AI has filled the most critical gap, and this batch of encryption projects deserves a second look.

Written by: Rex St. John

Translated by: AIdidiaoJP, Foresight News

Most good crypto projects did not fail due to a wrong direction, but because of three fundamental infrastructure gaps: judgments were too costly, there was no stable currency, and settlements were too slow. After 2023, all three have changed.

From 2014 to 2019, a large number of ambitious projects emerged in the crypto industry: on-chain courts, universal oracles, data markets, computing power markets, micropayments, identity systems, curation registries, insurance, traceability, job markets, autonomous organizations... aiming to transfer the entire economy onto the chain. Most of them have already disappeared.

The problem was not the ideas themselves, but that the three things they relied on did not exist at the time: mechanisms to solve vague real-world problems at low cost, stable units of account, and chains fast enough to settle micropayments.

Old Bottleneck: All Contracts Ultimately Require Human Judgment

Smart contracts can only verify what a computer can verify—signatures, balances, hashes. But for almost every meaningful contract, the key issues are off-chain: Has the work been completed? Is this statement true? Was the flight delayed?

Every project ultimately had to rely on humans for adjudication: token holder voting, staked jurors, curators. Yet humans are slow, expensive to hire, easily bribed in small groups, and unlikely to go to court for a $4 dispute. As a result, the market could never grow large, and the economic models could only be propped up by tokens. When the cycles of 2018 and 2022 came, projects collapsed.

Four Changes, All After 2023

Judgments have become cheaper. Large language models can read contracts, deliverables, and evidence in seconds for just a few cents and make rulings. Using different vendor models to form a jury, it is harder to bribe than a single human juror, and they are never absent.

AI agents have become users. AI agents are working, buying services, and signing agreements. They need contracts, payments, and credibility more than humans do; transaction amounts are small to the point that humans cannot arbitrate.

Stablecoins are in place. By mid-2026, the supply will be around $310 billion (Spark data), allowing contracts to be paid in dollars instead of volatile governance tokens.

Chains have become faster and cheaper. Solana confirms transactions in less than a second with fees of less than a cent, making per-transaction payments and arbitration on-chain viable.

Finding project patterns: An old idea, whose bottleneck was "a trusted party is needed for judgment." This is precisely the bottleneck that AI has eliminated.

Fifteen Directions, Ranked by Revenue Potential and Readiness

The income hierarchy is the author's estimate of the market share that new entrants might capture, not a prediction.

1. Online Commercial Courts (Kleros, OpenBazaar Arbitration)

How it used to work: Staking, randomly selecting jurors to resolve disputes between strangers, replacing platform customer service.

Why it got stuck: Each case requires manpower and time, making costs too high, suitable only for disputes over hundreds of dollars. Volume could not get up, and small juries were easily bribed.

What AI changes: Multi-model juries can decide $4 disputes in seconds for just a few cents, with each vote and reason public. Transactions between AI agents will generate millions of small disputes unmanageable by humans.

Revenue potential: Large. In 2025, merchants lost about $33.79 billion due to 261 million chargebacks (Mastercard data); the online dispute resolution software market is about $1.76 billion (Mordor Intelligence). Even charging a small fee for a fraction of the rulings represents real business.

2. Problem Oracles Without APIs (Augur, Truthcoin, Futarchy)

How it used to work: Allowing anyone to bet on any real-world issue, with token holders reporting the results.

Why it got stuck: Subjective questions (e.g., "Was the bill effectively passed?") require slow, manipulable human voting. Liquidity has gone to sports and elections—where APIs already have answers.

What AI changes: Model juries read adjudication standards and information sources to make rulings on vague outcomes, providing reasoning for appeal. This opens up today’s long-tail market that goes unlisted—because no one can settle.

Revenue potential: Large. By October 2026, the transaction volume of prediction markets in the US is expected to reach $120.7 billion (DeFi Rate). Settlement is a common weakness across all platforms, and can be sold to all platforms.

3. Insurance Without Adjusters (Etherisc, Nexus Mutual, P2P Mutual Aid)

How it used to work: Parametric insurance automatically compensates when measurable events occur (flight delays, insufficient rainfall); mutual organizations allow members to share risks and vote on claims.

Why it got stuck: Parametric insurance can only cover events with clean data sources. Mutual claims determined by member votes are slow and hard to gather a group for, and members have conflicts of interest—every payout comes from the pool. Using tokens valued by price volatility for premiums scared off buyers and regulators.

What AI changes: Claims are decided by a neutral jury based on policy wording and evidence, with no interest in the outcome. Trigger conditions can be written (e.g., "Crop loss in this region, determined by reports from ministries or two media outlets"), judged by the jury from documents, significantly increasing insurable risks. Payouts are made in stablecoin.

Revenue potential: Large but slow, requiring licensed insurers or regulated mutual organizations as partners. The parametric insurance market alone is expected to be about $19.4 billion in 2025, with an annual growth of about 12% (GM Insights). A faster entry point is to sell claim adjudication to existing insurance companies.

4. Custodied Job Markets (Colony, Gitcoin, Bounties Network)

How it used to work: Posting tasks, putting money in escrow, and releasing funds upon acceptance.

Why it got stuck: Someone still has to determine whether the work is completed. This person is the poster, bringing back all disputes from freelance platforms.

What AI changes: Acceptance criteria are judged by a neutral jury against actual deliverables (codebases, reports, designs). AI agents are both workers and buyers, at a scale unmatched by human reviewers.

Revenue potential: Medium to large. Upwork’s revenue is projected to reach $788 million in 2025, supporting over $4 billion of work volume (Yahoo Finance). The native market for AI agents can charge lower rates and handle more, smaller tasks.

5. Data and Media Licensing (Ocean Protocol, Ujo, Mycelia)

How it used to work: Creators and data owners sold access rights directly, with royalties executed by contracts.

Why it got stuck: There are few buyers, and no one can check how data is used after being sold.

What AI changes: AI labs have become buyers. Transactions between Reddit and Google are reported to be around $60 million annually, and the deal between News Corp and OpenAI exceeds $250 million over five years (Quartz). Terms can be written (e.g., "Not to be used to compete with the licensor’s models"), with the jury checking against audit evidence.

Revenue potential: Medium to large, but currently concentrated among a few major licensors. The opportunity lies in the long tail: millions of small creators unable to negotiate on their own.

6. Machine Payments (BAT, Lightning, IOTA)

How it used to work: Pay per article, per API call, per kilowatt-hour, with a few cents as tiny fees.

Why it got stuck: People dislike approving small payments, with fees and price volatility eating into profits.

What AI changes: AI agents don’t mind paying $0.26 a hundred times a day. The x402 standard is expected to settle about 198.9 million transactions, equating to approximately $52.7 million by September 2026 (TRM Labs).

Revenue potential: Currently small, but potentially large in the future. TRM's own screenings find that most transaction volumes are not real agent business. Payment rails are in place, what's missing is the reason to trust sellers—that involves the first and seventh ideas.

7. Reputation and Deposit Alternatives to Collateral (Bloom, ETHLend)

How it used to work: On-chain credit histories were used for uncollateralized lending.

Why it got stuck: Wallets can be created for free, and bad histories can be thrown out with the wallet, requiring lenders to still demand 150% collateral.

What AI changes: The judgment records of AI agents (completed deliveries, lost disputes) are difficult to forge, and the cost of abandonment is high; the jury's seizable deposits make commitments enforceable. This is the "agent deposit" template.

Revenue potential: Medium. The first clients are agent markets needing trust scores, not lenders.

8. Personality Verification (uPort, Sovrin, BrightID)

How it used to work: Portable identity proving you are the only real person.

Why it got stuck: There were no applications urgently needing it, and users were unwilling to go through the setup processes.

What AI changes: The demand has changed, not the solution. With robots and AI agents everywhere, having a "real person behind" suddenly becomes valuable. World reported 18 million verified real people by April 2026 (Decrypt).

Revenue potential: The market is large for winners, but competition is fierce and capital-intensive. The open gap is linking AI agents to accountable real people or companies.

9. Organizations with Enforceable Rules (The DAO, Aragon)

How it used to work: Using code and tokens to vote to run organizations.

Why it got stuck: Code cannot express the rules that people truly care about (e.g., “No payments to insiders unless disclosed”), with low voting rates and manipulation being easy.

What AI changes: Bylaws can be written in words, with every proposal checked by the jury against the treasury's signature. This is the “DAO Constitution” template.

Revenue potential: Small to medium. The treasury is large, but the space for charging tools is small; can be sold as risk control.

10. Curation Lists (Token Curation Registry, Steemit)

How it used to work: Staking tokens to maintain quality lists (advertisers, news sources).

Why it got stuck: The staking game rewards speculation rather than judgment, and voters do not read the entries.

What AI changes: Listing standards written in words are reviewed item by item by the jury, with reasons given, lowering the challenge costs.

Revenue potential: Small to medium; strongest when embedded in another product (app store, vendor registry, model market).

11. Content Provenance (Po.et, Ascribe)

How it used to work: Timestamping creative works to prove who created them first.

Why it got stuck: Timestamps prove a document's existence, not who created it, and no one verifies.

What AI changes: Generated media makes "Is this real and whose is it" a pressing issue for newsrooms, courts, and platforms. The jury can weigh the evidence chain, not just hashes.

Revenue potential: Medium, selling to publishers and platforms. Standards for content verification already exist and can be built upon.

12. Supply Chain Tracking (VeChain, IBM Food Trust)

How it used to work: Tracking goods from farms to shelves on a shared ledger.

Why it got stuck: Garbage in, garbage out. The ledger cannot determine if labels on boxes are real.

What AI changes: It has not inflated as much as hype suggested. Models can check document consistency (certificates, invoices, photos) on a large scale to catch lazy fraudsters. However, they cannot see inside boxes.

Revenue potential: Medium, but the sales cycle on the enterprise side is long. Start by checking compliance documents instead of ledger entries.

13. Tokenized Assets with Legally Binding Terms (Polymath, Harbor, tZERO, 2018 Security Tokens)

How it used to work: Stocks, real estate, and loans were tokenized for round-the-clock trading and instant settlement.

Why it got stuck: Tokens are just receipts. The true agreements (contracts, waterfall of payments, transfer restrictions, default procedures) exist in PDFs unreadable on-chain, and no court treats on-chain records as factual sources. Only simple assets have successfully been tokenized: by March 2026, about $23.6 billion of assets were tokenized, mainly government bonds and money market funds (Cointelegraph citing DeFiLlama).

What AI changes: Legal agreements themselves become contracts for jury reference. Whether a contract is violated, whether payments are due, whether transfers are allowed are judged by the jury against documents and evidence, with each ruling recorded on-chain. If the agreement specifies that this jury’s ruling is binding arbitration, the results can have legal force—arbitration awards are enforceable under the New York Convention in most countries.

Revenue potential: Large. Fees can be charged for the issuance and servicing of private credit, invoices, and real estate—these are the complex assets that have been overlooked in tokenization so far. Cannot be delivered individually: requires licensed issuers and lawyers to draft arbitration terms.

14. Shared Infrastructure and Computing Power Markets (Golem, iExec, Helium, Filecoin, Akash, io.net)

How it used to work: Tokens incentivized strangers to provide GPUs, storage, wireless coverage, or map data, sold cheaper than the cloud.

Why it got stuck: Many attempts have been made, but it still hasn't taken off. Token rewards bought supply, but paying demand did not materialize. Messari estimates the total on-chain revenue for the entire sector will be about $72 million by 2025, corresponding to about $10 billion in token market value (Decrypt citing Messari). The fundamental issue is trust: buyers cannot judge whether unknown suppliers genuinely completed tasks, maintained uptime, or delivered coverage; when issues arise, there is no adjudication. Enterprises purchase service guarantees, not tokens.

What AI changes: Two things. AI agents are the buyers, willing to assign tasks to the cheapest and most verifiable suppliers. The jury can adjudicate service claims based on evidence (logs, benchmarks, reasoning proofs), forfeiting the supplier's deposit for non-compliance—this is the "agent deposit" template applied to hardware. This provides guarantees that these networks have never had.

Revenue potential: If the trust issue is resolved, it could be significant because the demand for reasoning is huge. However, the currently proven paying demand is small, so start by offering a service guarantee for one type of task, rather than launching a market right away.

15. Intellectual Property Backed by Real Contracts (Mediachain, Ascribe, IPwe, Story Protocol)

How it used to work: Songs, art, patents, and brands were registered on-chain, using smart contracts for licensing and royalties.

Why it got stuck: Tokens have never been legal rights. Ownership lies with copyright offices, patent offices, and signed licensing agreements; anything entering those systems requires lawyers and application fees for each piece of work. No one legally relies on on-chain records; licensing is still negotiated via email.

What AI changes: AI agents can now write real contracts and prepare documents. Agents draft licenses in plain legal language, prepare copyright registrations, trademark applications, or transfer documents, submit them on behalf of others, and then hash the signed documents and receipts onto the chain. The same contract then becomes an agreement executable by AI agents: the jury adjudges whether use complies with the license, with royalties flowing based on rulings. On-chain records and legal records finally say the same thing.

Revenue potential: Large. Licensing goods retail is set to reach $389.8 billion by 2025 (Licensing International), and AI-trained licenses become new buyers (see idea 5). The opportunity lies in long-tail creators and small inventors who cannot afford lawyer fees for every application. Application services require lawyer review to avoid unauthorized practice, so start with copyright registrations and licenses that individuals can submit themselves.

How to Choose, and Possible Issues

Selecting a direction: The old bottleneck is human judgment, and today people are already paying for poor judgments. The focus is on: online commercial courts, prediction market settlements, and insurance claims.

Testing before taking action:

• Someone is already paying for slower human versions (chargeback teams, claims adjusters, market settlement agents).
• The problem can be answered with evidence visible to the jury. If a physical box needs to be checked, AI cannot help.
• It can run without tokens. Each idea above partially failed due to economic models relying on tokens rather than customers.

Risks:

• Model juries may be wrong together. Models trained on the same source data share blind spots. Public reasons, retain paths for appeals, and track consistency with human reviews.
• Evidence can be fabricated. Jurors can only judge based on what they see; fixing evidence by hash and treating parties' own words as untrustworthy.
• Regulatory constraints still apply in insurance, securities, and lending. AI eliminates judgment bottlenecks but does not eliminate licensing.
• Demand may still be thin. Machine payments and shared infrastructure have viable payment rails today, but proven paying demand is small. Measure before expanding.