Compare what AI handles, where you take over, what it costs, and what has actually been tested.
Revenue targets are scenarios you choose—not expected or observed income.
Researches buyer pain, alternatives, demand signals, and practical offer angles without treating search interest as purchase intent.
Builds downside, base, and upside math while keeping every revenue target visibly hypothetical.
Runs consent-based interviews, synthesizes evidence, and prepares a bounded pre-sell offer with human-controlled outreach and payment decisions.
Turns public business evidence into a small approval queue of tailored proposals and learns from actual replies.
Runs a primary-source-first research process with explicit scope, uncertainty, QA, and human delivery approval.
Drafts content from a founder-owned corpus, gates every publication, and learns only from observed audience response.
Captures current evidence, identifies friction, and produces five prioritized recommendations with explicit limits.
Converts a repeated job into tested templates, instructions, previews, and an honest marketplace listing.
Moves from interviews to one-job software, tests, instrumentation, and a human-controlled launch packet.
Makes uneven autonomy visible: AI can research, build, gather permitted data, and draft outreach; humans approve contact and own closing.
Maps the current process, builds reusable artifacts, and proves the flow with fictional data before client use.
Builds and tests a tiny x402-protected resource on Base Sepolia or Solana Devnet; it never uses mainnet or counts testnet settlement as revenue.
Level 5 means exception-only oversight—not no human involvement. The useful question is where the handoff happens.
Makes uneven autonomy visible: AI can research, build, gather permitted data, and draft outreach; humans approve contact and own closing.
AI researches the niche; the human decides whether evidence is sufficient.
AI can plan and build the directory under review.
AI qualifies and drafts outreach, but every contact requires approval.
AI prepares the lead packet; a human calls or closes.
Humans own contracts, collection, disputes, and ongoing quality.
Readiness, workflow proof, real revenue, and repeatability answer different questions. Income Lab keeps them separate.
These capabilities are real, but they are not launch recommendations or copyable workflows.
A dedicated Agentic brokerage account can place real long-equity and options orders. Trading is confined to that account, but a connected agent may read broader Robinhood account data.
Why it is not runnableReal trading can lose principal, and the broader read access means the dedicated account is not full data isolation.
Robinhood: Agentic Trading overviewAn agent can use a controlled virtual card within a user's Robinhood Gold Card. It is not a bank or checking account owned by an agent.
Why it is not runnableIt can initiate real purchases, which are prohibited for the launch experiment.
Robinhood: Agentic Credit CardAlpaca's official MCP v2 defaults to paper trading, but paper mode can still place simulated orders unless its tools are filtered.
Why it is not runnableThere is no global read-only switch. The official example includes an account toolset with a mutation tool, so the protocol must expose data-only toolsets explicitly.
Alpaca: Trading MCP Server