Most cannabis delivery operators in Michigan are not short on ideas. They are short on hours. Between dispatch, menu updates, customer texts, and driver schedules, the back office rarely gets the attention it needs. Many teams have started testing AI tools to draft product descriptions, reply to common questions, and build training materials, and the early results are uneven. The difference usually comes down to the instructions you give the tool. Some operators begin by browsing an ai prompt store to see how tested instructions are structured before writing their own.
Why generic prompts fail in cannabis delivery
A prompt like ‘write a product description for a gummy’ produces something bland, and worse, it may produce claims you cannot make. Cannabis marketing sits under strict rules, and a tool that has never read your license terms will not know which words are off limits. Generic prompts also ignore your actual operation. They do not know your delivery radius, your order minimums, your cutoff times, or the fact that your customers are adults who expect discreet packaging and clear answers.
A useful prompt for this industry does four things. It states the role, the audience, the constraints, and the output format. It also tells the model what to avoid. Written this way, the same request becomes far more reliable.
Core prompt categories for delivery operators
Menu and product copy
Ask the model to describe a product using only the facts you supply: THC and CBD percentages from your lab report, strain or product type as listed, serving size, and packaging details. Instruct it to avoid medical language, avoid implying treatment of any condition, and avoid words that appeal to minors such as candy-like references. Then have a staff member check every line against the label before anything goes live.
Customer support replies
Support is where delivery businesses gain the most time. Build a prompt that covers the common cases: order status, a missed delivery window, an ID verification question, a refund request, or a driver running late. Include your actual policy language and tell the model to escalate anything involving a complaint about safety, a suspected sale to a minor, or a medical question. The model should never answer dosing questions. It should direct those to a licensed healthcare provider.
Driver and dispatch training
New drivers need to learn ID checking, sealed-package handling, delivery documentation, and what to do when a customer is intoxicated or refuses to verify age. A training prompt can turn your written procedures into quiz questions, role-play scenarios, and checklists. Ask the model to generate scenarios where the correct answer is to refuse the sale. Those are the cases new staff most often get wrong.
Scheduling and internal summaries
Shift handoffs, end-of-day reports, and weekly inventory summaries are repetitive and easy to standardize. Give the model a fixed template and the raw notes, and ask for a summary in that template with any missing fields flagged rather than guessed.
Guardrails that protect your license
Treat AI output as a first draft from a junior employee who has never worked in your building. Several rules help keep that draft safe:
- Never enter customer names, addresses, order histories, or ID details into a public AI tool. Strip personal data before pasting anything.
- Confirm current requirements directly with the Michigan Cannabis Regulatory Agency and your attorney or compliance consultant. Rules change, and an old prompt can carry outdated language.
- Keep a human approval step for anything public-facing, including social posts, email campaigns, website copy, and delivery app listings.
- Log which prompt produced which published content, so you can trace and correct it later.
- Do not let AI make promises about effects, potency outcomes, or health benefits.
How to test and refine a prompt
Good prompts are tested, not hoped for. Run each new prompt against five or six realistic inputs, including a few messy ones: a customer who writes in all caps, a question about a product you do not carry, and a request that should be refused. Score the outputs on accuracy, tone, and compliance. Change one variable at a time so you know what caused an improvement. When a prompt works, save it with a version number and a short note about what it is for. To go deeper, explore The marketplace for AI prompts that actually work.
Many teams keep a shared document of approved prompts so that a closing shift and an opening shift produce consistent messages. Over time this library becomes a real operating asset, especially when a new manager joins and needs to understand how the business communicates.
A sample support prompt structure
Here is the general shape of a support prompt that works well for a delivery operation. Adapt the bracketed sections to your own policies:
You are a customer support assistant for [business name], a licensed cannabis delivery service in Michigan. Answer only using the policy text provided below. Be brief, friendly, and factual. Do not give medical advice or dosing guidance. If the customer asks about a health condition, reply that you cannot advise on that and suggest they speak with a licensed healthcare provider. If the message mentions a minor, a safety concern, or a complaint about a driver, label it ESCALATE and stop. Output a reply to the customer followed by a one-line internal note.
Notice what this does. It limits the source material, sets the tone, defines the escalation path, and specifies the output. Each of those controls reduces the chance of a damaging reply.
Measuring whether it is working
Judge AI tools by operational results rather than novelty. Useful measures include the time a support message takes to resolve, the number of messages that require human correction, the number of approved public posts that needed edits after review, and staff feedback on whether training materials are clearer. Track these for a month before deciding a prompt is worth keeping. If corrections are rising, tighten the constraints rather than adding more creative instructions.
Where to start this week
Pick one workflow that is repetitive and low-risk, such as order status replies or shift summaries. Write a prompt with the four elements described above. Test it on real but anonymized examples, have a manager approve the output, and run it for two weeks. Then expand to the next workflow. Operators who try to automate everything at once tend to lose trust in the tools quickly, while those who build slowly usually end up with a small, dependable set of prompts that fit their business.
AI will not replace the judgment a good dispatcher or compliance-minded manager brings. What it can do is handle the drafting so your people spend more time on the calls, checks, and customer relationships that actually keep a delivery business running well.
Leave a Reply