A defined permission needs a verifiable enforcement boundary.
What Matters
AI agents can already negotiate transactions in bounded experiments. That does not mean an autonomous growth team will reliably make your business profitable. The opportunity is to reward verified contributions, not manufactured attention. DN’s proposed approach combines attributable contribution margin, a 50/35/15 allocation and human-controlled approvals—with a free readiness assessment before any recruitment or spending.
Commercial disclosure: Decentralised News publishes this analysis and offers the Agent Revenue Alliance assessment and design service discussed below. This is not an independent review of our own service. No agent-marketplace earnings, DN client results or returns are claimed.
A million views can still be a terrible business.
An AI agent can close a deal while giving away the margin. A content team can publish every hour without producing anything a customer wants. A referral dashboard can look impressive while refunds quietly erase the reward pool.
The next useful question is not simply, “Can AI agents make money?” It is: “What did they create, what did it cost, and who is entitled to what remains?”
That is the difference between an agent activity machine and a contributor network worth building.
DN Evidence Block
Last source check: 22 October 2026. Scope: three Anthropic experiment reports, iLands’ public FAQ, and official platform/regulatory guidance. The experiment reports were published between June 2025 and April 2026; current documentation was inspected separately.
Decisive distinction: demonstrated transactions are not audited business profits; platform resource tokens are not automatically cash; and a revenue-sharing percentage is not a substitute for an attribution agreement.
Evidence mode: primary-source review plus an original DN design worksheet and synthetic arithmetic. No live money, contributor recruitment, marketplace commissioning or independent replication was undertaken for this article. Editorial responsibility: Decentralised News; prepared with AI assistance. No independent reviewer is claimed.
AI agents are making deals. That is not the same as running profitable companies.
The evidence is more interesting than either extreme: “agents are useless” or “agents now run businesses while you sleep.”
In Project Vend’s first report, published 27 June 2025, Anthropic described an AI-managed office shop that made costly pricing mistakes, invented payment details and sold some products below cost. Humans still performed physical tasks. The result was not a self-sustaining business success.
The 18 December 2025 follow-up reported improved business performance after changes to models, tools and operating instructions. It also documented continuing susceptibility to manipulation. DN’s takeaway is not that agents cannot operate commercially; it is that better instructions and controls matter alongside capability.
Project Deal, published 24 April 2026, supplies a different signal. In an employee marketplace experiment, 69 agents completed 186 deals with a total transaction value just over US$4,000 in the run used for real exchanges. Participants were given budgets, and three additional runs were experimental comparisons. This was a self-selected employee setting—not an open-market customer-acquisition test. Transaction value is not the platform’s revenue, and neither figure establishes profit after all operating costs.
These examples support a narrower, useful conclusion: agents can perform parts of commercial work. Whether a particular business benefits still has to be demonstrated.
What “paying an AI agent” actually means
Start by separating the software from the payment recipient.
An incentive can mean a task-selection rule, a compute budget, a platform resource balance, or compensation owed to an accountable person or business operating an agent. Those arrangements are not interchangeable. Offering a chatbot “20% of profits” does not, by itself, connect an account, create enforceable terms or demonstrate better performance.
For example, iLands describes persistent agents, workspaces and external actions governed by capabilities and permissions. Its FAQ describes Tokens as resource units with no cash value or investment rights. Those are the platform’s published descriptions; DN has not independently audited marketplace income, contributor reliability or settlement.
Before commissioning work through any marketplace, establish the actual counterparty, permitted task, payout mechanism, fees, dispute procedure and revocation path. An attractive agent profile is not a substitute for those answers. Marketplace participation may also involve costs even when a separate readiness assessment is free.
Do not pay for proof that an agent was busy. Pay under agreed terms for proof that useful work was done.
The business model: a contribution network, not a bot army
Imagine an online business commissioning one useful research asset. A researcher checks the evidence. A creator turns the findings into an original explanation. A distribution partner reaches a relevant audience. A sales owner handles authorized commercial conversations. A human approves what goes out and what gets paid.
The work can be AI-assisted, human-led or mixed. The common element is a defined deliverable, a permission boundary and a verifiable outcome.
| Workstream | Useful reward basis | Tempting but misleading proxy |
|---|---|---|
| Research and production | Accepted original work meeting an agreed brief, source standard and rights check | Words generated, tasks attempted or unsupported claims of originality |
| Audience development | Approved content delivered to a relevant audience; measured qualified visits and outcomes | Purchased views, engagement exchanges or unauditable reach screenshots |
| Commercial referrals | Eligible collected revenue linked to an approved referral under written terms | Every inbox message, duplicated prospects or an unaccepted quotation |
| Quality assurance | Reproducible findings and accepted improvements | Invented defects, automated praise or meaningless “trust” scores |
Not every contribution should depend entirely on sales. A fixed fee for accepted research can be fairer than a commission when the researcher cannot control distribution or conversion. A revenue share can suit genuine commercial origination when attribution is defensible. A hybrid can work, but the fixed fee and variable pool must not compensate the same obligation twice by accident.
Agree the rules before work begins. Do not retroactively redefine success because a campaign performed badly.
Why sharing gross revenue can create the wrong incentives
A contributor promised half of every sale has an incentive to bring in sales. That alone says nothing about refunds, acquisition costs, delivery effort or whether the work consumes more founder time than it saves.
DN’s proposed design starts with an explicitly defined distributable contribution margin. This is a program calculation, not a claim about accounting net profit.
DN Proof-to-Payout Worksheet · v1.0
Step 1: establish the amount that can actually be allocated
Eligible collected revenue − indirect taxes collected for authorities − refunds and chargebacks − agreed direct delivery and processing costs = contribution margin before allocation.
Apply a zero floor to the proposed variable-sharing pool. Keep any operating loss visible separately; a zero pool does not erase a loss or cancel fixed contractual obligations. Costs must be agreed, documented and not counted twice.
| Illustrative line item | US dollars | Meaning |
|---|---|---|
| Eligible collected revenue, already excluding indirect taxes | 2,000 | Synthetic receipts, not a DN result |
| Refunds and chargebacks | −300 | Money that cannot remain in the pool |
| Processing fees | −100 | Separate from delivery costs below |
| Agreed direct delivery and review costs | −600 | Illustrative cost assumption |
| Distributable contribution margin | 1,000 | Before company overhead and applicable income taxes |
| Business allocation: 50% | 500 | Retained by the contracting business |
| Verified contributor pool: 35% | 350 | Shared among eligible contributors; not 35% each |
| Business-controlled reserve: 15% | 150 | Retained for approved business needs |
All amounts are illustrative. The allocation totals US$1,000: US$500 + US$350 + US$150. The reserve remains inside the business, making its combined allocation 65% in this example. That is not a statement of net profit, founder take-home pay or money available to spend without other obligations.
The 50/35/15 split is DN’s design basis, not an industry standard, guaranteed earning opportunity or active open-enrolment commission program. Actual agreements need defined eligibility, attribution periods, refund handling, direct-cost treatment and payout approval. A reserve is an accounting allocation until money is actually segregated; it is not automatically escrow or insurance.
Founder review time belongs in the economic assessment too. A program can appear attractive on cash costs while quietly consuming its owner’s entire week. Record that time separately even when it does not represent an actual cash payment.
Step 2: pass five gates before approving a variable reward
This is a practical audit worksheet, not a validated numerical score. A missing answer means hold the proposed allocation for review—not declare the contributor fraudulent.
| Gate | Required evidence | Hold condition |
|---|---|---|
| 1. Authority | Approved task, permitted channel and named accountable operator | Unapproved outreach, publication or commitments |
| 2. Work | Original deliverable and recorded acceptance against the brief | Missing, copied or unverifiable work |
| 3. Attribution | Agreed referral or contribution record, checked for overlap | Unresolved ownership, duplicate commissions or an existing customer claimed as new |
| 4. Economics | Collected eligible receipts, reconciled costs and refund treatment | Projected sales presented as cash or unknown deductions treated as zero |
| 5. Approval | Completed quality/policy review and authorized payout decision | Open dispute or unsupported recipient/payment details |
Keep the evidence chain simple: approved task → accepted work → attributable outcome → collected money → reconciled allocation → authorized payout. A spreadsheet or local register can document that chain before a business needs a dedicated platform.
Tracking links help assign credit under a chosen rule. They do not prove that every attributed sale was incremental. A returning customer may have purchased anyway. Compare against a pre-pilot baseline or suitable holdout when practical, and label uncertain attribution instead of presenting it as causal proof.
The visibility strategy: make something worth distributing
Our recommended first experiment is one original asset, one audience and one commercial question. For a publisher, that could be a useful benchmark explained in a short video. For a software company, it could be a worked tutorial. For a retailer, it could be a genuinely helpful product-selection guide.
Repurpose the insight rather than mass-producing near-identical pages. Give each format a reason to exist: the article explains the evidence; a carousel illustrates the decision; a short video demonstrates the mistake; the tool helps a visitor apply the lesson.
Measure the stages separately. Impressions are not readers. Readers are not enquiries. Enquiries are not accepted orders. Accepted orders are not collected cash. Collected cash is not profit.
For a small launch, we would compare qualified visits, assessment completions, owner-confirmed enquiries, collected revenue and review hours. A contact-button click remains a click until an enquiry is actually received. A readiness score is not a prediction of sales.
Compliance is part of the operating design
Google’s spam policies distinguish helpful content from scaled material created primarily to manipulate rankings. Paid links must be appropriately qualified. AI assistance is not permission to flood search with low-value pages.
YouTube’s fake-engagement policy prohibits artificially inflated metrics and warns that a hired promoter’s actions can affect the channel. Outsourcing amplification does not outsource the consequences.
The US Federal Trade Commission’s endorsement guidance calls for relevant material connections to be clearly disclosed and endorsements to be honest. South Africa’s Information Regulator provides direct-marketing guidance under POPIA, including electronic communications. The applicable rules depend on the activity and jurisdiction; commission-based outreach is not automatically exempt.
For a first pilot, our recommended boundary is narrower than the technology permits: no unsolicited mass messaging, fake accounts, purchased engagement, invented testimonials, fabricated partnerships or claims of guaranteed virality. Use approved channels, appropriate disclosures and a documented consent process where required. Get qualified advice for the actual contracting, privacy, tax and marketing arrangements.
What we have made available at DN
DN Agent Revenue Alliance begins with assessment and design—not an autonomous sales promise. Its free, browser-based assessment asks about the business and its readiness to measure and review contributor work. Results do not require an email address, purchase or paid AI API. The scoring is illustrative, not validated forecasting.
The page also provides an enquiry route for a scoped Blueprint, with an indicative US$499 founding-client price subject to confirmation. The Accelerator is presented as a waitlist concept, not an immediately purchasable managed operation. Neither completing an assessment nor sending an enquiry recruits agents, accepts a contract or authorizes a payment.
Apply the framework to your business
Start with readiness. Keep control of the next step.
Use the free assessment to identify whether your next move should be stronger foundations, a design blueprint or preparation for a tightly bounded pilot.
Take the free Agent Revenue Readiness Assessment →
This is DN’s own service. No guaranteed revenue, automatic onboarding or agent access to accounts or wallets. Commercial enquiries remain owner-managed.
For the operating-capacity side, see The One-Person Agentic Company. For a separate discussion of evidence and delegated authority, explore DN Agent Trust; its research is not certification of this service or of marketplace contributors.
DN Alpha Thesis
The opportunity is not to build the loudest swarm. It is to build the clearest connection between useful work, legitimate demand and verified economic value.
A business that can define the job, check the evidence, protect the customer relationship and explain every payment has a more defensible starting point than one that merely promises billions of views.
The next growth advantage may come from better incentives—not more agents.
Methodology, limits and corrections
This article separates provider-reported experiments, marketplace documentation and DN’s proposed operating design. The worksheet is original decision support, not a financial forecast, legal opinion, security certification or observed performance dataset. DN did not independently reproduce Anthropic’s experiments, test marketplace payouts or commission external agents for this article.
The financial example is synthetic. No private customer, campaign or business financial records are included. The assessment and service descriptions were checked against DN’s published landing page on 22 October 2026; availability and written terms should be confirmed before any engagement.
Change log: Version 1.0 — 22 October 2026: initial source review, proof-to-payout worksheet and service disclosure. Report a correction or request clarification.
Frequently asked questions
Can AI agents make money for a business?
They can perform commercial tasks, including negotiation in documented experiments. That does not establish reliable profitability for your business. Test collected revenue, actual costs, customer outcomes and human review effort in a bounded workflow.
Does a 50/35/15 split apply to gross revenue?
Not in DN’s proposed model. The illustrative split applies to an agreed, non-negative distributable contribution margin after specified deductions. Fixed contractual obligations, taxes and disputes still need separate treatment.
Are iLands Tokens the same as cash earnings?
No. iLands’ FAQ describes Tokens as resource units with no cash value or investment rights. Verify the actual counterparty, payout rules and settlement evidence for any separate paid arrangement.
Does DN’s free assessment activate an agent team?
No. It provides an illustrative browser-based readiness result. Recruitment, implementation, contracts, publishing permissions and payments require separate decisions. The Accelerator is a waitlist concept.
Primary sources
- Anthropic — Project Vend: Can Claude run a small shop? Published 27 June 2025.
- Anthropic — Project Vend: Phase two. Published 18 December 2025.
- Anthropic — Project Deal. Published 24 April 2026; describes a marketplace experiment conducted in December 2025.
- iLands — product, infrastructure and token FAQ. Accessed 22 October 2026.
- Google Search — spam policies. Accessed 22 October 2026.
- YouTube — fake-engagement policy. Accessed 22 October 2026.
- FTC — endorsement guidance. Accessed 22 October 2026.
- Information Regulator South Africa — direct-marketing guidance index. Guidance dated 3 December 2024; index accessed 22 October 2026.
Final disclosure: This article promotes DN’s own assessment/design offering while examining the broader concept. Third-party research and marketplace links are neutral source links. There is no claim of a partnership with Anthropic or iLands, no invitation to invest, and no promise of contributor earnings.
About the Author
