Practical AI for Small Businesses

I have spent more than 30 years designing, improving, and explaining technology systems. Today I use AI tools hands-on to build practical applications and help people make sensible technology decisions.

By the end, you should be able to choose one small, low-risk AI experiment for your business.

1. UnderstandWhat an AI language model does, and why it can sound confident while being wrong.
2. CompareThe difference between asking a chatbot a question and assigning an agent a task.
3. InstructHow to explain a job clearly without learning secret prompt phrases.
4. ChooseA useful first project that is easy to review and safe to stop.

You do not need to be a programmer. You do need to know what a good result looks like in your own business.

LLM means large language model. In plain words, it has learned patterns from a very large amount of writing and uses those patterns to create a useful response.

Remember: confident language is not proof that an answer is correct.

Both use an AI model. The practical difference is what you allow them to do.

A chatbot is like asking someone for advice across a desk. An agent is more like giving an assistant a defined assignment, access to selected materials, and a place to leave the finished work. That shift from answering to acting is what makes boundaries important.

An agent should not mean unlimited access. Give it only the files and tools required for the task, and keep approval before sending, buying, publishing, or deleting.

An agent works in a loop: decide what to do, try one step, look at what happened, and choose what comes next.

The checking step is the part people often miss. A useful agent does not simply make a longer answer. It looks at what its last action produced and decides whether to continue, ask a question, correct the work, or stop.

1. GoalUnderstand the result you want.
2. PlanChoose a sensible next step.
3. ActUse an approved tool or file.
4. CheckLook at the result and compare it with the goal.
5. Continue or stopTake another useful step, ask you a question, or deliver the work for review.

Think of it as a careful assistant following a checklist, not as magic.

If the destination is vague, an agent can move quickly in the wrong direction.

A useful assignment names the work, the desired result, and what the agent must not do.

1. Find and readOpen only the three approved proposal files.
2. ExtractPull out prices, dates, terms, and promised services.
3. Notice gapsMark information that is missing or unclear.
4. ComparePlace like information side by side.
5. CheckVerify the comparison against the original proposals.
6. DeliverPrepare a recommendation for a person to review.

A good agent does not keep going merely because it can. It pauses and checks whether the next step is useful and allowed.

• Did the tool or step work?
• Is the important information complete?
• Does the result match the assignment?
• Are important facts supported by the source?
• Is a person’s approval required now?
• Has the task reached a sensible stopping point?

Stop when the work is complete, blocked, unsafe, missing essential information, or waiting for a human decision.

These words describe different parts of the working arrangement. You do not need to memorize them.

The goal is not to give an agent every tool. Give it the smallest useful set with clear boundaries.

“Prompt engineering” is mostly the skill of explaining a job clearly. You do not need secret phrases.

The best prompts usually sound less like computer commands and more like a thoughtful handoff to a new colleague. Explain the situation, point to the right information, describe the finish line, and say where the person should pause for you.

1. GoalWhat should be accomplished?
2. ContextWhat should the AI understand about your business or situation?
3. InformationWhich notes, files, or facts may it use?
4. BoundariesWhat must it avoid or ask permission to do?
5. ResultWhat should it produce, and in what format?
6. CheckHow should it verify the work before giving it to you?

A useful shortcut: goal + background + information + boundaries + result + check.

A little context and one clear boundary can turn a vague request into work you can safely review.

Useful AI work is often ordinary work made easier, clearer, or more consistent.

Begin with work that is useful, repeated, easy for you to judge, and safe to redo.

The more serious the consequence of a mistake, the more human review and restricted access you need.

This is not a reason to avoid AI. It is a reason to match the freedom you grant with the risk of the job. Drafting a checklist can be low risk. Sending a promise to a customer, moving money, or changing an important record deserves a deliberate human decision.

• Work from copies rather than irreplaceable originals.
• Never paste passwords, banking credentials, or secret keys.
• Use the least customer-identifying information needed.
• Review names, facts, prices, dates, and promises.
• Require approval before sending or publishing.
• Require approval before deleting or changing records.
• Ask for sources when an important claim must be correct.
• Keep a person responsible for the final decision.

Hermes is one example of a personal AI agent. It runs on your computer, connects an AI model to tools you approve, and works through tasks in steps.

For a first session, use a new practice folder and a task that cannot affect a customer or business system.

On Windows or macOS, the official Hermes Desktop installer is the easiest route. Use the command-line installer only if you are comfortable opening a terminal.

Use only the current instructions at the official Hermes documentation. On Windows, open a new terminal after installation so the hermes command is available.

If your ChatGPT account includes Codex access, Hermes can connect through a browser sign-in. You do not need a separate OpenAI API key for this path.

1. Choose OpenAI CodexSelect it from the model setup list.
2. Open the sign-in pageHermes displays a web address and a short one-time code.
3. Sign inUse your OpenAI or ChatGPT account and approve the code.
4. Return to HermesChoose an available Codex model, then start with hermes.

A separate Codex command-line installation is not required. Available models and usage limits depend on the OpenAI account and plan. See the official Hermes provider guide.

Hermes offers several AI providers, but a beginner only needs one that already fits their account and budget.

If cost is uncertain, stop during setup and confirm the provider’s billing rules before continuing.

Provider requirements can change. Check the current official Hermes provider guide.

Start in a new practice folder. Nothing in this first exercise should affect a customer, business record, or irreplaceable file.

1. CreateMake an empty folder named AI workshop practice.
2. OpenOpen Hermes in that folder and ask what tools are available.
3. Set the boundaryTell it not to change or delete existing files.
4. Make one test fileAsk for workshop_test.md with three facts you provide.
5. VerifyAsk Hermes to read the file back and confirm that every requested item is present.

Let the agent ask questions before it recommends a tool or an automation.

• You explain how the work really happens.
• The agent asks before it advises.
• The result separates easy improvements from automation.
• The first recommendation stays small and testable.

Turn rough conversation notes into organized work without allowing the agent to contact the customer.

Before using the draft: check the customer’s name, dates, prices, promises, and every statement about what happens next.

Do not automate a whole department. Run one small experiment and keep notes about what happened.

1. Day 1: ChoosePick one repeated task you understand well.
2. Day 2: MeasureRecord how long it takes and what makes it difficult.
3. Day 3: ExplainWrite the goal, background, information, boundary, result, and check.
4. Day 4: TryAsk AI to prepare one draft or analysis.
5. Day 5: ReviewRecord what was useful, wrong, missing, or awkward.
6. Day 6: ImproveRevise the instruction and try once more.
7. Day 7: DecideIs the improvement worth repeating with human review?

Look for time saved, fewer missed details, easier follow-up, better consistency, or less stress—not a dramatic “AI transformation.”

You can use this guide on your own. If you would like a second opinion about a process, a safe first experiment, or whether a custom solution is justified, you are welcome to contact me.

No obligation and no technical vocabulary required. Bring the business problem, and we can begin by talking through how the work happens today.