogues

20 transitions sharing the mechanism · expand a row for the eight similarity dimensions · rows below 40 are dimmed and excluded from the base rate

Analogue	Period	Mechanism	Similarity	Occurred	Yrs to mainstream	Lesson
Credit cards + Fair Credit Billing Act liability cap	1958–1985	Consumers delegated payment authority to an instrument they did not control at point of use; adoption inflected only after the 1974 statutory $50 fraud liability cap and network-level dispute machinery socialized the downside.	82	✓yes	16	Trust infrastructure (capped liability + reversibility) preceded mass convenience by roughly a decade; capability was never the constraint. This is the closest structural template for agentic payments.
Ride-hailing: algorithmic dispatch plus stranger-trust	2010–2018	Two simultaneous trust leaps (stranger in a private car, algorithm choosing price and route), solved by ratings, GPS traceability, cashless settlement and insurance.	80	✓yes	7	Trust was manufactured by observability plus insurance, not by reputation of the provider — audit logs and indemnity are the agentic equivalents.
Voice commerce via smart speakers ('Alexa, order me…')	2015–2024	Same delegation promise — speak an intent, the agent transacts — but with weak error-recovery, poor merchant integration and no liability framework.	78	✗no	0	[EVIDENCE AGAINST] The single closest failed analogue. Capability adjacency is not enough: without reversibility, catalogue trust and a dispute path, users reverted to information-only use (timers, music, weather) — exactly the failure mode the thesis must avoid.
Google Duplex agentic phone calls	2018–2024	AI acting as the user's representative to third parties; blocked by counterparty consent, disclosure rules and narrow merchant coverage.	75	✗no	0	[EVIDENCE AGAINST] The counterparty side is the network bottleneck. Agents that must negotiate with unwilling institutions stall; agents operating inside a single institution's rails ship.
Direct deposit and automatic bill payment (ACH autopay)	1972–1998	Standing authorization for a third party to move money without per-transaction consent; pure delegation of recurring administrative labour.	74	✓yes	20	Delegation of recurring obligations became invisible default behaviour, but took two decades and required employer/utility-side adoption — the counterparty side, not the user side, was the bottleneck.
Amazon 1-Click / card-on-file commerce	1997–2010	Stored credentials plus removal of the confirmation step; complexity and risk hidden behind an interface, with merchant-absorbed fraud liability.	74	✓yes	10	Removing a single confirmation step produced outsized conversion gains — evidence that consumers will trade control for friction removal when the loss is bounded and reversible.
Card-network fraud scoring and proactive transaction blocking	1995–2012	Institutions given unilateral authority to decline/freeze on the customer's behalf; consumers accepted false positives in exchange for protection.	74	✓yes	10	Protection was adopted universally and invisibly because it was opt-out and institution-funded. The 'protect' half of the thesis has a materially higher base rate and faster clock than the 'act' half.
ATMs replacing teller interaction	1969–1990	Control over cash handling traded for 24/7 access; required trust that an unattended machine would correctly debit an account.	72	✓yes	18	Adoption followed network interoperability (shared ATM networks) more than machine quality — agent-readable counterparty networks matter more than agent intelligence.
Turn-by-turn GPS navigation replacing self-routing	2005–2015	Consumers surrendered situational judgement to an opaque optimizer, accepting occasional catastrophic misroutes for consistent time savings.	72	✓yes	9	Delegation of judgement scales fast when the feedback loop is immediate and the failure is visible and correctable in real time — agents with observable intermediate steps will be trusted sooner.
Robo-advisors and automatic 401(k) enrolment/target-date defaults	2006–2020	Discretionary authority over money granted to an algorithm; regulatory safe harbour (PPA 2006 QDIA) plus fiduciary framing unlocked default-level adoption.	69	✓yes	12	A regulatory safe harbour converted an opt-in trust problem into an opt-out default — the single highest-leverage policy analogue for agentic payments liability.
Password managers and passkeys	2012–2026	Users handed the entire credential set to a single intermediary — maximum trust concentration for maximum convenience and security.	69	✓yes	12	Adoption stayed sub-mainstream until OS vendors bundled it; standalone trust products underperform bundled ones. Distribution via Apple/Google/banks is decisive.
Spam filtering and automatic email quarantine	1998–2010	Software given unilateral authority to delete/hide messages on the user's behalf — a protective delegation with false-positive risk.	68	✓yes	8	Protective delegation is adopted far faster than transactional delegation because errors are cheap and reversible; this is the likely first beachhead for AI agents (scam detection, account defense).
ADAS: cruise control → adaptive cruise → lane-centering	2005–2025	Incremental, bounded delegation of physical control with explicit hand-back protocols — trust built by scope-limited autonomy.	68	✓yes	15	Bounded autonomy with a clear re-take-control mechanism scaled; unbounded autonomy did not. Scoped spend limits and revocation are the design pattern that wins.
Level-4/5 consumer self-driving promised for 2018–2020	2015–2026	Full delegation of a high-consequence physical task; capability timelines were badly over-forecast while trust/regulatory timelines were under-forecast.	68	✗no	0	[EVIDENCE AGAINST] Base rate for 'delegation becomes mainstream within 5 years of credible demos' is poor when liability is unresolved. Argues for slippage of the thesis into 2033–2035 for high-consequence categories, while low-consequence categories (subscriptions, scheduling, scam alerts) hit 2029–2031.
Mobile check deposit and mobile banking	2009–2018	Bank-side liability absorption plus smartphone distribution converted a trust-heavy financial behaviour into a default within a decade.	68	✓yes	8	When the trusted incumbent ships the feature inside an existing app, the trust question is pre-answered and adoption is fast — favouring banks, card networks and OS vendors over standalone AI brands.
Nest learning thermostat / delegated home control	2011–2022	Device given autonomous authority over comfort and spend, with a utility-side economic incentive (demand-response rebates) subsidizing adoption.	65	✓yes	10	Third-party subsidy (utility) accelerated delegation; expect banks/insurers/carriers to subsidize consumer agents for their own loss-ratio reasons.
Elevator operators eliminated after safety infrastructure	1945–1960	Automation existed decades before acceptance; the 1945 NYC operator strike plus emergency stop, auto-doors, voice reassurance and signage built trust.	64	✓yes	12	Trust artefacts (visible stop button, explanation, human fallback) mattered more than technical capability — the UI of revocation is the product.
Commercial aviation autopilot and autoland certification	1930s–1980s	High-consequence delegation gated by certification, redundancy, logging and clear liability allocation rather than by capability.	58	✓yes	30	Where consequences are severe, adoption is certification-paced. Health, vehicle and large-value financial delegation in the thesis will follow this slower clock than scheduling or shopping.
Automatic transmission adoption in US autos	1940–1970	Consumers gave up direct mechanical control for convenience once reliability was demonstrated; a pure control-for-ease trade.	56	✓yes	22	Control-surrender in a safety-relevant domain takes a generation and is driven by OEM default-fitting, not consumer demand — distribution defaults beat preference.
Microsoft Bob / Clippy / 1990s 'intelligent agents'	1995–2001	Premature delegation interface without underlying capability or user-controlled scope; produced durable brand damage to the 'agent' concept.	48	✗no	0	[EVIDENCE AGAINST] Interface-first delegation without reliability creates backlash and delays the real transition by years. A visible agent failure at scale (mass mis-purchase, prompt-injection theft) is the main tail risk to the 2031 date.

Reference class20 analogues with similarity ≥ 40 · occurred in 16 · base rate (Laplace) = 77.3% = prior

Consumer delegation-of-agency transitions: cases where a technology is granted standing authority to take consequential actions (financial, physical, informational) on an individual's behalf, where the binding constraint is trust architecture — bounded liability, reversibility, auditability, authentication, institutional indemnity — rather than raw capability. Twenty cases spanning 1930–2026, of which 15 realized the pattern and 5 did not. Observed regularities in the class: (i) median time from credible capability to mainstream is ~12 years, interquartile roughly 8–18; (ii) protective/defensive delegation (spam filters, fraud blocking, ADAS) mainstreams 2–4x faster than transactional delegation because errors are cheap, reversible and institution-funded; (iii) adoption inflects on a liability or safe-harbour event (FCBA 1974, PPA 2006 QDIA, Reg E), typically 3–8 years before the mainstream threshold, not on a model/hardware milestone; (iv) bundled distribution through an already-trusted credential holder (bank, OS, carrier, employer) beats standalone trust brands in every realized case; (v) the counterparty/merchant side, not the user side, is the recurring bottleneck — the failures (voice commerce, Duplex) died on counterparty integration and error recovery, not user willingness; (vi) scoped autonomy with visible hand-back (ADAS, spend limits) scales where unbounded autonomy does not.

Caveats: [CAVEATS] (1) Selection bias toward documented, eventually-successful delegations; many small agentic products failed without leaving a measurable record, so the true realization rate is below 15/20. (2) The thesis bundles heterogeneous categories with very different clocks — scam detection and bill management plausibly clear by 2029; vehicle, health and large-value financial delegation follow the aviation/ADAS certification clock and likely miss 2031. A composite metric can pass on the protective leg while the transactional leg fails, so the 'more value than generation' comparison is partly an accounting choice. (3) The adversarial dimension is genuinely novel: no historical analogue had an intelligent attacker able to hijack the delegated agent via its own input channel (prompt injection). This has no base rate and could either slow adoption sharply or, symmetrically, be the reason protective delegation becomes mandatory. (4) Historical delegations were deterministic and auditable; LLM agents are stochastic, which weakens the transferability of the reversibility mechanism — an undo does not repair a leaked credential or a disclosed medical fact. (5) Consumer survey evidence on 'most valuable AI use' is highly correlated across vendors and prone to framing effects; vendor-published agent-usage statistics should be scored low on independence. (6) The 2026 macro regime in the dossier (≈5% risk-free rate, hawkish Fed) is a mild tailwind to protection/cost-saving delegation and a headwind to discretionary standalone AI subscriptions, which biases the comparison in the thesis's favour for reasons unrelated to technology. (7) Base rates come from an era of falling capital costs and rising trust in institutions; both conditions are weaker now, and declining institutional trust is the most under-weighted drag on the delegation clock.

Evidence

23 active of 24 items · each weighted by quality × independence · retired items are struck through

#	Claim	Direction	Strength	Quality	Indep.	Weight	LR	Wtd log-LR	Kind	Source	Date
0	Fastest-rising consumer search clusters in the dossier are delegation-shaped, not generation-shaped: 'ai phone call', 'ai phone assistant', 'ai phone answering service' (Edge AI +25.6) and 'ai agents', 'ai agent builder' (AI Software & Agents +28.5) — people are querying for systems that handle obligations, not produce content.	supports	moderate	0.55	0.70	0.39	2.00	+0.267	Observed fact	Internal attention-mover panel, 2026-09-25 brief	2026-09-25
1	Card networks and payment processors have shipped agentic-commerce rails with scoped, revocable credentials (Visa Intelligent Commerce, Mastercard Agent Pay, Stripe/OpenAI Agentic Commerce Protocol, Google AP2), i.e. the tokenization and mandate plumbing for delegated purchase exists before mass demand — trust infrastructure preceding convenience, exactly the mechanism's prediction.	supports	strong	0.80	0.85	0.68	3.00	+0.747	Observed fact	Network/processor product announcements 2025–2026	2026-06-30
2	Distribution is being pre-wired by entities that already hold credentials and liability: Microsoft at ~30M Copilot seats and repackaging Copilot, Meta's Muse agent plus a human-concierge test, Apple pushing on-device inference on cost-per-token grounds. Delegation rides incumbents rather than standalone apps.	supports	moderate	0.70	0.60	0.42	2.00	+0.291	Observed fact	Briefs 2026-09-14 / 09-24 / 09-25	2026-09-25
3	AI-enabled fraud is already re-pricing security as non-discretionary: CRWD/ZS +12%, PANW +11% on a single AI-safety headline; a single cyberattack broke Boston Scientific's guidance. The 'protection' half of the thesis has a live, funded budget line today.	supports	moderate	0.65	0.70	0.45	2.00	+0.315	Observed fact	Briefs 2026-09-13 / 09-14	2026-09-14
4	Edge inference economics are improving fast enough to make always-on background agents affordable and privacy-defensible (Apple claiming new Macs beat cloud AI on cost per token; NPU-equipped phone base installed and growing), removing the cost objection to continuous monitoring rather than episodic prompting.	supports	moderate	0.60	0.75	0.45	2.00	+0.312	Observed fact	Brief 2026-09-24; device silicon roadmaps	2026-09-24
5	Regulatory tailwind for delegated cancellation/dispute: FTC click-to-cancel/negative-option rules and EU PSD2 SCA infrastructure mean the highest-friction consumer tasks (retention flows, disputes) are becoming machine-addressable and legally supported.	supports	weak	0.50	0.80	0.40	1.30	+0.105	Observed fact	US/EU regulatory record	2026-01-01
6	Reference-class mechanism has repeatedly completed: credit-card delegation scaled only after the 1974 $50 liability cap; ACH autopay, robo-advice/auto-enrolment, spam quarantine and card-network fraud blocking all reached majority adoption once the institution, not the user, absorbed the error cost. The pattern is that liability socialization, not capability, is the gate.	supports	strong	0.75	0.90	0.68	3.00	+0.742	AI inference	Historical analogue set	2026-09-25
7	Measured consumer AI usage today is still overwhelmingly informational: the large-scale usage studies of the dominant chat assistant show ~70–80% of consumer messages are practical guidance, information seeking and writing — action/transaction use is a small single-digit share. The thesis requires a >20pp shift in stated most-valuable use by 2031.	contradicts	strong	0.75	0.85	0.64	0.33	-0.700	Observed fact	Vendor/academic usage telemetry studies 2025–2026	2026-03-01
8	Voice commerce is the closest failed analogue: a decade of Alexa/Google Assistant installed-base ubiquity produced negligible delegated purchasing because consumers refused to hand over the purchase decision without visual confirmation. Ubiquitous distribution did not convert to transaction authority.	contradicts	strong	0.80	0.90	0.72	0.33	-0.791	Observed fact	Smart-speaker commerce history 2017–2024	2026-09-25
9	Prompt injection and agent hijacking remain unsolved at the architecture level; public red-team results repeatedly show agents with tool access can be steered to exfiltrate data or transact. Until an agent can be distinguished from an attacker deterministically, institutions will not extend chargeback parity.	contradicts	strong	0.70	0.80	0.56	0.33	-0.615	Observed fact	Security research 2025–2026	2026-08-01
10	Enterprises are actively restricting agentic model access over data-exposure fears — Nvidia, Palantir and Booz Allen restricting Anthropic model use — showing that the most sophisticated buyers reduce delegation as capability rises, which is the opposite of the thesis's trust trajectory.	contradicts	moderate	0.70	0.70	0.49	0.50	-0.340	Observed fact	Brief 2026-09-14	2026-09-14
11	No top-10 US bank or top-3 card network has yet published an explicit consumer liability/chargeback framework for agent-initiated transactions; pilots operate under existing card rules with merchant-side risk. Metric leg (5) is currently at zero.	contradicts	strong	0.70	0.75	0.52	0.33	-0.577	Observed fact	Network rulebooks / bank disclosures	2026-09-25
12	The long-horizon-reliability base rate is poor: Google Duplex was narrowed to a niche after launch, and consumer L4/L5 autonomy promised for 2018–2020 arrived ~7–10 years late. Multi-step agentic task benchmarks still show materially sub-95% end-to-end completion, below the threshold where unattended execution is tolerable.	contradicts	moderate	0.65	0.80	0.52	0.50	-0.360	Observed fact	Agentic benchmark literature; autonomy history	2026-07-01
13	A ~5% risk-free rate, hawkish Fed and a visibly financing-driven AI capex complex (Oracle force majeure, circular vendor financing, Burry shorts) raise the odds of a 2027–2028 AI funding air-pocket that delays consumer agent rollout by one to two product cycles even if the destination is unchanged.	contradicts	moderate	0.60	0.65	0.39	0.50	-0.270	Consensus expectation	Briefs 2026-09-17 / 09-24 / 09-25	2026-09-25
14	Merchant-side incentives are asymmetric: retailers and airlines earn margin from friction, upsell and retention flows, so a meaningful share of counterparties will block or degrade agent traffic (robots.txt/bot-blocking, agent-hostile checkout) rather than expose agent-readable interfaces — the network-effect leg is contested, not automatic.	contradicts	moderate	0.55	0.70	0.39	0.50	-0.267	AI inference	Bot-management and agent-traffic policy disputes 2025–2026	2026-05-01
15	The capability precondition is largely met for the narrow, high-frequency tasks the thesis names (bill anomaly detection, subscription cancellation, rebooking, scam screening) — these are bounded, reversible and already scripted; the residual gap is authorization and liability, not intelligence.	supports	moderate	0.50	0.60	0.30	2.00	+0.208	AI inference	Engine assessment	2026-09-25
16	Agent identity, delegated authority, tamper-evident action logging and millisecond containment are shipping as a commercial stack (Nvidia Open Agent Safety Platform, 100+ partners; DigiCert agent identity; IBM identity integration; Vera CPU for agents on CoreWeave; Arm partnership), explicitly framed as what would have stopped the Hugging Face hack. Two of the thesis's missing prerequisites — auditability standards and agent-vs-attacker authentication — are moving from research to purchasable infrastructure, enterprise-first.	supports	moderate	0.70	0.60	0.42	2.00	+0.291	Observed fact	Briefs 2026-09-28 / 09-29 / 09-30	2026-09-30 added 2026-10-01
17	The US federal response to agent risk was a voluntary 'morally binding' accord rather than statute, leaving the private sector to define agent governance. This forecloses the fastest legislated route to a Fair-Credit-Billing-Act-style consumer liability cap and makes the thesis dependent on card networks and issuers volunteering chargeback parity against their own loss ratios.	contradicts	moderate	0.65	0.75	0.49	0.50	-0.338	Observed fact	Brief 2026-09-30	2026-09-30 added 2026-10-01
18	Institutional reaction to agent risk is converting from refusal into priced containment: the same security complex that restricted model access is now buying per-agent identity and permissioning as a recurring line item, with Cybersecurity flow ranking above Cloud and Semis. Delegation is being made underwritable rather than being abandoned.	supports	moderate	0.60	0.50	0.30	2.00	+0.208	AI inference	Briefs 2026-09-29 / 09-30	2026-09-30 added 2026-10-01
19	Updated financing-shock read: the 10-year at 5.33–5.34% (highest since 2002), 30-year through 5.6%, UK 30-year gilts above 6%, a 16-month-high dollar, Micron beating on revenue without multiple expansion, and credit scrutiny of Nvidia's vendor-financing guarantees to neoclouds. The term-premium repricing is persisting through a soft core PCE print, which raises the probability of a 2027–2028 AI funding air-pocket that pushes consumer agent rollout one to two product cycles later even if the destination holds.	contradicts	moderate	0.70	0.60	0.42	0.50	-0.291	Observed fact	Briefs 2026-09-29 / 09-30 / 10-01	2026-10-01 added 2026-10-01
20	Consumer confidence at a 12-year low (81.9) with 9.3m student-loan borrowers in default while aggregate spending holds on a top-quintile AI wealth effect. The mass-market willingness to pay directly for a new standalone AI capability is weakening, reinforcing that consumer delegation must arrive bundled into bank, carrier, OS or insurer relationships rather than as a paid subscription — consistent with the thesis's distribution mechanism but negative for any direct-pay adoption path.	supports	weak	0.60	0.70	0.42	1.30	+0.110	Observed fact	Briefs 2026-09-29 / 10-01	2026-10-01 added 2026-10-01
21	AI-run healthcare billing is already imposing costs on patients without transparency, and California has banned AI 'robo bosses' — the first consumer/worker-protection rules aimed at machine decisions made about people rather than by them. This creates demand for a defensive counterparty agent (audit, dispute, appeal) and establishes the legal vocabulary of machine accountability the protection half of the thesis needs.	supports	weak	0.50	0.70	0.35	1.30	+0.092	Observed fact	Briefs 2026-09-30 / 10-01	2026-10-01 added 2026-10-01
22	SEC approval of tokenized stock trading with nearly half of volume occurring outside US market hours shows consumers already accepting unattended, machine-executed financial action in a high-stakes domain, weakening the claim that visual human confirmation is a hard requirement for delegated transactions.	supports	weak	0.45	0.60	0.27	1.30	+0.071	Observed fact	Brief 2026-09-29	2026-09-29 added 2026-10-01
23	Six weeks after the thesis was recorded, metric leg (5) remains at zero: no top-10 US issuer or top-3 network has published agent-transaction liability or chargeback-parity rules, and Visa/Mastercard pilots still run under existing card rules with merchant-side risk. Network rulebook changes land on fixed April/October cycles, so the realistic earliest publication is now April 2027 — compressing the remaining window for the 2031 composite.	contradicts	moderate	0.65	0.50	0.33	0.50	-0.225	Observed fact	Network rulebooks / issuer disclosures as of 2026-10-01	2026-10-01 added 2026-10-01

Stage of adoption

Where the transition sits on the historical pathway

innovators

Still credit cards circa 1969–1972: instrument exists, rails laid, fraud is the dominant objection, liability cap unlegislated — but with one refinement this month. The identity/audit/containment layer is being built in the enterprise first (closer to EDI and corporate purchasing cards in the 1980s) and will propagate to consumers only after it has an actuarial loss history. Read-only assistive AI remains early majority; delegated-action AI with payment or account-control authority remains innovators, now with a clearer institutional supply chain but no clearer consumer liability rule.

Prerequisites met

✓Capable tool-use models reliable on bounded, reversible tasks
✓Scoped-credential primitives shipped (passkeys, network tokens, delegated OAuth, Visa Intelligent Commerce / Mastercard Agent Pay / ACP / AP2)
✓Distribution owners committed (OS assistants, bank apps, card networks, carriers, Microsoft Copilot at ~30m seats)
✓Edge inference cheap enough for always-on background monitoring
✓Funded, non-discretionary anxiety budget: AI-enabled fraud and scam losses rising; security re-rating reflexively
✓Consumer intent visible in search: delegation queries outgrowing generation queries
✓NEW: commercial agent identity, delegated-authority and containment stack shipping at enterprise scale (Nvidia OASP + DigiCert + IBM), i.e. the audit/authentication layer is now purchasable rather than theoretical

Prerequisites missing

○Published consumer liability / chargeback-parity framework for agent-initiated transactions — still zero of five institutions, and the statutory route was closed by Washington's voluntary accord
○Prompt-injection defence strong enough for unattended consumer execution (enterprise containment is mitigation, not solution)
○~99% end-to-end multi-step task reliability unattended execution demands
○Merchant-side agent-readable interfaces and a non-hostile bot-policy equilibrium
○Insurance/underwriting market pricing agent error with an actuarial loss history
○Dispute-grade evidentiary status for agent action logs (logs now exist; legal weight does not)
○A consumer-side willingness-to-pay channel that survives a weak mass-market household balance sheet (must be bundled, not sold)

Scenarios

Bull / base / bear · weights normalised to 100%

Bull25%

A liability safe harbour arrives early (2027–2028), either via card-network rulebook change granting chargeback parity for agent-initiated transactions or via a Reg E/PSD3 clarification. Once error cost is socialized onto issuers and platforms, adoption follows the autopay/robo-advice curve: OS- and bank-level agents ship with default-on spend ceilings, insurers underwrite agent error, and merchants capitulate to agent-readable checkout to avoid losing basket share. All five metric legs clear by 2030; agent-initiated checkout passes 10% of US e-commerce GMV, standalone chat subscriptions are bundled to zero marginal price, and the scarce asset becomes the credential custody + identity graph, not the model. Payment networks, device OS owners, identity/fraud vendors and workflow-data incumbents capture the margin; model vendors become substitutable suppliers.

Base45%

Bifurcated adoption. Protection scales first and fast — scam screening, account defence, bill and subscription monitoring, fraud alerts — because it is read-mostly, low-blast-radius, and already funded by banks, carriers and insurers who cut loss ratios by deploying it. Transaction authority scales slowly and only inside narrow, bounded, high-trust corridors: reorder of known SKUs, travel rebooking with an existing PNR, utility and telecom bill disputes, in-network appointment scheduling, vehicle and device control. By 2031 the composite hits 3 of 5 legs — the survey leg and the enabled-agent leg clear, protection spend exceeds content-subscription spend, but agent-initiated GMV lands in the 3–8% range and liability frameworks are partial and merchant-segmented. The thesis's directional claim (trust, not capability, is the binding constraint) is validated; the quantitative threshold is met narrowly and late. Two-track market: whoever owns the credential and the liability relationship wins; whoever owns only the intelligence gets commoditized.

Bear30%

Two failure modes, either sufficient. (1) A high-profile agentic fraud or injection catastrophe — a hijacked assistant draining thousands of accounts — triggers a regulatory freeze and bank policy reversal, repeating the voice-commerce arc where ubiquitous distribution never converts to transaction authority. (2) A 2027–2028 AI financing air-pocket (the briefs' circular-financing, Oracle force-majeure, rate-shock channel) starves the rollout of capex and patience, pushing agentic commerce a full product cycle right. In either case, AI's consumer value stays informational and creative through 2031; agents remain a confirm-each-step copilot; protection grows as a security budget line but is captured by incumbent banks and cyber vendors rather than becoming a new consumer product category. The thesis is directionally right but arrives 2034–2037, and the pure-play agent vendors funded in 2025–2027 are the dead cohort.

Indicators

What to watch · the next confirmation signal is what the monthly review looks for first

Next confirmation signalUnchanged and now more precisely dated: a top-10 US issuer or top-3 card network publishing an explicit consumer liability and chargeback-parity rule for agent-initiated transactions, with a stated spend ceiling and revocation standard. Because network rulebook changes ship on April/October cycles, the realistic first window is the April 2027 release — absence then would be a genuine negative update, not noise. Secondary tells over the next 90 days: (a) whether the voluntary federal AI accord's working groups produce an agent-transaction liability annex or explicitly decline to; (b) an insurer or reinsurer publicly quoting a priced agent-error product for consumer transactions; (c) Apple/Google shipping an OS agent with stored payment authority on by default; (d) Adobe Analytics or Shopify disclosing agent-originated GMV as a separate line.

Supporting

·Delegation-shaped consumer queries outgrowing generation-shaped queries (Edge AI +25.6, AI Software & Agents +28.5, with 'ai phone answering service' and 'ai agent builder' in the top rank)
·Payment networks shipping scoped agentic-credential standards ahead of measurable demand
·Microsoft Copilot ~30M seats and repackaging — delegation distributed via incumbent workflow, not standalone apps
·Security complex re-rating reflexively on AI-risk headlines (CRWD/ZS +12%, PANW +11%) — protection is becoming a standing category
·Edge inference cost claims (Apple cost-per-token) enabling always-on background monitoring economics
·Enterprise willingness to pay for measurable labour substitution while punishing per-seat pricing — the pricing model is already migrating to outcomes/actions

Contradictory

·Zero of the required bank/network consumer liability frameworks for agent transactions published as of today
·Measured consumer AI usage still ~70–80% informational/creative
·Voice-commerce precedent: ubiquitous distribution, negligible delegated purchasing
·Prompt injection unsolved; sophisticated enterprises (NVDA, PLTR, Booz Allen) restricting rather than expanding model access
·Merchant bot-blocking and agent-hostile checkout — the network effect is being actively contested
·Rates/financing regime raising the odds the buildout slips a cycle (Oracle force majeure, circular vendor financing)

Biggest variables

·Whether liability for agent error is socialized onto institutions or left with the consumer
·Whether prompt-injection defence reaches a standard an underwriter will price
·Merchant cooperation vs. blockade — does agent-readable inventory become table stakes or a paywalled API
·The timing and depth of an AI financing air-pocket (a capex/funding event, not a demand event)
·Whether the winning agent sits at the OS layer (AAPL/GOOGL), the bank/network layer (V/MA/JPM), or the model layer (OpenAI/Anthropic) — determines who captures the take-rate
·Regulatory divergence: EU AI Act high-risk classification of consumer agents could fragment the rollout geographically

Value pools

Where economic value accumulates if the behaviour happens · universe tickers only

Layer	Emerging need	Opportunity	Becomes	Tickers
Delegated authorization & agent-vs-attacker identity	Proving that the entity acting is the authorized agent of the right human, under a scoped, revocable, auditable mandate that survives social engineering and injection	Device-bound identity, behavioural biometrics, mandate issuance and revocation, tamper-evident action logs with dispute-grade evidentiary status; sold to banks, OS owners and merchants as mandatory plumbing	mandatory infrastructure	AAPL AKAM CRWD FICO FTNT GEN GOOGL MA MSFT NET PANW V ZS
Liability absorption & underwriting of agent error	Someone must pay when the agent buys the wrong thing, cancels the wrong policy or is hijacked; consumers will not delegate until they cannot lose	Chargeback parity products, agent-error insurance riders, issuer-funded guarantees as a customer-acquisition weapon; balance sheet plus actuarial data becomes the moat	gains pricing power	ACGL AIG AIZ AJG ALL AON AXP BRO CB CINF EG HIG MA MRSH PGR PYPL TRV V XYZ
Credential custody & the identity/preference graph	The agent needs standing access to email, calendar, bank, health, vehicle and payment credentials; whoever holds them holds the switching cost	Early credential capture at the OS, bank or super-app layer — the default-payment-instrument lock-in of this decade	becomes scarce	AAPL AMZN FICO GOOG GOOGL INTU MA META MSFT PYPL V
On-device / edge inference silicon	Always-on background monitoring and private action require cheap, low-latency, privacy-preserving local compute rather than metered cloud tokens	NPU content per device, agentic power management, secure enclaves for mandate storage; inference margin migrates toward device silicon	gains pricing power	AAPL ADI AVGO MCHP MPWR MRVL NXPI ON QCOM SWKS TXN
Workflow data & systems of record that the agent must act through	An agent can only act where it has structured state: invoices, payroll, tickets, CRM, tax, benefits	Repricing from per-seat licence to per-action / outcome take-rate — the same revenue base at higher realized value	gains pricing power	ACN ADP APPF CRM IBM INTU NOW ORCL PLTR SNOW TYL
Agent-readable commerce & transaction inventory	Merchants must expose price, availability and cancellation terms to machines or lose basket share to those that do	Scale retailers with clean catalogues, transparent pricing and logistics win; opaque, friction-monetized merchants lose. Agents optimize on price and reliability, which favors private label and warehouse economics	gains pricing power	ABNB AMZN BKNG CRCL DASH EBAY FIS FISV GPN MA PYPL UBER V WMT XYZ
Discovery & persuasion (ad-funded attention)	Advertising assumes a human eye at the decision point; agents do not see banners, do not build brand affinity and do not pay for convenience defaults	Only those who own the agent itself (and can sell placement inside its choice set) retain the rent; independent ad-tech intermediaries lose the impression	loses pricing power	APP CMCSA DIS FOX FOXA GOOG GOOGL META NWS NWSA OMC PSKY RDDT TTD
Retention-flow and inertia-dependent subscription revenue	Consumers delegate cancellation and renegotiation; churn that was suppressed by friction becomes realized churn	Short/underweight businesses whose ARPU depends on forgotten subscriptions and auto-renewal inertia; conversely, genuinely-preferred services gain share	new risk	ADBE CMCSA DIS GEN INTU NFLX T TMUS
Travel & experience intermediation	Booking, rebooking, disruption management and price monitoring are the highest-ROI delegable tasks and the first to go agentic at scale	Whoever holds the PNR and the loyalty relationship keeps the customer; pure comparison/meta layers get disintermediated by the agent doing the comparing	new risk	ABNB BKNG CCL DAL EXPE HLT LUV MAR NCLH RCL TNL UBER
Inference capacity for continuous, background, per-user agents	Agentic workloads are persistent and event-driven rather than session-based — token volume per user rises an order of magnitude even as per-token price falls	Memory, optics, networking and power for inference-weighted (not training-weighted) buildout; contracted physical suppliers survive a financing event better than equity-funded compute landlords	becomes scarce	ALAB AMD AMZN ANET AVGO CIEN COHR CRDO DELL DLR EQIX GOOGL HPE MSFT MU NVDA ORCL SMCI VRT
Consumer protection as a bundled financial product	Scam detection, identity defence, account-takeover monitoring and dispute automation become a non-discretionary household line item, distributed by banks, carriers and insurers who profit from lower loss ratios	Subsidized distribution with high attach rates; the revenue is bundled, which is why metric leg (4) is the hardest to measure but the easiest to monetize	mandatory infrastructure	ALL AON AXP BAC C CB FICO GEN JPM MA PGR T TMUS TRV V WFC
Agent-initiated finance & settlement	Standing instructions require programmable, 24/7, reversible settlement with spend ceilings — the SEC tokenization exemption and stablecoin rails are the substrate	Brokers, custodians and exchanges that can honour machine-initiated, rule-bounded orders and prove authorization	gains pricing power	BLK CBOE CME COIN CRCL FIS FISV GPN GS HOOD ICE JPM MA MS NDAQ PYPL SOFI V XYZ

Second-order thesis

The first-order read — 'agents will do things for us' — is nearly consensus and nearly unpriceable. The second-order read is that this transition converts intelligence from the scarce input into the abundant one, and promotes a boring, balance-sheet-heavy asset to the scarce position: the right to be trusted with a credential and the capacity to absorb the loss when the agent is wrong. Every completed analogue in the reference class (credit cards, ACH autopay, robo-advice, autoland, spam quarantine, card-network fraud blocking) scaled on the same day the institution, not the user, took the downside; none scaled on the day the technology got better. That means the investable edge is not in model vendors or agent startups but in the three layers the mechanism forces into mandatory infrastructure: (1) authorization and identity — distinguishing my agent from an attacker becomes a per-transaction service that every bank, merchant and OS must buy, which turns cybersecurity from a discretionary IT line into a payments-rail cost of goods sold; (2) liability underwriting — the entity with the actuarial data and the balance sheet to offer chargeback parity for agent transactions acquires customers at negative marginal cost and locks in the credential, which is why the card networks and the largest issuers, not the model labs, are the likely toll collectors; (3) on-device compute — always-on background monitoring and protection is a volume business at near-zero marginal cost per event, which pulls inference economics toward the handset and away from metered cloud rental, making the terminal value of centralized GPU rental more contestable than the current order book implies while making memory, optics and NPU content structurally more valuable. The overlooked negative is the reflexive one: delegation is a solvent applied to every revenue stream that was secretly friction revenue. Auto-renewal inertia, retention call-centre saves, price opacity, brand recall at the shelf, and the entire ad-funded discovery stack all assume a tired human at the decision point. Remove the human and a large slice of consumer-facing gross margin — in streaming, telecom, insurance renewal, timeshare, ad-tech and mid-tier branded goods — is revealed to have been a tax on attention rather than a payment for value. Scale winners with transparent pricing and clean machine-readable catalogues (AMZN, COST, WMT, TJX) absorb that share; the losers do not get a warning quarter, they get a step-change in churn once the agent is default-on. Finally, the dossier's dominant macro fact — electrons as the binding constraint — compounds here rather than competing with it: agentic workloads are persistent and event-driven, not session-based, so per-user token consumption rises even as price per token falls, which keeps memory, power and interconnect demand firm through the exact financing air-pocket that will kill the equity-funded compute landlords. Own the constraint and the liability, rent the intelligence.

Losers

·Standalone consumer chat/content subscriptions — the $20/month 'ask it anything' product becomes a bundled feature with zero marginal price
·Independent ad-tech and attention intermediaries whose impression depends on a human at the decision point (TTD, APP, RDDT); search and social ad rents compress unless the platform owns the agent
·Subscription businesses whose ARPU is propped up by cancellation friction and forgotten auto-renewals (NFLX, DIS, CMCSA, ADBE, TNL, gym/telecom-style recurring models)
·Travel and comparison metasearch layers that exist to do the comparing — the agent absorbs that function; PNR- and loyalty-holders survive, aggregators without inventory do not
·Call-centre BPO, retention-desk and outbound-telephony labour economics
·Mid-tier consumer brands whose premium rests on shelf recall and habit rather than measurable performance — agent-optimized reordering favours private label and price-transparent scale retail (pressure on PG, KO, MDLZ, NKE brand rents)
·Per-seat software pricing models that cannot demonstrate per-action labour substitution
·Pure-play agent startups without credential custody, distribution or a liability balance sheet — they become features of whoever holds the payment instrument
·Equity-funded compute landlords and second-tier neoclouds exposed to a financing event rather than a demand event
·Merchants and service providers who monetize opacity — price-obscured insurance renewals, variable-rate utility plans, resort fees, surprise billing

Probability history

2 observations · created 2026-09-25 · 1 monthly update

posteriorcurrent range 45–78%reference-class prior

Updates log

Monthly reviews, newest first · what changed, and whether the new evidence argued for moving the probability

Date	Before		After	Δ	New	Retired	What changed	Should change mind?
2026-10-01	57.2%	→	61.7%	+4.5	8	1	·Nvidia's Open Agent Safety Platform expanded from announcement (09-28) to an emerging de facto standard (09-29/09-30): 100+ partners, per-agent identity with DigiCert, IBM identity/storage integration, millisecond containment, Vera CPU for agents on CoreWeave, Arm partnership. The identity/delegated-authority/audit-log layer the thesis names as the binding constraint is now being productized and sold as a recurring budget line — but enterprise-first, not consumer-first. ·Washington chose a voluntary 'morally binding' AI accord over statutory safety legislation. This removes near-term regulatory drag but also removes the most plausible fast path to a legislated consumer liability cap; the agent-liability standard will now be set privately (networks, issuers, insurers) or not at all. ·Macro cost-of-capital deteriorated materially rather than stabilized: 10-year at 5.33–5.34% (highest since 2002), 30-year through 5.6%, UK 30-year gilts above 6%, dollar at a 16-month high, Brent near $100 on a refining/distillate crisis, and the term-premium move is occurring despite a softer core PCE (3.0%). Micron beat on revenue and was not rewarded. ·A new and specific AI credit-cycle channel appeared: Nvidia's financing guarantees to chip buyers plus rising neocloud debt are drawing credit-risk scrutiny; Anthropic's IPO filing carries an existential-risk warning and OpenAI delayed its IPO. Cost of capital for model-layer narrative assets is rising while distribution-owning incumbents are unaffected — mildly reinforcing the thesis's 'delegation rides incumbents' claim and its financing-air-pocket risk simultaneously. ·Consumer stress hardened (confidence 81.9, a 12-year low; 9.3m student-loan borrowers in default; spending funded by a top-quintile AI wealth effect), which cuts both ways: it suppresses discretionary standalone AI subscriptions while raising the appeal of agents that cancel, dispute and protect. ·Still zero of the five required bank/card-network consumer liability frameworks for agent-initiated transactions. The thesis's single most important confirmation signal has not moved in six weeks.	Net: essentially no move, a hair up at most — I would hold at ~57% rather than defend it upward. The genuinely new positive is real but indirect: per-agent identity, delegated authority and tamper-evident logging went from research problem to purchasable enterprise stack in one month, which is the thesis's own named binding constraint being worked on by the parties who would have to underwrite it. That is mechanism confirmation, not adoption confirmation. Against it sit three new negatives of comparable weight: the voluntary federal accord closes the fastest legislated path to a consumer liability cap, leaving the thesis dependent on networks volunteering to absorb loss against their own P&L; the term-premium shock is now persisting through soft inflation data, which raises the odds of a one-to-two-cycle delay; and the household balance sheet (confidence at a 12-year low, 9.3m student-loan defaults) argues the consumer cannot pay directly, so everything must route through bundling, which is slower and controlled by fewer gatekeepers. Most importantly, the single highest-information signal — metric leg (5) — has not moved in six weeks and now has a structural earliest date of April 2027. The thesis is intact and its mechanism is getting cleaner, but its timeline is drifting, and the gap between the 2031 horizon and the real pace of liability socialization is the thing I would downgrade on if April 2027 passes empty.