Can an AI Agent Pay for Itself?

01 / The question

Can the machine create enough value to cover its own cost?

The obvious answers were bad ones: low-paying surveys, spam, speculative trading, mass-produced “AI content,” or another software-as-a-service business that would require months of engineering, sales, and support.

The constraint changed the problem. The owner cannot take on substantial extra work and does not write code. The agent can research, create files, build software, operate browsers, test systems, and maintain documented workflows. The experiment therefore asks whether that capability can be packaged into one honest, low-maintenance digital product.

02 / The rules

No fake traction. No hidden labour.

• 01
  Legitimate work only. No fabricated reviews, purchases, followers, credentials, or testimonials.
• 02
  No automated spam. No unsolicited posting, mass outreach, or platform-rule workarounds.
• 03
  No paid advertising for 30 days. The product must first demonstrate organic interest.
• 04
  Owner workload stays bounded. Routine involvement targets five minutes per week, plus unusual customer or platform requests.
• 05
  Authority remains human. Identity, banking, publication, customer replies, refunds, and financial changes require explicit owner approval.
• 06
  Failure remains a result. Zero sales will be reported as zero sales, not reframed as “brand awareness.”

03 / The product

Control the Build

The first product is a plain-language safety system for non-coders managing AI-built projects. It addresses a specific moment: the prototype works, but every new request risks breaking behaviour the owner already accepted.

The package includes a 21-page playbook, nine editable Markdown templates, a worked fictional example, and a printable change-safety card. It helps buyers define the accepted baseline, constrain changes, request evidence, preserve recovery, and decide whether to accept, revise, roll back, or escalate.

04 / Division of labour

The agent works. The human remains accountable.

05 / Measurement

The experiment has a stop condition.

The scorecard is reviewed weekly. The primary 30-day measures are:

• Search impressions and listing visits
• Favourites, sales, and conversion rate
• Customer questions and refunds
• Marketplace fees and net revenue
• Human support time per sale

06 / Day zero

Everything is live. Nothing is proven.

This is the useful starting point: one product exists, the storefront is operational, the monitoring rules are active, and there is no evidence yet that strangers want to buy it.

Update / 30 July 2026

Route 02 is live: sell the work, not another product.

The experiment now includes a bounded Fiverr service for testing and auditing vibe-coded web applications before launch. It tests whether the agent can convert its existing testing, evidence-capture, and reporting capability into paid work while the human retains identity, financial, customer, and acceptance authority.

The first inbox activity was not demand. Seven notifications led to zero legitimate orders and six suspicious on-platform conversations. No supplied links were opened. The dashboard still showed zero impressions, zero clicks, and zero orders at the checkpoint.

Update / 9 August 2026

Checkpoint 02: low discovery, zero clicks, and a narrower second test.

The first two weeks did not produce enough buyer attention to judge the quality of the work. Etsy recorded six visits in 30 days, including only two from Etsy search, three listing views, zero favourites, and zero orders. Fiverr recorded 77 impressions, zero clicks, and zero orders.

The diagnosis is upstream of delivery: the Etsy shop has one search door, while the Fiverr card asks a cold buyer to begin with a broad audit at a high entry price. Pivot V2 replaces that with countable outcomes and smaller commitments.

Three focused Etsy downloads are prepared: a project handoff, a bug and regression kit, and an app launch checklist. The Fiverr five-bug offer is now live at US$15 with new copy and gallery art. The separate bug-fix gig remains staged because Fiverr presented a new human-verification gate.