How Can a Small Online Store Use AI for Product Listings Without Getting the Facts Wrong?

AI language models do not know your product. They know patterns from text they were trained on. When asked to write a product listing, they produce plausible-sounding copy by combining patterns. The problem is that plausible-sounding is not the same as accurate.

Four categories of invented claims create the most risk for a small online store, each tied to a specific regulatory or commercial consequence:

The fix is not to avoid AI for product listings. The fix is to build the workflow so the AI can only use facts the owner has verified, and so a person checks the output before it is published. The five-node workflow in the next section does exactly that.

In Dave by voolama (hellodave.ai), a product listing workflow uses five node types to prevent invented claims. Dave by voolama has ten workflow node types: Start, End, User Interaction, Human Review, Agent Interaction, Agent Review, Routing, API Call, Information, and Safety. Four of those ten are used in this workflow.

1. User Interaction node: staff submits the product name and SKU. The staff member opens the Task Inbox (navigate to Tasks in the sidebar), opens the User Interaction task, fills in the product name, SKU, and a reference to the verified spec sheet, and clicks Approve. Sourced from the tasks help documentation, DEV v0.11.0, confirmed 2026-08-30.
2. Information node: the verified spec sheet. The Information node holds the full verified spec sheet: dimensions, weight, materials, compatibility, certifications, and any health or safety claims the business has verified. The agent reads this node's content as part of the workflow context. The agent cannot access any source outside the workflow context. The workflows help documentation states the Information node 'does not affect execution, use it to add notes, context, or instructions directly on the canvas.' Sourced from the workflows help documentation, DEV v0.11.0, confirmed 2026-08-30.
3. Agent Interaction node: the AI drafts the listing. The agent's systemPrompt instructs it to write a product listing using only the facts in the workflow context, to use [NOT IN SPEC SHEET] as a placeholder for any claim it cannot source from the spec, and not to invent dimensions, materials, compatibility claims, or health benefits. Setting temperature to 0.3 produces consistent, factual output. The agents help documentation states temperature controls randomness in responses (0.0 = deterministic, 1.0 = creative). Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.
4. Agent Review task: a reviewer checks for invented claims. The tasks help documentation (DEV v0.11.0, confirmed 2026-08-30) lists four task types: user_interaction, human_review, agent_review, and safety_check. An agent_review task contains output from an Agent in its payload for the reviewer to verify or override. The reviewer opens the agent_review task, reads the draft in the Payload section, and checks every [NOT IN SPEC SHEET] placeholder. If the draft contains no invented claims, the reviewer clicks Approve. If it contains a placeholder or an invented claim, the reviewer clicks Request Changes and adds a Comment identifying the specific claim to fix.
5. Human Review task: the owner signs off. After the Agent Review step approves, a Human Review task is created for the owner. The owner reads the final draft in the Payload section, confirms it matches the spec sheet, and clicks Approve. The decision is recorded in the audit log with the owner's user ID and a timestamp. Nothing is published until the owner decides. Sourced from the tasks help documentation, DEV v0.11.0, confirmed 2026-08-30.

The agent's Config JSON is the primary control for preventing invented claims. Three fields matter most. Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.

systemPrompt with four explicit instructions:

You are a product listing writer for a small online store. Write a product listing using only the facts in the workflow context. Do not add, invent, or infer any specification, dimension, material, compatibility claim, or health benefit that is not explicitly stated in the workflow context. If you cannot find a specific fact in the workflow context, use [NOT IN SPEC SHEET] as a placeholder in the listing. Do not remove the placeholder: the reviewer needs to see it. Write in plain American English. Keep the listing under 150 words.

temperature: 0.3. The agents help documentation states temperature controls randomness in responses (0.0 = deterministic, 1.0 = creative). For a factual product listing, 0.3 produces consistent output with low variation. On 2026-08-15, the voolama content team observed that temperature 0.3 produced consistent placeholder output across five consecutive test runs on the production tenant, compared to two invented specifications without placeholders across five runs at temperature 0.7. These observations are from a single internal test on one content type on one date and should not be read as a performance guarantee.

Prompt locking. The agents help documentation states: 'When locked, workflow designers cannot override it, the agent always uses its locked prompt.' Save the systemPrompt above as a Prompt Library entry (navigate to Agents in the sidebar, switch to the Prompt Library tab, click New Prompt, select Category: Safety). Lock the prompt to the agent. After locking, the prompt selector in the Workflow Editor is greyed out for any workflow that uses this agent. Only Admin and Create roles can create agents and lock prompts. NIST SP 800-53 Rev 5 AC-6 section 6a (csrc.nist.gov/pubs/sp/800/53/r5/upd1/final) requires least-privilege role assignment.

This article covers the five-node workflow pattern in Dave by voolama (hellodave.ai) that prevents AI from inventing product facts. Four things it does not cover:

• This is not legal advice. The FTC 16 CFR Part 255 and FDA 21 CFR Part 101 citations are to publicly available federal rules. Whether a specific product listing violates those rules requires a qualified attorney's assessment. Health claim compliance for food and dietary supplements is a specialized area of regulatory law: consult a licensed attorney before publishing any health-related product claim.
• Dave by voolama does not publish to any e-commerce platform. The platform generates the draft and records the approval. Publishing the listing to the store is a separate step outside Dave by voolama, using the store's own product management interface. The help documentation does not describe any integration with named e-commerce platforms.
• The Information node does not verify the spec sheet. The Information node holds whatever the owner puts into it. If the owner's spec sheet contains an error, the agent will reproduce that error in the listing. The Human Review task is the control point where the owner confirms the spec sheet itself is accurate before approving the listing.
• The internal test observations are not a controlled study. The 2026-08-15 temperature comparison (five runs at 0.3 vs. five runs at 0.7 on the voolama production tenant) was a single internal observation on one content type. It is reported as context for the temperature recommendation, not as a performance guarantee.