As If Running a Cannabis Business Isn’t Hard Enough

Questions about artificial intelligence have become a near-daily occurrence at mg Magazine, ranging from which tools to use and how to get started to more nuanced concerns about data privacy, workplace security, and everything in between. This article addresses some of the most common questions that come up in many of these conversations: how to build a working AI system that actually increases productivity on a consistent basis and doesn’t break the bank.

Artificial intelligence tools can make routine work faster, but the first few attempts often create as much frustration as progress. A person may spend hours teaching a chatbot the preferred tone for emails, explaining a project, or correcting repeated mistakes. Then the model changes, a new conversation begins, a feature moves, or the carefully developed context appears to be gone.

That experience can make the technology seem unreliable. Often, however, the underlying problem is not that the person chose the wrong chatbot. The problem is that the workflow depends too heavily on one conversation, one feature, or one model.

ChatGPT, Claude, Gemini, and similar services offer memory, projects, custom instructions, Gems, and other ways to preserve context. Those features can be useful, but they vary by product, account type, subscription, and workplace license. They also continue to evolve. A sustainable approach keeps the essential instructions, examples, and quality controls in a place the user controls.

The goal is not to “train” a chatbot once and expect permanent perfection. It is to build a repeatable work system that can travel from one model to another.

Start With a Job, Not a Brand

The most useful starting question is not “Which AI is best?” It is “What specific job needs to be done?”

Early experiments should involve work that is routine, reversible, and easy to check. Useful examples include:

• Reorganizing nonsensitive notes into an outline.
• Producing a first draft of a routine email.
• Summarizing a document the user is authorized to share.
• Generating questions for an upcoming meeting.
• Converting a long explanation into a checklist.
• Comparing two versions of text and identifying changes.

High-consequence work requires greater caution. An AI-generated answer should not become the sole basis for a legal, medical, financial, hiring, safety, or disciplinary decision. The National Institute of Standards and Technology’s generative AI guidance emphasizes documented processes, risk awareness, evaluation, and human oversight. Those principles apply even when the user is an individual employee rather than a technology specialist.

Build a Portable AI Brief

One long chat can feel as though a model has learned everything about a person’s work. In reality, the system’s access to previous information depends on the product and settings being used. A new chat may not include the same context. Memory can be limited or disabled. Project information is stored within each product’s environment and may not carry over to a new session or a different tool. Deleted conversations may no longer be available.

The safest response is to maintain a short AI brief outside the chatbot. A word-processing file, notes document, or approved company knowledge system can hold the information needed to restart the work.

A practical brief may include:

The brief should contain only information the person is authorized to store and share. It should not become a convenient dumping ground for confidential material.

Use a Prompt That Resembles a Good Assignment

Effective prompting does not require secret commands or elaborate technical language. It requires the same qualities as a good assignment given to a colleague: a clear task, sufficient context, a defined audience, a requested format, and boundaries.

A useful structure is:

Vendor guidance follows similar principles. Anthropic recommends clear, direct instructions, relevant context, examples when useful, and explicit output requirements. Google’s guidance for custom Gems encourages users to specify a persona, task, context, and output format. The language may differ among platforms, but the practical lesson is consistent: Clarity usually matters more than cleverness.

Prompts also work better when large assignments are divided into stages. Instead of requesting a finished report in one step, a user can ask the system to:

• Identify the main points in the source material.
• List missing facts and questions.
• Propose an outline.
• Draft one section at a time.
• Review the complete draft against a checklist.

This process makes errors easier to notice and reduces the temptation to accept a polished response without examining how it was assembled.

Keep the Source Material Separate From the Instructions

AI systems can become confused when instructions, examples, background material, and the requested output are mixed together.

A clearly marked structure also makes the prompt easier to reuse. The stable instructions remain the same, while the source material changes with each assignment.

Users also should tell the model what to do when information is missing. A direction such as “Do not guess; insert [INFORMATION NEEDED]” is more dependable than asking for a confident, complete response regardless of the evidence available. Anthropic’s documentation on reducing hallucinations similarly recommends allowing the model to express uncertainty and grounding answers in provided material and verifiable citations.

Protect Information Before Pasting It

Convenience can encourage people to paste first and consider privacy later. That order should be reversed.

Before entering workplace material into an AI service, an employee should know whether the employer permits the tool and whether an approved business account is available. Consumer accounts and workplace accounts may have different data controls, retention rules, administrative protections, and terms.

Google states that Gemini’s feature availability and data handling can depend on the user’s Workspace license. OpenAI provides data controls for ChatGPT users, while Anthropic offers separate controls and retention options across its products and plans. Those settings are important, but they do not replace workplace policy.

Removing a name may not be enough. A combination of job title, location, date, transaction amount, and unusual circumstances can identify a person or organization. When sensitive information is unnecessary, it should be replaced with neutral placeholders such as [CUSTOMER], [CITY], or [AMOUNT].

The Federal Trade Commission’s own AI use policy prohibits unauthorized disclosure of nonpublic information to generative AI systems and permits internal use only when tools meet agency privacy and security standards. The policy is written for a federal agency, but the principle is broadly useful: authorization and information sensitivity should be considered before the prompt is submitted.

Treat Every Answer as a Draft

Generative AI produces language by predicting plausible responses. Plausible is not the same as verified.

A chatbot may invent a quotation, merge two people, cite a source that does not exist, use outdated information, or state an uncertain conclusion with confidence. It also may perform well on one version of a task and poorly on a nearly identical one.

Before an output enters a workplace document, the user should check:

• Names, titles, dates, prices, percentages, and calculations.
• Quotations against the original recording or transcript.
• Links and citations by opening the source.
• Claims against authoritative, current references.
• Whether confidential information appears in the response.
• Whether the wording introduces stereotypes, assumptions, or unfair conclusions.
• Whether the result actually follows the assignment.

The final reviewer should be able to explain and defend the work without pointing to the chatbot as the authority. AI can assist with a decision, but accountability remains with the person using the output.

Save the Process That Worked

A useful personal AI library does not need hundreds of prompts. A small set of proven workflows is easier to maintain.

For each recurring task, the user can save:

• The reusable prompt.
• A blank input template.
• One approved example.
• The human-edited final version.
• A checklist of frequent errors.
• The date the workflow was last tested.

The date matters because tools change. A prompt that performed well several months ago may produce a different result after a model update. Periodic testing with the same sample assignment can reveal whether the workflow still meets expectations.

When possible, important conversations and outputs should be exported or copied into an approved storage system. ChatGPT, Claude, and Gemini provide various history, activity, export, project, or customization tools, but those platform features should be treated as working conveniences rather than the only archive.

Build for Change

No AI workflow is permanent. Models, interfaces, limits, pricing, privacy controls, and product names can change. The durable elements are the user’s judgment, source material, instructions, examples, and review process.

A sustainable system begins with one low-risk task. It uses a portable brief, a clear prompt, staged work, careful privacy decisions, and a human verification checklist. Once that process works consistently, it can be adapted to another assignment.

The measure of success is not whether a chatbot produces an impressive answer on the first try. It is whether the user can repeat the work, understand the result, detect mistakes, and move the process to another approved tool without starting from zero.