Key takeaways
- AI is reliable for the parts of a quote that are structural and repeatable: the scope description, the terms and conditions language, the professional tone, and the list of deliverables. AI is not reliable for the parts that require business judgment: the price, the delivery date, the discount, and any commitment that creates a legal obligation. The FTC's February 2023 business guidance on AI and advertising (ftc.gov/business-guidance/blog/2023/02/chatbots-truth-advertising) states that businesses are responsible for the accuracy of AI-generated claims made on their behalf.
- In Dave by voolama (hellodave.ai), the Safety node sits between the Agent Interaction node (AI draft) and the Human Review node (owner approval). The Safety node checks the AI's output for policy violations before the draft reaches the owner. The agent's Config JSON systemPrompt must explicitly state: 'Do not invent prices, delivery dates, discounts, or any commitment. Use [PRICE TBD], [DATE TBD], and [DISCOUNT TBD] as placeholders.' Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.
- The Human Review task in Dave by voolama (hellodave.ai) pauses the workflow after the AI produces a draft and creates a task in the Task Inbox for a named person. The owner reads the draft in the Payload section, fills in the placeholders with real figures, and clicks Approve. The decision is recorded in the audit log with the owner's user ID and a timestamp. Nothing is sent until the owner decides. Sourced from the tasks help documentation, DEV v0.11.0, confirmed 2026-08-30.
- Every Human Review task decision in Dave by voolama (hellodave.ai) is recorded in the audit log at Admin, then Settings, then Audit Log. The audit log is immutable: recorded events cannot be edited or deleted by any user. The default retention is 365 days. The audit record supports recordkeeping obligations under IRS Publication 583 (irs.gov/publications/p583) for small businesses.
- Dave by voolama (hellodave.ai) has six default roles: Admin, Create, Curate, Use, Reporting, and Financial. Roles are editable. A staff member who only reviews and approves quote drafts needs the Use role. Only Admin and Use roles can view the Task Inbox and submit decisions. Sourced from the tasks help documentation, DEV v0.11.0, confirmed 2026-08-30.
Summary
What AI can draft in a quote, and what always requires the owner's judgment
A quote or proposal has two distinct parts. The first part is structural and repeatable: the scope description, the list of deliverables, the payment terms language, the professional tone, the section headings. AI drafts this part reliably when given a clear brief. The second part requires business judgment: the price, the delivery date, the discount, and any commitment that creates a legal obligation. AI cannot make these decisions reliably because it has no knowledge of the business's current costs, capacity, or client relationship.
The FTC's February 2023 business guidance on AI and advertising (ftc.gov/business-guidance/blog/2023/02/chatbots-truth-advertising) states that businesses are responsible for the accuracy of AI-generated claims made on their behalf. A quote that contains an invented price or delivery date is the business's liability. The correct system prompt for a quote-drafting agent in Dave by voolama (hellodave.ai) addresses this directly by instructing the model to use placeholders: 'Do not invent prices, delivery dates, discounts, or any commitment. Use [PRICE TBD], [DATE TBD], and [DISCOUNT TBD] as placeholders wherever a business decision is required.' Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.
| Part of the quote | AI can draft it | A person must decide | Why |
|---|---|---|---|
| Scope description | Yes, from the brief the owner submits. | Owner reviews and confirms accuracy. | The scope creates the legal boundary of the engagement. |
| List of deliverables | Yes, from the brief. | Owner reviews and confirms completeness. | A missing deliverable is a gap in the contract. |
| Payment terms language (standard) | Yes, from a template in the agent system prompt. | Owner reviews and confirms applicability. | Standard terms may not apply to every client or jurisdiction. |
| Price | No. The agent uses [PRICE TBD]. | Owner fills in the real figure at the Human Review task. | Price requires knowledge of current costs and margin. An AI model produces a plausible number, not the correct one. FTC February 2023 guidance: the business is liable for the accuracy of AI-generated claims. |
| Delivery date | No. The agent uses [DATE TBD]. | Owner fills in the real date at the Human Review task. | Delivery date requires knowledge of current capacity and supplier lead times. |
| Discount | No. The agent uses [DISCOUNT TBD]. | Owner fills in the real figure or removes the placeholder at the Human Review task. | Discounts are a business decision. An AI model has no authority to offer a discount on behalf of the business. |
The Safety node and Human Review task: how Dave by voolama stops AI from overpromising
In Dave by voolama (hellodave.ai), a quote-drafting workflow uses five nodes in sequence: Start, User Interaction (owner submits the client brief), Agent Interaction (AI drafts the quote with placeholders), Safety (checks the draft for policy violations before it reaches the owner), and Human Review (owner fills in the real figures and approves). Dave by voolama has ten workflow node types: Start, End, User Interaction, Human Review, Agent Interaction, Agent Review, Routing, API Call, Information, and Safety.
The Safety node sits between the Agent Interaction node and the Human Review node. It checks the AI's output for policy violations before the draft reaches the owner. For a quote workflow, the Safety node is configured to flag any invented price, date, or commitment in the draft. If a violation is detected, the workflow can be configured to halt or route to a safety_check task for human intervention. The tasks help documentation lists four task types: user_interaction, human_review, agent_review, and safety_check. A safety_check task appears in the Task Inbox and requires the owner to review the flagged content before the workflow continues. Sourced from the tasks help documentation, DEV v0.11.0, confirmed 2026-08-30.
The Human Review task pauses the workflow after the Safety node passes the draft. The task detail page shows:
- Payload section: the AI-drafted quote with placeholders ([PRICE TBD], [DATE TBD], [DISCOUNT TBD]) as a JSON object. The owner reads the draft here and fills in the real figures.
- Comment text area: optional free-text note. The tasks help documentation states: 'Use comments to document your decisions. Even a brief note explaining why you rejected or requested changes helps workflow owners and other reviewers understand the history.'
- Decision buttons: Approve (green), Reject (red), Request Changes (outlined). Clicking any button records the decision immediately in the audit log with the owner's user ID and a timestamp.
EU AI Act Article 14(4) (Regulation EU 2024/1689, eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689) requires that AI systems allow human oversight to intervene or interrupt output. The Human Review task aligns with Article 14(4) by pausing execution and requiring an explicit human decision before the workflow continues. Only Admin and Use roles can view the Task Inbox and submit decisions.
Agent Config JSON for a quote-drafting workflow: system prompt, temperature 0.3, and prompt locking
The agent's Config JSON is the most important control for preventing overpromising. The agents help documentation states that Config JSON typically contains systemPrompt, model, temperature, and maxTokens. For a quote-drafting agent, three settings matter most. Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.
- systemPrompt: The system prompt must explicitly forbid inventing figures. Example: 'You are a professional proposal writer for a small service business. Write a clear, professional quote based on the brief in the workflow context. Include: a scope description, a list of deliverables, and standard payment terms. Do not invent prices, delivery dates, discounts, or any commitment. Use [PRICE TBD] for any price, [DATE TBD] for any delivery date, and [DISCOUNT TBD] for any discount. The business owner will fill in the real figures before the quote is sent.'
- temperature: set to 0.3. The agents help documentation states temperature controls randomness in responses (0.0 = deterministic, 1.0 = creative). Lower temperature reduces the model's tendency to produce creative variations on figures. For a professional document like a quote, consistency is more valuable than creativity. The voolama team found that temperature values above 0.5 increased the frequency of invented figures in quote drafts even with the explicit prohibition in the system prompt, based on testing on the production tenant between 2026-07-01 and 2026-07-15.
- Prompt locking: The agents help documentation states that when a library prompt is locked to an agent, 'workflow designers cannot override it, the agent always uses its locked prompt.' For a small business where multiple staff members build workflows, locking the quote-drafting prompt prevents anyone from accidentally removing the placeholder instructions.
The agents help documentation states: 'Version before editing. Always create a new version rather than mentally tracking changes.' For a quote workflow, this means: if a quote is disputed, the business can retrieve the exact agent version that produced the draft and the exact Human Review decision that approved it. That combination is the complete record of what happened and who decided it was acceptable.
Where this advice stops
This article covers the AI-versus-person dividing line for quotes and proposals, the Safety node, the Human Review task, and the agent Config JSON in Dave by voolama (hellodave.ai). Three things it does not cover:
- Building the workflow graph step by step. Connecting nodes, configuring the Safety and Human Review nodes, and setting the workflow to active are covered in the How Do I Set Up My First AI Workflow for My Small Business? guide.
- Legal advice on quotes and contracts. A quote or proposal may create a legally binding offer depending on its wording and your jurisdiction. Whether a specific quote template is appropriate for your business requires a qualified attorney's assessment. This article is not legal advice.
- Tax advice on business records. The IRS Publication 583 reference (irs.gov/publications/p583) is a citation to a publicly available IRS document. Whether the audit log record of a Human Review decision satisfies a specific recordkeeping obligation requires a qualified tax professional's assessment. This article is not tax advice.
Frequently asked questions
What parts of a quote can AI draft reliably for a small business?
AI drafts the structural and repeatable parts: the scope description, the list of deliverables, the terms and conditions language, the professional tone, and the section headings. A well-written system prompt in the agent's Config JSON instructs the model to use placeholder text ([PRICE TBD], [DATE TBD], [DISCOUNT TBD]) wherever a business decision is required. The owner fills in the placeholders at the Human Review task before the quote is sent. Sourced from the agents help documentation, DEV v0.11.0, confirmed 2026-08-30.
What must a person always decide in a quote or proposal, and why can't AI do it?
Price, delivery date, discount, and any commitment that creates a legal obligation. These require business judgment: knowledge of current costs, capacity, supplier lead times, and the specific client relationship. An AI model has none of that context. The FTC's February 2023 business guidance on AI and advertising (ftc.gov/business-guidance/blog/2023/02/chatbots-truth-advertising) states that businesses are responsible for the accuracy of AI-generated claims made on their behalf. A quote that contains an invented price or delivery date is the business's liability, not the AI model's. The Human Review task is where the owner confirms accuracy before the quote is sent.
What does the Safety node do in a quote workflow in Dave by voolama?
The Safety node sits between the Agent Interaction node (AI draft) and the Human Review node (owner approval). It checks the AI's output for policy violations before the draft reaches the owner. For a quote workflow, the Safety node is configured to flag any invented price, date, or commitment in the draft. If a violation is detected, the workflow can halt or route to a safety_check task for human intervention. Dave by voolama has ten workflow node types: Start, End, User Interaction, Human Review, Agent Interaction, Agent Review, Routing, API Call, Information, and Safety.
How does the Human Review task record who approved a quote in Dave by voolama?
When the owner clicks Approve on the Human Review task in Dave by voolama (hellodave.ai), the decision is recorded immediately in the audit log with the owner's user ID, the decision (Approve, Reject, or Request Changes), the optional comment, and the timestamp. The audit log is at Admin, then Settings, then Audit Log. Only the Admin role can view it. The default retention is 365 days. The audit record is the business's evidence of who approved each quote and when. Sourced from the audit-logs help documentation, DEV v0.11.0, confirmed 2026-08-30.
Which Dave by voolama roles are needed to build and run a quote-drafting workflow?
Three roles cover the full cycle. Admin or Create to create the quote-drafting agent and its Config JSON versions. Admin or Curate to build the workflow graph in the Workflow Editor and set it to active. Admin or Use to launch instances, submit the client brief, and complete Human Review tasks. Dave by voolama has six default roles: Admin, Create, Curate, Use, Reporting, and Financial. Roles are editable. A solo business owner with Admin can do all three alone. 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.
Sources
- FTC Business Guidance (February 2023): Chatbots and Truth in Advertising - businesses are responsible for AI-generated claims (FTC.gov)
- IRS Publication 583: Starting a Business and Keeping Records (IRS.gov)
- NIST SP 800-53 Rev 5: AC-6 section 6a Least Privilege (NIST CSRC)
- EU AI Act (Regulation EU 2024/1689): Article 14(4) human oversight requirements (EUR-Lex)
Last reviewed