Chatbot
A conversational interface that answers questions or helps with a task. The chat may use a language model, search, business data, or tools.
BrainIT Consulting · Free Field Guide No. 0
AI in Plain English
What AI is, why it matters, how to try it safely, and how human choices shape its impact.
No coding requiredMostly vendor neutralOne safe first experiment
What you will leave withA plain-language mental model, a five-question Ethics Compass, and one completed AI First-Try Card.
You do not need to be technical
AI is often introduced through technical vocabulary, bold predictions, or a demonstration that moves too quickly. None of those is a good starting point for deciding whether it belongs in your business.
Begin with work you already understand. You know what a useful customer reply looks like. You know which facts in a brochure are approved. You know when an employee record is private. You know which decisions require your judgment. Those are the foundations of responsible AI use.
You do not have to trust AI in general. Trust should be earned for a particular job, with particular information, under particular controls.
In this guide, AI means a machine-based system that receives an input and produces an output such as a prediction, recommendation, decision, or piece of content. That broad description includes many kinds of systems. When we discuss tools that write or converse, we will use the more precise term generative AI.
Why AI seems to be everywhere
AI is not new. Businesses have used pattern recognition, recommendations, fraud detection, forecasting, image recognition, speech recognition, and other machine-learning methods for years.
What changed for most people is the interface. Generative AI can respond to ordinary language. Instead of learning a specialist command or preparing a formal data model, a person can describe a task, provide examples, ask a follow-up question, and revise the result.
BEFORESpecialist interfaceMenus, commands, forms, or purpose-built systems
NOWOrdinary languageDescribe, demonstrate, question, and revise
That accessibility makes AI feel sudden. It also makes different technologies look deceptively similar. A chat box may sit in front of a writing model, a search system, a database tool, or an agent that can take actions. The conversation is the interface; it does not tell you what the system can reach or do.
Ask five questions
- What information can it receive?
- What result can it produce?
- What tools or systems can it reach?
- Can it act, or only suggest?
- Who checks the result and remains responsible?
What modern generative AI actually is
A language model learns patterns from large collections of examples. When you provide a prompt, it uses those learned patterns and the current context to generate a likely continuation, piece by piece.
That sounds modest compared with the smooth paragraphs it can produce, but the result can be remarkably useful. The model can reorganize text, propose alternatives, extract a structure, explain an unfamiliar idea, draft from supplied facts, compare options, or transform one format into another.
01 INPUTYour requestInstructions, examples, and permitted information
02 PATTERNSModel workInfer a useful response from context and learned patterns
03 OUTPUTA generated resultDraft, summary, classification, recommendation, image, or code
04 DECISIONHuman reviewCheck accuracy, suitability, fairness, and consequence
The model is not retrieving a perfect stored answer. It is generating a response. That is why the same ability that helps it draft a friendly message can also produce a plausible sentence that is wrong.
A simple mental modelAI can be very good at producing a possible answer. A person still decides whether it is a suitable answer.
What AI is not
Fluent language can make a system feel more certain, informed, or human than it is. Keep these distinctions visible.
NOT AUTOMATIC TRUTHConfidence is not evidenceAn answer may be incomplete, outdated, unsupported, or false.
NOT YOUR BUSINESS MEMORYContext must be suppliedIt does not know which price, promise, policy, or exception is current.
NOT A RESPONSIBLE PERSONDuty stays humanIt does not carry your duty of care, professional judgment, or accountability.
NOT ONE TECHNOLOGYDifferent systems differA recommender, language model, classifier, image generator, and agent have different abilities and risks.
Treat polished AI output as a capable draft, not as proof.
AI also does not need broad access to be useful. A first experiment should use the least information and power needed to produce a reviewable result.
A small-business map of AI
The labels overlap, and vendors do not always use them consistently. This map is meant to orient you, not settle every technical definition.
Generative AI
A system that creates new content such as text, images, audio, video, or code from instructions and context.
Coding assistant
A generative tool focused on explaining, writing, reviewing, or changing software.
Workflow automation
A defined sequence that moves information or work between steps. It may use AI for one step or none at all.
AI agent
A system that can work toward a goal, use approved tools, observe results, and decide what to do next.
MCP server
A connection that lets a compatible AI application discover approved information or tools.
Start smaller than the technology allowsDo not begin by choosing the most powerful category. Begin with the smallest useful job.
What AI can do, and what remains human work
AI is strongest when the task can be described, the inputs are available, and the result can be checked.
Useful AI assistance
- Turn approved notes into a first draft.
- Summarize a non-sensitive document.
- Reformat information into a checklist or table.
- Suggest headings or descriptions.
- Compare two versions and identify differences.
- Extract possible actions for review.
Human responsibility
- Decide why the work should be done.
- Decide what information may be used.
- Recognize private or regulated material.
- Check facts, tone, context, fairness, and consequences.
- Make commitments and important decisions.
- Accept responsibility for the final result.
This is not a contest between people and machines. A useful design assigns each part of the work to the party best equipped to handle it.
Why a small business should care
A small business may not need an “AI strategy.” It may need a better way to prepare a weekly update, organize inquiries, explain a procedure, reuse approved product information, or find what changed between two documents.
AI can be worthwhile when it reduces the blank-page problem, helps a person see patterns, or makes existing information easier to use. It is less attractive when reviewing the output takes longer than doing the work, the necessary information cannot be shared safely, errors would be hard to detect, or the task depends on trust and judgment that should stay personal.
A promising first use
- Repeats often enough to learn from.
- Has a describable input and useful result.
- Has a person who knows how to judge the result.
- Allows mistakes to be caught before they reach a customer or change a record.
- Can be tested with public, fictional, or non-sensitive information.
The goal is not to “add AI.” The goal is to improve one piece of work without giving up control.
Shaped by the past, influencing the present, shaping the future
AI does not arrive outside history. People choose what information is collected, which examples are included, how a system is trained, what it is optimized to produce, where it is deployed, and which voices are heard when problems are found.
PASTPatterns enterKnowledge, creativity, omissions, assumptions, and unfair treatment can all appear in historical information.
PRESENTOutputs influenceDrafts, images, rankings, and recommendations affect what people notice and how they decide.
FUTUREPeople chooseBuilders, owners, workers, customers, and governments shape what is allowed and who benefits.
A model can reproduce or amplify an old pattern even when no one asks it to be unfair. A summary can preserve an important warning or quietly omit it. A recommendation can save time while directing attention away from alternatives.
The future is not chosen by the model alone. It is shaped by the people who build, buy, govern, use, question, and sometimes decline AI systems.
Ethics belongs inside the decisionEthics is not an obstacle added after the useful work. It is how people decide which uses are worth pursuing and under what conditions.
A practical ethics compass
You do not need an ethics department to ask five useful questions.
01TruthCould a person mistake the output for verified fact? What must be checked, sourced, or marked as a draft?
02PermissionDo we have the right to use this information, voice, image, document, or record for this purpose?
03FairnessWho could be helped, overlooked, misrepresented, or unfairly burdened?
04TransparencyWould someone reasonably want to know that AI assisted or influenced the result?
05ResponsibilityWhich named person checks the evidence, owns the decision, and handles a mistake?
These questions do not produce one universal answer. They make the decision visible before convenience quietly becomes policy.
Where AI can go wrong
The most dangerous output is not always absurd. It may be a plausible answer that slips through because it sounds polished.
Confidently wrong
NIST uses the term confabulation for generative AI that confidently presents false or erroneous content. Fluency and factual accuracy are different qualities.
Missing contextThe system does not know an exception or current business rule.
Sensitive informationPrivate data is placed into a tool without understanding storage, access, or terms.
BiasPatterns in data or design produce uneven results.
Automation surpriseA person thinks the system is drafting when it can also send, save, or change records.
Fabricated sourcesA citation, quote, person, product, or policy is invented.
Over-relianceReview becomes a quick approval ritual instead of a genuine check.
Keep the first experiment low consequenceAvoid medical, legal, financial, employment, safety-critical, or similarly consequential decisions. Do not use confidential customer, employee, payment, credential, or health information. Keep the output as a draft and make human review real.
Choose one safe first experiment
Pick the exercise that feels most relevant. You need only one.
Draft a customer message
Use fictional customer details and approved facts. Ask for a courteous draft. Check every fact, commitment, name, and next step. Do not send it from the AI tool.
Summarize a public document
Choose a public page or non-sensitive document you know. Ask for a short summary plus questions or omissions. Compare the result with the original.
Develop marketing ideas
Provide an existing public business description. Ask for three headings and the audience each serves. Reject invented credentials, prices, guarantees, or customer claims.
The experiment passes when…You can inspect the source, see the draft, identify what must be checked, and decide whether the result is actually useful.
Complete the AI First-Try Card
Write the plan before opening an AI tool. You can type directly into the card, copy it, clear it, or print it. This page does not send or store your answers.
AI First-Try Card
One useful draft · limited information · visible human decision
1. Choose the exercise
2. Define the job and information boundary
3. Check the result
4. Use the Ethics Compass
5. Decide the next step
Choose where to go next
One experiment is enough for today. If it was useful, choose the guide that matches the next question it raised.
01Build with an AI coding agentUnderstand how an agent can help create a first app.
02Find the first processChoose repeated work worth improving.
03Protect the dataSet practical information boundaries.
04Give an agent one safe jobDefine purpose, permission, proof, and recovery.
05Build a business MCP serverConnect approved business facts to compatible tools.
06Keep the human voiceUse real and synthetic voice responsibly.
Use the card alone—or invite a second pair of eyes
You can complete the AI First-Try Card entirely on your own. If a second pair of eyes would be helpful, BrainIT can review one proposed workflow with you: what goes in, what should come out, what must stay human, and how you will know whether the experiment helped.
Sources and limits
This guide draws on the OECD definition and principles for AI systems, the NIST Generative AI Profile, UNESCO's Recommendation on the Ethics of Artificial Intelligence, and U.S. Federal Trade Commission guidance on protecting personal information.
The AI First-Try Card, three exercises, and five-question Ethics Compass are BrainIT teaching devices, not official checklists from those organizations.
This is general educational guidance, not legal, employment, privacy, security, financial, medical, or compliance advice. Requirements depend on the information, consequences, contracts, industry, tools, and jurisdiction involved.
Emile du Toit · BrainIT Consulting