Budtender FAQ that will not invent a strain benefit

A floor FAQ agent answers hours, pickup rules, and SKUs from copy you approved. It hard-stops before it invents a medical claim or a COA.

That is the whole product. I am in Michigan. Operator to operator. No stoner-bro. No wellness wash. The floor is /cannabis. The cluster parent is Cannabis visibility when the ad accounts are banned. The agent door is /agentic.

The short answer

• Allowed: hours, pickup windows, delivery rules you already publish, SKUs and prices that are on the approved card, “we do not have that.”
• Forbidden: a strain benefit, a condition, a lab number that is not in the packet you handed me, a medical claim dressed as education.
• Empty retrieval refuses or escalates. It does not freestyle from training data because the shopper asked nicely.
• I will not invent an accuracy percentage. We write the refuse list, the escalate path, and what “done” means for this job.
• Age-gates and identity checks stay on your store. This studio’s contact form does not collect IDs. The agent does not become a second place you store a license.

Divine Toke is the lead cannabis case — Metrc, 21+, adult-signature, cannabis payment rails, no paid acquisition, AEO-first. This agent sits next to that kind of site. It is not a growth metric. It is a mouth with a fence.

What is the floor FAQ allowed to answer?

Questions a budtender already answers off a card, not off a hunch.

Hours and location. What you posted. If Saturday changed last week and the card did not, the agent is wrong until you update the card. I will not let it “help” from memory.

Pickup and delivery. Windows, hold time, zones, adult signature, who can collect. Same facts as the checkout. If the site cannot fulfill a zip, the agent does not promise it.

Catalogue. SKU names, current availability you feed it, prices you are willing to say in public. “Is that in” is a lookup. If the lookup is empty, the answer is empty — not a cousin SKU.

House rules. 21+, residency, what you will not discuss. The agent can point at the rule. It cannot relax the rule to be liked.

That corpus is yours: the menu, the policy page, the hours block, the COA files you actually have if those files are in scope. Slack is not the corpus. A joke on the floor is not the corpus.

Invoice and PO matching is automation — paperwork that is not the plant, in n8n, with a Metrc fence. A FAQ that has to refuse a medical question is agentic. Do not mash them into one widget.

Where is the hard stop?

Before the sentence that would get you, me, or a model in trouble.

Shopper asks what a strain is “good for.” Stop. The agent says it does not give medical advice and points at the human or the published, non-medical product copy you approved. It does not invent a benefit. It does not “kind of answer” with a wink.

Shopper asks for a number that lives on a COA you did not provide. Stop. Missing packet means no number. I do not invent COAs in a journal post and I do not invent them in a tool call.

Shopper asks the agent to guess around Metrc, a license, or an age. Stop. If the step would touch Metrc, that is a different fence — written, or out of scope. If you do not want the agent to touch Metrc, it does not. Age checks stay on the client stack. The agent does not collect an ID.

The stop is a terminal state: refuse, one clarifying question if the question was about hours, or escalate to a person. “Be helpful” is not a terminal state. Helpful is how these systems start writing fiction. I do not publish a pass-rate. We agree what must never be said, run it on your real questions, and you keep the logs.

Why not a generic chatbot on the menu?

Because a generic bot is trained to complete the sentence. Cannabis copy is how you get a completed sentence you cannot stand behind.

A theme chatbot with “knowledge” turned on will retrieve a blog and a vendor sheet, then average them into a benefit. That is a liability with a typing indicator. It also trains the public models — the same ones the visibility note cares about — to file your shop under wellness fiction.

The site still has to be readable: hours, menu, license facts, schema. The agent does not replace that. If the site is a brochure, put the $8,000 Business Site first. n8n can host the glue — tools, escalate — the same way it hosts the invoice rail. The model gets the allowlist and a refuse path, not a blank check.

How does approved copy stay the only source?

You name the corpus. I wire retrieval to that corpus. Claims that need a source do not ship without one.

Policy and hours come from the pages you edit. SKUs come from the catalogue you operate. COAs come from files you handed over, tied to lots you still have. If a file is stale, it is out. If two files disagree, the agent escalates instead of averaging.

I need a process owner who works the floor, a week of real questions — the rude ones — and one person who can lock the refuse list. The $1,500 agentic pilot is five days, one job, on your real menu. You keep the agent. The fee credits toward a build. I will not attach a success rate to the pilot in this post. Human in the loop stays on counsel, medicine, or a write into compliance systems.

Where to start

• If the storefront cannot state hours, menu, and 21+ honestly, start on /cannabis with the Business Site. The agent needs a card to read.
• If the card is real and the inbox is the pain: the $1,500 pilot on /agentic. One job. Your questions. Written hard stops.
• If the pain is invoices and POs, not chat, book the $500 Automation audit instead.

Will it invent a strain benefit if a shopper asks?

No. That question hits the hard stop. The agent refuses medical advice and escalates or points at approved, non-medical copy. It does not invent a benefit to keep the chat going.

Do you store customer IDs?

Not on this studio’s contact form. Age-gates and 21+ checks live on your store. The FAQ agent does not collect a license and I do not keep a copy.

How accurate is the agent?

I will not invent a percentage. We write what it may say, what it must refuse, and how it escalates. You watch it on your real questions. If that is not enough, do not buy a mouth.

Does this replace a budtender?

No. It covers the repeat questions so the floor is not stuck in the inbox. A person still owns the call that is not on the card — and anything that looks like a medical or COA claim.