Unsolved problems, evidenced

Why it was considered unsolvable

The party operating the ad-decisioning engine and the party measuring its success are typically the same party, or use the same underlying signals — a self-grading conflict of interest, not a lack of measurement sophistication. Rigorous randomized geo-holdout testing is slow, expensive, and against the platform's commercial interest.

What the evidence shows

The one independent check found — a study of 640 Meta geo-lift tests — found Advantage+, Meta's most-automated campaign type, beating manual campaigns in only 42% of tests, and delivering 12% lower incremental return even when it won #54. That single independent finding stands against dozens of platform-reported, self-selected case studies making the opposite claim #48 #50 #51 — every one of them the ad server reporting on its own campaign.

The gap

No standardized, industry-trusted, external-audit layer for incrementality exists — nothing that sits outside the ad server's control the way a financial auditor sits outside the company it audits. No source in a 112-claim evidence base showed any platform submitting its AI-optimized campaigns to independent adversarial testing as standard practice.

Why it's an opportunity, not just a flaw

Advertisers are already cutting spend they cannot audit #53 and 62% of practitioners name black-box platforms their top challenge #55 — the demand side of an audit market already exists. Whoever builds the trusted, standardized holdout methodology becomes the layer every AI-spend decision has to pass through.

Verdict

Open opportunity— the measurement infrastructure is unbuilt; the conflict of interest it would resolve is adjudicated fact in the auction layer and documented pattern in the measurement layer.

Deep dive: Ad Servers Research Materials → Unsolved Problems

Why it was considered unsolvable

Whoever operates the infrastructure clearing the auction has both the ability and the incentive to see more of it than the parties transacting on it. This isn't a modeling problem; it's a conflict of interest baked into who owns the pipes — and it is now adjudicated fact: a US court found Google illegally tied its publisher ad server and exchange #71, and the EU found Google informed its own exchange of rival bids it needed to beat, since at least 2014 #73.

Binding constraint

Structural information asymmetry held by the infrastructure owner, which that owner has no commercial incentive to surrender voluntarily.

Does AI remove the constraint?

No. The AI "transparency" features shipped since — reporting expansions #57, conversational copilots #1 — add visibility into what already happened after the auction cleared. And the agent-to-agent protocols marketed as "negotiation" turn out, in their own technical documentation, to be discovery-then-approve sequences with no counter-offer step, executing through conventional RTB underneath #98 #99. Richer discovery on an unchanged auction is automation of the lookup, not removal of the asymmetry.

Verdict

Automation in disguise— for everything shipped as "AI transparency" to date. The strongest force acting on the actual constraint is regulatory, not technological: remedies rulings pending on two continents.

Deep dive: Ad Servers Research Materials → Unsolved Problems

Why it was considered unsolvable

Recognizing the same user across publishers and platforms requires either a shared ground truth no single company controls, or invasive tracking that browsers and regulators are closing off. No modeling skill substitutes for parties who won't share data.

Binding constraint

A coordination and consent problem across many independent controllers of the user relationship — not a technical resolution problem.

Does AI remove the constraint?

No — and the six-year browser-native attempt just failed: Google reversed its cookie-deprecation plan and retired more than ten Privacy Sandbox APIs for low adoption #76 #77. What scaled instead is deterministic, consent-based identity — UID2 #78, LiveRamp's graph, which Publicis is buying outright for $2.17B #81. AI sits on top of that already-solved consent layer as an access interface; it is not the mechanism that resolves identity. The sharpest confirmation is negative: Morningstar's moat downgrade of The Trade Desk reasoned specifically that UID2, as an open multi-tenant standard, is not an exclusive asset #101 #106.

Verdict

Automation in disguise— identity was solved by consent and acquisition, not by AI. The moat is ownership of the consented graph, and open standards branded "proprietary" are not the same asset.

Deep dive: Ad Servers Research Materials → Unsolved Problems

Why it was considered unsolvable

An adversarial arms race, not a static classification problem — any detection signal, once learned, gets reverse-engineered and evaded. A fraud model decays the moment it succeeds.

Binding constraint

Adversarial adaptation speed. The defense's historical tooling advantage disappears once both sides hold similarly capable AI.

Does AI remove the constraint?

By the defense's own published numbers, no — the offense is accelerating faster. One verification vendor's telemetry shows bot traffic growing ~8x faster than human traffic, with AI-agent traffic up 7,851% year over year #83 (vendor-reported, flagged as such). The industry's newest response is triage, not elimination: classifying declared legitimate AI agents apart from evasive scrapers #84. The standard waste baseline — only $0.36 of every DSP dollar reaching the consumer #82 — has not been shown to improve despite years of AI detection tooling deployed at scale.

Institutional-capability formulation

Genuine constraint removal would change what gets paid for — restructuring the incentive that makes fraud profitable — rather than detecting bad impressions faster after the fact. No one in the evidence base is attempting that.

Verdict

Automation in disguise— and currently losing the acceleration race it entered. The domain's clearest case of AI scaling both sides of a problem at once.

Deep dive: Ad Servers Research Materials → Unsolved Problems

The structural fact

In every agent-to-agent advertising deal documented in this domain's research, both agents represent institutions — a platform's seller agent, an agency's buyer agent. The consumer, whose attention is the actual product, remains what they have always been: inventory. Individuals have never had a mechanism to negotiate collectively over the terms on which their attention and data are sold.

The proposal

The same protocol infrastructure now being standardized for institutional agents could, in principle, let individuals' agents aggregate into consortia — millions of consumer-side seller agents negotiating as a bloc. That would not automate the existing market; it would seat a new party at the table — the first genuinely new market structure this industry would have seen since the auction itself.

Status

This is the authors' original proposal (Sudhir Patavardhan, 2026-08), offered as such — nothing in the domain's 112-source evidence base shows any vendor building it, and it was deliberately not researched as a market claim to validate. It is logged here because the plumbing being standardized today is exactly the plumbing that would make it possible tomorrow.

Verdict

Authors' proposal— forward-looking, unbuilt, and flagged for anyone who wants to build it.

Deep dive: the report's Blindspots & open fronts section

What's now possible

Provenance ground truth exists at manufacture: the Aura Blockchain Consortium (LVMH, Prada, Richemont/Cartier, OTB) reports 40M+ products carrying a verifiable digital identity assigned at origin #22, and the EU's Digital Product Passport mandate is on track to make garment passports compulsory for every item sold into the EU, realistically around 2028 #75.

The gap

No evidence was found that anyone queries that registry at the point of resale — the thrift stores, consignment desks, and resale marketplaces where secondhand counterfeiting actually lives. TikTok Shop's authentication program runs on Entrupy-style post-hoc inspection, per the vendor #21. The constraint was removed upstream; the removal sits unused downstream.

Binding constraint

Not a research problem — an integration and adoption problem. The unbuilt piece is the query/verification layer at the point of resale: marketplace integrations, consignment tooling, whatever turns a registry lookup into a habit.

Why it's an opportunity, not just a flaw

The registry exists, the EU mandate will make passports universal, and whoever builds the resale-side verification layer completes the network — provenance's missing last mile — and captures the value of being the default answer to "what is this garment?" This entry was promoted to the public registry directly from the fashion article's adversarial grill session (2026-08) — the first entry sourced that way.

Verdict

Open opportunity— flagged for problem solvers and investors. Watch item: the final text of the EU textiles delegated act (if self-declared brand databases count as compliant passports, shared verifiable provenance loses its premium).

Deep dive: the report's Blindspots & open fronts section · Fashion Research Materials

Why it was considered unsolvable

Verifying a luxury good's authenticity always required scarce human expert attention, which cannot scale as resale/secondhand volume grows. Counterfeits are estimated at ~2.5% of global trade, ~$464B #36.

Binding constraint

Expert attention doesn't scale, and authenticity historically could only be established after the fact, by inspection — there was no ground truth to check against.

Does AI remove the constraint?

Split evidence, and the split is instructive. Entrupy-style computer vision #21 automates the expert's pattern-matching — faster inspection, but still probabilistic, still after-the-fact, still purchasable by any single reseller. That's automation within the old model. The Aura Blockchain Consortium's digital product passports #22 — co-founded by LVMH, Prada Group, Richemont/Cartier, and OTB, covering 40M+ products — instead assign each item a verifiable digital identity at the point of manufacture. Authenticity is no longer inferred after the fact from an image; it's established at origin. That removes the actual constraint.

Evidence anyone is attempting it

eBay's Certilogo acquisition #20; Entrupy's adoption including TikTok Shop's handbag program #21; Aura's 40M+ products across four major luxury groups #22.

Institutional-capability formulation

Intelligence should live upstream, at manufacturing/point-of-origin, as shared infrastructure across a consortium of brands — which is exactly why it's hard to copy. It requires industry-wide cooperation, not a vendor contract a single competitor could also sign.

Verdict

Genuine invention candidate— Aura-style provenance infrastructure. The CV-inspection approach (Entrupy/Certilogo) is automation in disguise by direct contrast within the same problem.

Deep dive: Fashion Research Materials → Unsolved Problems

Why it was considered unsolvable

Fashion demand is volatile and trend-driven, but production lead times (often months, often overseas) force commitment decisions long before real demand is known. The mismatch drives markdowns and waste — an estimated 2.5–5B excess items worth $70–140B in 2023 alone #45 (an industry estimate, not an audited figure — see the domain's Known Gaps).

Binding constraint

A slow feedback loop — the gap between a demand signal appearing and a garment reaching the shelf — not a lack of prediction skill per se.

Does AI remove the constraint?

Mostly not yet. Most deployed "AI forecasting" #8 improves the accuracy of a prediction made months in advance — a better guess within the same slow production cycle. That's accelerating work within the constraint, not removing it. 75% of executives are prioritizing AI forecasting for exactly this #45, which is telling: the industry default is "predict better," not "shorten the loop." The one model that would actually attack the constraint — short lead times plus AI reading in-season sell-through — is the Inditex/Zara story, but the specific KPI figures behind it are unverified trade-press numbers #10.

Evidence anyone is attempting it

PVH×OpenAI's demand-planning scope #1 #2 is framed as forecasting-layer automation, not lead-time compression.

Institutional-capability formulation

Intelligence shouldn't live in a better one-shot forecast — it should live in a continuous sense-and-respond loop spanning design, sourcing, and store-level sell-through, short enough that decisions can be revised mid-season. That requires re-architecting the supply chain itself, not adding a forecasting model on top of an unchanged one.

Verdict

Automation in disguiseas most commonly deployed today — with a plausible-but-unconfirmed invention-candidate exception. Needs primary-source verification before that exception can be asserted confidently.

Deep dive: Fashion Research Materials → Unsolved Problems

Why it was considered unsolvable

Individual body-shape variation is combinatorially large; standard sizing charts are population averages built for manufacturing convenience, not individual fit. Retailers historically had no cheap way to capture a shopper's actual 3D body data.

Binding constraint

No scalable, cheap way to capture individual body data — fit was always a guess against an average.

Does AI remove the constraint?

Partially, and the evidence is genuinely encouraging rather than automation dressed up as invention: Zalando's AI-built 3D avatars report up to a 40% return-rate reduction in an early pilot #6; ASOS attributes part of a 160bps returns-rate cut to try-on tech #7. A shopper's actual body shape is now informing a fit decision that used to be pure average-based guesswork. Caveat: bias evidence found elsewhere (runway/imagery representation narrowing to 0.3% plus-size #25; generative models reproducing training-data body-type skew #26) raises an open, unverified question of whether these 3D body models and their training data are themselves representative. No study was found answering this specifically for fit/sizing algorithms.

Evidence anyone is attempting it

Zalando #6, ASOS×AIUTA #5 #7, Stitch Fix Vision #11.

Institutional-capability formulation

The compounding version isn't "sell a try-on widget" — that's currently automation, purchasable from vendors like AIUTA by any retailer. It's closing the loop from fit/return data back into design and pattern-making, so garments themselves are shaped by accumulated body data over time. No evidence was found of any retailer doing this yet.

Verdict

Promising but unproven— real signal, currently deployed as commoditized automation; graduates to a genuine invention candidate only if a retailer closes the design loop, which hasn't been observed.

Deep dive: Fashion Research Materials → Unsolved Problems

Why it was considered unsolvable

Automated UI test scripts encode the exact steps of one implementation. Every UI refactor breaks them, so teams either drown in maintenance or quietly stop trusting their test suite.

Binding constraint

Scripts encode implementation details, not the tester's actual intent — so there's nothing stable for the script to track when the implementation changes.

Does AI remove the constraint?

Yes — when intelligence lives in understanding intent (what the test is actually trying to verify) rather than in a fixed sequence of clicks and selectors, the test can survive a UI refactor because it's no longer coupled to the old implementation's shape. The obvious path — AI generating test scripts from natural language — was a legitimate productivity win, but generated scripts couldn't keep pace with application change: same constraint, faster output.

Institutional-capability formulation

Intelligence should live in intent understanding, not script replay. This is the reference case the Two Hills framework itself was built from, kept here as the calibration example for every domain that follows.

Verdict

Solved— genuine capability, not automation dressed up as one.

Deep dive: Methodology — the Two Hills framework's reference case. This entry predates the domain pipeline, so it links to Methodology rather than a domain page.