The Next Fifteen Years

A forecast built from first principles
Section future / 03-domains / cognitive / startups / ideas-institutions.md

Ideas - institutions, risk, and human formation#


Contents

From ~2028 the binding constraints are institutional, not technical (thesis). These ideas sell into licenses, liability, underwriting, and the novice→expert pipeline - the clocks that do not move at model-release speed.

1. AI liability insurance that is a real product line#

Why now. Underwriters will price AI risk before legislatures settle it (insurance, U6). Early policies are soft-market noise; the company that builds loss data, exclusions that match reality, and risk engineering wins the hard market.

Build this. MGA or carrier partnership for AI E&O / product liability with mandatory logging/checkpoint standards, vendor indemnity parsing, and reinsurance pathways. Publish rate structures that become the de facto deployment code.

Own this. Underwriting data, binder relationships, and the inspection/standards process (Hartford Steam Boiler pattern).

Not this. A chatbot that fills ACORD forms.

Dead if. Model providers indemnify customers so broadly that private AI liability stays a thin rider forever - watch enterprise contract terms (C6).

2. Correlated-risk measurement for model monoculture#

Why now. Finance and cyber already misprice correlated failure; foundation-model concentration makes portfolio-level herding a systemic object (finance). Nobody sells a clean measurement of "how many of your vendors share a model scaffold."

Build this. Software and advisory that maps model/provider concentration across a firm's stack and a fund's book, stress scenarios for synchronized failure, and remediation (diversity requirements, kill switches). Sell to CROs, regulators, and reinsurers.

Own this. Cross-client concentration graph and scenario libraries.

Not this. Generic GRC text generation.

Dead if. Supervisors mandate and provide the measurement themselves as a public utility - you may still implement.

3. Outcome-priced work with escrow and dispute#

Why now. Seat pricing breaks when agents do the work (software, B5). Buyers want outcomes; sellers fear unbounded liability; both need attribution.

Build this. Contracting + metering + escrow + arbitration for agent and human-AI work products: definitions of done, evidence packs, chargebacks, and optional insurance wrap.

Own this. The marketplace or rail where outcome contracts clear, and dispute precedent data.

Not this. Usage-based billing lipstick on a seat product.

Dead if. Reliability stays too low for anyone to accept outcome risk - then seats persist and this rail starves (still a signal).

4. Compliance operations rebuilt as AI-native infrastructure#

Why now. Monitoring, change-tracking, audit trails, and multi-jurisdiction licensing are text-and-deadline machines currently staffed with expensive humans (YC RFS: AI-native compliance). Complexity grows faster than revenue for multi-state and multi-country firms.

Build this. System of record for licenses, policies, controls, and evidence - agents draft and watch; humans and officers sign. Start in one painful regime (money transmission, healthcare privacy, contractor licensing).

Own this. The compliance system of record and regulator-recognized evidence formats.

Not this. "Ask the bot what GDPR means."

Dead if. Incumbent GRC suites ship good enough agent layers and keep the customer relationship - integrate or out-vertical them.

5. Apprenticeship products that preserve the pipeline#

Why now. AI eats the junior task set that trained experts (Game 4, U3). Firms that stop hiring juniors save a year and lose a decade. Software is the canary; law and consulting follow.

Build this. Structured practice environments with grounded feedback (tests, shadow cases, simulated clients) plus employer partnerships that restore a paid novice rung. Sell to firms as talent pipeline insurance, not to students as another course.

Own this. Employer offtake, credential trust in the profession, and longitudinal competence data.

Not this. Uncritical chatbot tutoring that accelerates homework collapse (education).

Dead if. Dense AI mentorship alone compresses novice→expert without institutional redesign - then pure product wins and employer partnerships matter less (still build the product).

6. The Primer path (adaptive foundations for children)#

Why now. Tutoring quality was never scalable; models make adaptive practice possible at consumer prices (YC RFS: The Primer). The corpus warns: education's bottleneck is also childcare, motivation, and credentials (education) - so product design must not assume school vanishes.

Build this. Mastery product for reading, writing, numeracy with parent/teacher controls, long-horizon learner models, and evidence that beats high-dose human tutoring on defined skills. Expand upward only after fundamentals work.

Own this. Learner longitudinal data, parent distribution, and (eventually) school district procurement relationships.

Not this. A skinnable chat wrapper on a homework database.

Dead if. Open models + parent diligence make paid foundations software a race to zero and schools refuse to integrate - hard market, not impossible niche.

7. Law: authorization and negotiation layers#

Why now. Drafting compresses; standing, negotiation, and signed accountability do not (law). Consumer and SMB legal need is vast and underserved; full firm replacement is gated on liability.

Build this. Workflow products where models prepare and humans (or insured entities) authorize - filings, contract negotiation assistance with explicit principal, court or agency submission rails where rules allow.

Own this. Bar-compliant delivery structure, E&O, and channel into SMBs or in-house teams.

Not this. "Replace your lawyer" autonomous advice with no balance sheet.

Dead if. Unauthorized-practice enforcement blocks the wedge or full provider indemnity enables true autonomy and skips your human-in-the-loop design - different deaths, both real.

8. Synchronous evaluation for hiring and credentials#

Why now. Resumes, take-homes, and diplomas are cheap to generate. Live, time-boxed, proctored, job-relevant performance is expensive to fake (Game 5, education).

Build this. Assessment platforms that run realistic work trials with integrity guarantees, portable score reports employers trust, and anti-proxy design. Verticalize (software, nursing skills, trades).

Own this. Employer trust in the score and the item bank that stays ahead of generators.

Not this. Another async LeetCode clone with an AI proctor sticker.

Dead if. Employers decide formal evaluation is hopeless and hire only from trusted networks - then you need to become infrastructure for those networks.

9. Small-software cloud (deploy the one-user tool)#

Why now. Agents make bespoke internal tools cheap to write and still annoying to host, auth, and share (YC RFS: cloud for small software). This is formation infrastructure for every company, not only startups (formation).

Build this. Opinionated runtime: auth, permissions, secrets, audit, one-click share, cost caps - Google-Docs-easy for agent-written apps, with enterprise guardrails.

Own this. Default deploy target for internal agents and the admin relationship.

Not this. Yet another Kubernetes control plane for the same Big Software complexity.

Dead if. The major clouds ship a true "small software" tier that deletes the complexity tax - race to become the layer or the vertical specialist.

10. Defense and dual-use autonomy components#

Why now. Acquisition is opening to commercial modular systems; cost-per-kill, sensors, resilient logistics, and attritable drones are explicit demand (YC RFS: American defense). The corpus treats warfare as cheap ground truth and fast iteration (warfare) - ugly, real, and complement-heavy (manufacturing, test ranges, clearances).

Build this. Specific components with open interfaces, extreme-environment validation, and a path through actual procurement - not slideware "AI for defense."

Own this. Manufacturing, test data, program relationships, and clearances.

Not this. A fine-tuned model demo without hardware or a buyer.

Dead if. You cannot get a real user (service or prime) onto a test event within two years - the market is relationship-gated; treat that as the product problem.


Related: Insurance · Game 4 · Education · Law · Uncertainty 3 · Uncertainty 6

View markdown source

select · Enter open · Esc close