- Home
- Skills
- Workflow Automation
- Support & Sales Chatbot That Does Not Invent
Support & Sales Chatbot That Does Not Invent
Grounded support, after-sales and pre-sales chat: prices fetched not recalled, identity before order data, real handoff, measured cost.
$49
Works with the AI Agents you already use
Support & Sales Chatbot That Does Not Invent
Example session with this skill installed
Analyze my support bot configuration in ./bot-config and tell me if it's safe to deploy for order tracking.
- Read your context and instructions
- Compiled the support & sales
The commercial guardrail scan for ./bot-config is complete. The configuration was evaluated against requirements for identity verification, tool grounding, and refusal paths.
Bot Security and Grounding Audit
| Requirement | Status | Observation |
|---|---|---|
| Identity Check | PASS | Bot requires order_id and postcode before calling order_status. |
| Tool Grounding | FAIL | Found hardcoded delivery times in policy.md. Should use shipping_api. |
| Refusal Path | PASS | Instruction found to offer human handoff after two failed retrieval attempts. |
| Citation Logic | WARNING | Citation events are enabled but help_center_index lacks URL metadata. |
| Injection Defense | PASS | User content is correctly delimited using XML tags in the system prompt. |
| Disclosure | PASS | Mandatory AI assistant identification is present in the greeting. |
Next steps
- Remove the static delivery estimates from
./bot-config/policy.mdto prevent legal binding to stale data. - Update the
help_center_indexschema to include source URLs for the structuredciteevent. - Run
python scripts/dimensionner_bot.py plan.jsonto verify the cost-per-conversation for this specific flow.
Connects securely to your tools. The creator never sees your data.
About this skill
Prices fetched, not recalled. Handoff that carries the conversation.
Put a support, after-sales and pre-sales chatbot on a website and keep it honest - grounded answers with citations, the four moving commercial facts fetched by tool call instead of recalled, identity checked before order data, clean handoff to a human with full context, prompt-injection defence on user content, measured containment and cost per conversation. Use when a site adds a chat assistant, when a bot invents prices or delivery times, or when support volume must come down without angering customers.
Any team ships a chat widget in an afternoon. Three decisions separate it from a production assistant, and all three are made before the first prompt. Price, stock, delivery and the return window are fetched at answer time, never recalled - in most consumer law what your channel states binds you, so a bot quoting last quarter's price has made an offer. The bot must be able to say it does not know and hand over, because a model with no refusal path answers anyway, fluently. And the business case is arithmetic, not faith.
Integration is spelled out with the code: the widget that calls your backend so the model key never reaches the browser; the streaming endpoint with token, tool, citation and done events; narrow tool contracts where every commercial response carries an as_of timestamp, writes need explicit confirmation and an idempotency key, and a failed tool returns a typed error instead of an empty result the model will fill in; the grounded-answer contract rejected at the application layer when it carries a number with unknown confidence; the handoff payload; and two CI gates.
Includes an EXPERT.md reference (exact parameters, thresholds, failure diagnosis): retrieval chunking, top-k and the relevance floor that is the actual refusal mechanism, the five prompt-injection defences in order of effect, the golden-set evaluation with four scored metrics, operational metrics defined so they cannot flatter you, a four-stage rollout, and a ten-row failure table.
Tested scripts. scripts/dimensionner_bot.py computes tokens, cost, containment economics and felt latency: on 20 000 conversations a month at 62 % containment it returns 4 936 input tokens per conversation, 0.0087 per conversation, and 0.0464 per contained conversation against 3.50 for a human contact - and it refuses a plan that loses money, assumes containment above 80 %, or takes over two seconds to show a word. scripts/garde_commerciale.py scans prompts and content for prices, stock, delivery times, warranties and absolute promises written as text, and for eight missing guardrails: no price tool, no refusal instruction, no citation, no handoff, no AI disclosure, no identity check before order data, no injection defence.
Acceptable use: assisting your own customers on your own service, with the assistant identified as automatic and a route to a human. The skill refuses presenting the bot as human when asked, dark patterns that hide the human route, pressure selling or invented scarcity, profiling by health, origin, religion, politics or sexual orientation, training on transcripts without a lawful basis and notice, and answering regulated medical, legal or financial advice as if qualified.
Not legal advice: consumer law, DSA and AI Act transparency duties depend on your market and size.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Workflow Automation.
Requires backend for tool execution. Frontend widget not included. Does not support regulated medical, legal, or financial advice. Validation scripts require Python 3.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 27 days ago
- Software
Python