This report was generated using the special AI prompt included in our new book, The Shelf Test.

The Shelf Test

Commercialization Assessment · Revision 5 · October 23, 2026

Invention

AI Food Investigation Agent

Target customers

Journalists, researchers, and health-conscious consumers

Development stage

Concept and workflow design

Commercial model

Paid software access, subscription, or API

1. My invention

The proposed product is a specialized agentic AI system that investigates the origin, manufacturing location, authenticity, and safety history of individual food products.

A customer supplies information about a food product, including its brand, packaging, barcode, or production codes. Multiple AI agents conduct a persistent investigation, searching public websites, corporate records, regulatory information, industry databases, historical sources, and accessible supplier records.

Unlike a conventional AI search, the system is designed to follow leads, investigate inconsistencies, connect information across sources, and continue investigating when evidence is inconclusive.

The resulting report contains:

• Investigative findings and potential leads.
• Links to supporting sources.
• Confidence levels reflecting the strength of the evidence.
• Explicit distinctions between verified findings, reasonable inferences, and unresolved questions.
  The system does not replace professional judgment. Journalists and researchers remain responsible for assessing evidence and drawing conclusions. Consumers gain access to a level of investigative research that would otherwise require considerable time and expertise.
  The initial market
  The invention now addresses two customer groups with different motivations.

Journalists and researchers need credible, traceable findings that accelerate investigations, expose previously overlooked information, and reduce the time spent searching and connecting evidence.

Health-conscious consumers and concerned citizens want to know more about the food they buy and eat. Their questions may concern where a product was actually manufactured, who made it, whether origin claims are accurate, and whether credible safety concerns exist.

For consumers, the underlying motivation can be personal and immediate: What am I feeding myself and my family, and can I trust what the package tells me?

The proposed product gives both groups access to the same investigative capability, potentially delivered through different interfaces.
The potato chip experiment
You attempted to identify the actual manufacturing facility of a bag of potato chips using two general-purpose AI systems.

• Gemini: Given a consumer-style question and packaging information, it could not identify the manufacturing location.
• ChatGPT: You pushed it to investigate further. It also failed to identify the factory, eventually describing the search as a dead end and explaining why manufacturers may not disclose this information.
  Neither investigation produced the desired answer.

This is a useful competitive baseline. Two conventional AI attempts failed to answer a specific, apparently straightforward consumer question.

It also illustrates a distinction between explaining why information is unavailable and conducting a deeper investigation to try to uncover it.

Working assumption: I’ll assume the invention works as described. These are provisional judgments; more detail would increase confidence. The system is still a concept, and the two unsuccessful AI searches are not demonstrations of the proposed system’s performance.
2. All seven gates
Gate Status Confidence Assessment

1. The Gatekeeper POTENTIAL TROUBLE Medium Some definitive manufacturing information may be controlled by food companies, suppliers, or specialized database operators. Investigative agents may find alternative evidence, but restricted access could prevent certain conclusions.
2. The Unlearning PASS Medium The design adds investigative capability while preserving professional judgment. Source links and confidence levels let researchers verify the work, while consumers can examine the basis for findings rather than accepting unsupported answers.
3. The Phantom POTENTIAL TROUBLE Medium Professional investigations provide recurring opportunities, but consumer concern may spike after a food-safety incident or purchasing question and subside afterward. This could limit ongoing consumer subscriptions.
4. The Duct Tape POTENTIAL TROUBLE High Conventional AI and web searching are convenient alternatives. The Gemini and ChatGPT experiences show a specific unanswered question, but the new workflow must demonstrate better investigative results to justify payment.
5. The Valley NOT APPLICABLE High No specific funding gap or approaching interruption to development or commercialization has been identified.
6. The Iron Gate NOT APPLICABLE Medium No identified government approval, qualifying incentive, legal restriction, or IP obstacle currently controls adoption.
7. The Handoff NOT APPLICABLE High No transfer of technical knowledge, succession, or continuity problem is presently indicated.

Assessment: One provisional pass, three potential adoption obstacles, and three gates not presently applicable. No gate is established as BLOCKED.
3. Where to work first
Primary obstacle: The Duct Tape
The commercial challenge remains persuading customers that the specialized system offers something worth paying for beyond existing AI tools.

The potato chip experiment makes this problem concrete.

An ordinary customer expects that entering a product name and a code from its packaging should reveal where the product was made. Yet neither Gemini nor ChatGPT could establish the manufacturing facility.

Your proposed system would investigate the unanswered question using additional research paths, such as possible contract manufacturers, corporate relationships, facility information, product codes, regulatory filings, and historical records.

The distinction is between a system that searches for readily available answers and one deliberately designed to pursue an investigation.

For journalists, a successful investigation could save substantial research effort or uncover a publishable lead.

For consumers, it could resolve a concern about the origin or history of a food they routinely purchase.

However, two separate issues remain:

1. Can the specialized workflow obtain useful evidence that ordinary AI missed?
2. Will customers pay enough for that improvement to support the service?
   The first concerns the proposed research advantage. The second is the adoption question at the heart of The Duct Tape.
   Connected obstacle: The Gatekeeper
   A manufacturer’s refusal to disclose its factory locations does not necessarily prevent independent investigation.

Agents might connect evidence from public records, local reporting, industry relationships, and other sources to identify a likely facility.

But sometimes the definitive answer may exist only in private production records. A persistent investigation cannot guarantee access to information that is not available.

The important product design choice is to distinguish a verified manufacturing location from an inferred location.

The system should not claim certainty merely because many indirect sources point toward one answer.
4. Ways through
The Duct Tape: Demonstrate a better investigation
Strategy 1 — Establish a repeatable comparison.

Use the potato chip product as the first benchmark.

Retain the unsuccessful Gemini and ChatGPT investigations, then compare their results with those of the specialized workflow.

Evaluate whether the new system discovers additional manufacturing information, identifies previously overlooked leads, provides stronger sources, or produces a more useful explanation of unresolved evidence.

Strategy 2 — Show different customers the same investigative result.

Give the report to a food journalist and a concerned consumer.

For the journalist, ask: “What investigative work did this save, and would the findings help you pursue a story?”

For the consumer, ask: “Does this report answer a question that matters to you, and would you pay to investigate another product?”

Next test: Complete one agent-driven investigation and obtain reactions from both customer groups.

What changes the rating: A demonstrated advantage over existing research methods that prospective customers value enough to adopt would justify a PASS.
The Gatekeeper: Find alternative paths to evidence
Strategy 1 — Map information dependencies.

Identify which sources are publicly accessible, which require subscriptions or licenses, and which are held privately by manufacturers.

Determine whether independent evidence can establish useful findings without relying on one company’s cooperation.

Strategy 2 — Pursue authorized access where necessary.

If specialized data is essential, investigate appropriate database licensing or supplier partnerships.

Useful question for a data provider: “Can your records be used in a paid investigative service, with citations to the underlying information?”

Next test: During the potato chip investigation, record which leads can be pursued independently and which encounter an actual information-access restriction.

What changes the rating: A demonstrated route around essential access dependencies, or permission to use needed information, would support a PASS.
The Phantom: Give consumers a reason to return
The consumer market introduces an obstacle that was less pronounced when journalists were the sole target.

People may become highly interested in food origins after a recall, contamination story, or troubling discovery. When the immediate concern passes, they may return to familiar shopping habits.

An individual consumer may also investigate a product once and have little reason to investigate it again.

Strategy 1 — Evaluate repeated product questions.

Test whether consumers want investigations of multiple products they purchase regularly, rather than only one product that has attracted attention.

Strategy 2 — Make the investigative benefit immediately visible.

Even when a complete answer remains unavailable, a report can distinguish known facts from uncertain claims and show what was investigated. It must not present the absence of a reported safety problem as proof of safety, or an undisclosed factory as evidence that a food is unsafe.

Useful question for consumers: “After seeing this investigation, which other food products would you want examined, and how often would you use a service like this?”

Next test: Show consumers the potato chip report and ask them to name additional products they would actually investigate.

What changes the rating: Evidence of recurring consumer demand, with benefits users can readily observe, would support a PASS.
5. Why The Unlearning passes
The proposed system has an important adoption advantage: it does not require journalists or researchers to abandon the skills that make them effective.

Instead, it reduces portions of the investigative workload while preserving their ability to question, corroborate, and interpret results.

Three design elements support that relationship.

Source links allow an investigator to inspect original material and determine whether a finding has adequate support.

Confidence levels make uncertainty visible, provided those levels reflect evidence quality rather than merely how certain the AI sounds.

Human control preserves the investigator’s authority to evaluate evidence and decide what deserves further attention.

For consumers, these same elements can make an investigation understandable and inspectable. However, confidence levels must be carefully explained so that consumers do not mistake a tentative finding for a verified food-safety judgment.

The system’s additive design supports a provisional PASS, with Medium confidence until users have evaluated actual reports.
6. What to reread
Primary: Chapter 10, “The Death of the Fire Demon”
This is the strongest match for how your invention intends to work with existing expertise.

In the chapter, changed chemistry eliminated hazardous solvent operations while preserving the required product. Workers welcomed the trial and used their own experience to inspect the result.

Your system similarly aims to reduce labor-intensive investigative work while preserving the journalist’s essential responsibility for evaluating findings.

Look for how eliminating burdensome work made a new approach easier to adopt without making established expertise irrelevant.
Supporting: Chapter 4, “Dale’s Shermanator”
Dale’s process incorporated recycled material while preserving the finish and strength customers expected.

The parallel is your need to improve investigative speed and depth while preserving the evidence quality, accountability, and source traceability that researchers already value.

The lesson is not merely to do more research. It is to deliver a better result without sacrificing the characteristics that make the existing method acceptable.

The Conclusion provides a useful framework for connecting these gates and deciding which work should come next.
7. My next moves
1

Use the potato chip problem as the first benchmark.

Preserve both failed general-purpose AI attempts. Define what would count as verified factory identification, a credible lead, or a well-supported unresolved result.

2

Design and demonstrate the investigative agent workflow.

Focus on deeper research, following leads, corroborating evidence, linking sources, and assigning defensible confidence levels. Identify where restricted records actually constrain the investigation.

3

Evaluate the results with professionals and consumers separately.

Determine whether journalists value the research assistance and whether consumers have sufficient recurring interest to pay for investigations.

Three questions for the next assessment

1. Should health-conscious consumers be able to purchase individual investigations, or do you envision an ongoing subscription?
2. Would the consumer report use the same detailed evidence format as the professional report, or a simplified presentation with access to the full sources?
3. Beyond identifying the factory, should the first version investigate ingredients, contamination history, recalls, and potentially misleading origin claims?

Shelf Test conclusion
The proposed AI Food Investigation Agent has a plausible adoption route because it addresses a recognizable information problem while preserving the researcher’s judgment.

The failed Gemini and ChatGPT searches strengthen the case for investigating whether existing consumer AI tools leave important questions unanswered. They do not yet prove that a deeper agentic workflow can answer them.

Adding health-conscious consumers expands the potential market but introduces a meaningful Phantom concern: individual consumers may want investigations only occasionally. Professional researchers may offer more regular usage, while consumers could provide a broader audience for accessible product-level reports.

The most important next work is to demonstrate that the specialized agents can uncover and substantiate findings that conventional AI research missed.

If that advantage is demonstrated, the commercialization discussion can move from what the invention might accomplish to which customers value the result enough to pay for it.