Running a cannabis delivery service in Calgary means writing a steady stream of copy: order confirmations, driver arrival updates, product descriptions, FAQ answers, and replies to customers asking why their order is late. Many owners have started testing AI writing tools to handle some of this work, and the gap between a useful answer and a useless one usually comes down to the instructions you give the tool. Sets of prompt engineering templates can give your team a tested starting structure instead of a blank text box, which saves time and keeps tone consistent across staff members.
Why most AI replies sound wrong for delivery
A generic prompt such as "write a reply to a customer whose order is late" produces a generic answer. It might apologize too much, promise a specific arrival time the dispatcher cannot guarantee, or use language that does not match the brand. For a delivery business, those problems matter more than they do in most industries because customers are often anxious about discretion, timing, and whether the order will arrive at all.
A working prompt usually answers five questions before the model writes a single word: who is the writer, who is the reader, what is the goal of the message, what information is available, and what must never be said. When any of those are missing, the output drifts. Adding those five elements turns a vague request into something a dispatcher or support lead can use with light editing.
Start with a role and a boundary
The most useful instruction in any delivery prompt is the boundary. Tell the model what it is not allowed to do. For example, a support reply prompt might say that the assistant cannot confirm a delivery time unless the order status field includes one, cannot discuss medical effects, and cannot suggest that a product is safe or unsafe to consume. Without those lines, AI tools tend to fill gaps with confident-sounding claims, which is the last thing a licensed business needs.
Next, define the role. "You are a customer support writer for a licensed delivery service in Calgary. You write in plain, friendly English, avoid slang unless the customer uses it first, and keep replies under 80 words" gives the model a voice and a length target. Those two details alone remove most of the rambling that makes AI copy hard to scan on a phone.
Build a small library around the tasks you repeat
Rather than writing one master prompt, most teams do better with a short library of prompts, each tied to a single recurring task. Useful starting points include:
- Order status replies for orders that are preparing, out for delivery, or delayed
- Address and building-access clarifications, such as apartment buzzer instructions or condo lobby procedures
- Refund or substitution requests, with a triage step that flags anything needing a manager
- Product description drafts, limited to the verified facts in your inventory system
- FAQ answers for hours, service areas, payment methods, and delivery windows
Keep each prompt focused. A prompt that handles refunds, product copy, and driver messages at once will do all three poorly. Store the prompts in a shared document with a name, the task it covers, the input fields a staff member must fill in, and the date it was last reviewed. That last field matters because rules, service areas, and menu offerings change.
Use variables instead of retyping details
Good prompts contain placeholders such as [ORDER_STATUS], [ESTIMATED_WINDOW], [CUSTOMER_FIRST_NAME], and [ISSUE_SUMMARY]. Staff fill these from the order system, then paste the completed prompt into the AI tool. This keeps the model from guessing and reduces the risk of a message quoting the wrong window or the wrong name. It also makes the prompts easy to audit, because a reviewer can see exactly what information the reply was based on.
If a placeholder is empty, the prompt should instruct the model to say so rather than invent a value. A line such as "If ESTIMATED_WINDOW is blank, tell the customer that a dispatcher will confirm the timing shortly, and do not offer a time" prevents a common and costly error. To go deeper, explore The marketplace for AI prompts that actually work.
Keep humans in the loop for compliance
Cannabis is a regulated product, and the rules governing advertising, age verification, and packaging come from provincial and federal sources that change over time. An AI tool does not know your current licence conditions, and it will not be held responsible if a message breaks them. For that reason, treat AI output as a draft. A trained staff member should review anything that mentions potency, effects, health, pricing promotions, or age requirements before it goes to a customer.
It also helps to write these restrictions into the prompt itself and into your internal checklist. Ask the model to avoid health claims, to avoid implying that a product treats any condition, and to leave out promotional language that your legal adviser has not approved. Those instructions reduce the amount of editing needed, but they do not replace review. If you are unsure about a specific claim, check it with a lawyer familiar with cannabis advertising in Alberta before publishing.
Test prompts the way you would test a menu
Before a prompt goes live, run it through a set of realistic test cases. Include a plain order update, a late order, a customer who is angry, a customer who asks a question the business cannot answer, and a request that should be refused. Read every output and mark whether it is accurate, on-brand, and safe to send. If three out of five outputs need the same correction, the prompt needs a new rule, not a human rewrite every time.
Keep a simple log of changes. When a prompt is revised, note what problem it solved. Over a few months, this log becomes a practical record of what your business has learned about communicating with customers, and it makes onboarding new staff much faster.
Common mistakes to avoid
- Letting the model invent delivery times, prices, or stock levels that are not in your system
- Using one prompt for every channel, when a text message, email, and app notification each need different length and tone
- Skipping the boundary section, so the model discusses effects or makes safety claims
- Never updating prompts after a change in hours, service zones, or product lineup
- Publishing AI-written product descriptions without checking them against verified product data
- Assuming a polished reply is a correct reply
A simple workflow for your team
A practical workflow looks like this. A staff member selects the prompt for the task, fills in the variables from the order or inventory record, generates a draft, and checks it against a short review list: Is every fact in the draft present in the source data? Does it avoid restricted claims? Does it match the brand voice? Is the length right for the channel? Only then does the message go out. Logging who approved each message creates accountability and gives you a trail if a customer later disputes what they were told.
Over time, the prompts become one of the more valuable internal assets a delivery business owns. They capture how your best dispatchers and support leads communicate, and they let newer team members produce consistent replies from their first week. The technology is only as good as the instructions behind it, so the effort is worth putting into clear, specific, well-tested prompts from the start.
Final thoughts for Calgary operators
AI tools can reduce repetitive writing for a cannabis delivery service, but they work best when they are constrained by clear roles, firm boundaries, verified data, and human review. Start with two or three prompts for your most common messages, test them against realistic scenarios, and expand only after the first set performs reliably. Keep records of what you change and why, and revisit the library whenever your licence conditions, service area, or product range changes. A modest, well-maintained set of prompts will serve your customers better than a large collection nobody checks.

Leave a Reply