AI foundations for professionals

Why this module

Most people's first week with an AI assistant goes the same way: a few impressive answers, one embarrassing mistake, and then the tab quietly closes. The people who get real work out of these tools are not more talented. They carry a few mental models about what the tool is and where its edges are.

This module gives you those models: what a language model actually does, which of your tasks it should touch, which three tools to set up, what never to paste into them, and a daily loop that keeps you using them after the novelty wears off.

Lesson 1.1: What a language model actually does

A modern AI assistant like ChatGPT, Claude, or Gemini is built on a large language model, which does one thing: given some text, it predicts what text should come next. It learned to do that by reading an enormous amount of writing during training, then being tuned by people to be helpful and follow instructions.

That single fact explains almost everything you will notice. It explains the fluency: prediction over billions of pages of human writing produces text that reads like a competent person wrote it. It explains the knowledge cutoff: the model learned from text up to a certain date, and unless the tool has web search (most now do), it knows nothing after that date, or about your company.

Most of all, it explains why the model sometimes states things that are flatly wrong with total confidence. It is not looking anything up in a database of facts; it is producing the most plausible continuation. Usually the plausible answer is also the correct one. When it is not, nothing in the tone tells you. People call this a hallucination. Treat it as a permanent feature of the technology, not a bug awaiting a patch.

Four more terms you will run into:

• Tokens. Models read and write in chunks called tokens, roughly three-quarters of an English word each. Usage limits and prices are set in tokens.
• Context window. How much text the model can hold in view at once: your prompt, attached files, and the conversation so far. Large, but not infinite; in a very long chat, early details slip.
• Reasoning models. Models tuned to work through a problem before answering. Better at math, logic, and planning, and slower. Most assistants let you switch a thinking mode on for hard problems.
• Multimodal. Most assistants also handle images, voice, and files, so you can photograph a whiteboard or talk while you drive.

A worked example

Suppose you ask an assistant, "What does our refund policy say about damaged goods?" Without the policy, it can either admit it does not have the document or produce a plausible-sounding policy that a company like yours might have. Older models did the second thing constantly; current ones are better about admitting gaps, but the pull toward a fluent answer never goes away.

Now paste the policy in and ask again. The model answers from the text in front of it. That is the pattern behind almost every reliable use of AI at work: give it the real material and ask it to work from that.

Exercise

Answer its questions honestly and save the response for Lesson 1.2.

You are an experienced coach helping a professional understand what an AI assistant can and cannot do reliably in their job.

My role: [YOUR JOB TITLE AND WHAT YOU ACTUALLY DO]
My industry: [YOUR INDUSTRY]
Tools I use most: [E.G. OUTLOOK, EXCEL, SALESFORCE]

Before answering, ask me up to four clarifying questions about my typical week and wait for my reply. Then give me:
1. Five tasks in my job where an AI assistant is reliably useful, and why.
2. Five tasks where it is likely to be wrong, incomplete, or risky, and what I should verify.
3. Two habits that would help me catch its mistakes.
Three short headed lists, plain language, under 400 words.

Lesson 1.2: What AI is good at, bad at, and what needs a human check

The fastest way to get value from AI is to stop asking whether it is smart and start sorting tasks by type.

Where it is strong. First drafts of anything written. Summarizing a document you give it. Rewriting for tone, length, or audience. Explaining an unfamiliar concept at the level you ask for. Brainstorming when you have no options. Turning messy notes into structure. Converting formats: bullets to prose, a paragraph to a table. Writing spreadsheet formulas and short scripts. Playing a critic or a skeptical reader.

Where it is weak. Facts it has not been given, especially numbers, dates, names, and citations. Arithmetic done in its head across long lists (it should write code for that; see Module 5). Recent events, unless it has search. What is true inside your organization: who owns what, what was decided last week, what the client actually said. Judgment calls that depend on context it lacks.

The rule of thumb. AI is most valuable on tasks that are expensive to produce and cheap to verify. A proposal draft takes you an hour to write and five minutes to read critically: a great trade. A legal citation takes the model two seconds to produce, takes you twenty minutes to confirm, and may be fabricated: a bad trade unless the tool is grounded in sources you can click.

The three buckets.

1. Delegate and skim. Low stakes, easy to check. A meeting recap from a transcript, a rewrite of your own paragraph, a first-pass agenda.
2. Draft and verify. Medium stakes. A customer-facing email, a summary of a contract you will still read, a formula you will test on a copy.
3. Human only, AI as sparring partner. High stakes or regulated: final numbers in a board report, medical or legal conclusions, anything involving a named person's private data. AI can help you think, but a human owns every word that leaves the building.

A worked example

Suppose a customer service team lead's week includes a weekly performance summary, knowledge base updates, and prep for a one-on-one with a struggling rep.

• Weekly summary: draft and verify. Paste the raw numbers, get a draft, check every figure against the source.
• Knowledge base updates: delegate and skim. Paste the resolved ticket, ask for an article in the house format.
• One-on-one prep: human only. Never paste the rep's name or HR notes into a consumer tool. Describe the situation generically and ask for coaching questions.

Exercise

Use the task list from your Lesson 1.1 answer, or write a fresh one.

You are a practical operations advisor. Sort each task below into exactly one of three buckets:

A. Delegate and skim (low stakes, easy to check)
B. Draft and verify (medium stakes; I will edit and confirm)
C. Human only, AI as sparring partner (high stakes, regulated, or involves private data about named people)

My role: [YOUR JOB TITLE]
Constraints I work under: [E.G. HIPAA, CLIENT CONFIDENTIALITY, NONE THAT I KNOW OF]
My tasks:
[PASTE YOUR TASK LIST, ONE PER LINE]

Return a table with columns: Task, Bucket, What I should give the AI, What I must verify afterward. If a task is unclear, ask me instead of guessing. Keep each cell under 20 words.

Lesson 1.3: The three tools everyone should have

You do not need twenty AI tools. You need three, chosen deliberately, and you need to actually open them.

1. A general assistant you pay for. ChatGPT, Claude, or Gemini. All three have a free tier; the paid individual plans add the stronger models, higher limits, and features like projects and memory. Pick one and go deep; Module 3 helps you choose.

2. The AI that already lives in your work suite. If your company runs Microsoft 365, that is Copilot, inside Outlook, Word, Excel, Teams, and PowerPoint, seeing your files and mail with the permissions you already have. On Google Workspace, it is Gemini inside Gmail, Docs, and Sheets. It works with your real documents without pasting, and IT has usually already reviewed it. Availability depends on what your organization licenses, so ask.

3. A source-grounded research tool. For questions about the outside world, Perplexity answers with citations you can click. For questions about your own documents, NotebookLM lets you upload the sources and answers only from them, again with citations. Both tie every answer to something you can check.

A worked example

Say you are a project manager on a Tuesday. A stakeholder asks about a vendor's compliance certification: you ask Perplexity, click the citation, and confirm it on the vendor's site before replying. A 40-page requirements document lands: you add it to NotebookLM with the last three status reports and ask which requirements changed since July, with page references. You need a status email: Copilot in Outlook drafts it from the thread and you fix the tone. Three tools, three jobs, nothing pasted where it should not be.

Exercise

Every major assistant has a settings area for custom instructions or personalization. Draft your profile with this prompt, then paste the result there.

Help me write a short personal profile for the custom instructions of an AI assistant, so it gives better answers without my repeating myself.

About me:
- Role: [YOUR JOB TITLE AND WHAT YOU ACTUALLY DO]
- Industry and company size: [E.G. REGIONAL ACCOUNTING FIRM, 40 PEOPLE]
- Who I usually write for: [E.G. CLIENTS, MY BOSS, MY TEAM]
- How I like answers: [E.G. SHORT, BULLETS FIRST, PLAIN ENGLISH, NO HYPE]
- Things I never want: [E.G. EMOJI, EXCLAMATION POINTS, MADE-UP STATISTICS]

Ask me two questions first if anything is too vague. Then write the profile in first person, under 150 words, in two parts: "About me" and "How to respond to me", including a line telling the assistant to flag uncertainty rather than guess.

Lesson 1.4: Privacy ground rules

Every useful thing AI does at work involves showing it information. The question is never "is AI safe" in the abstract; it is "what am I allowed to show this tool, on this plan, in this job." Five rules cover most situations.

Rule 1: Know your organization's policy before you need it. Many companies now have a written AI policy, an approved tool list, or both. Read it. If none exists, ask your manager or IT lead: "Which AI tools am I allowed to use, and what am I allowed to put in them?" Get the answer in writing.

Rule 2: Never paste confidential data into a consumer AI tool unless your organization has approved it. Confidential means client and customer records, employee information, unreleased financials, contracts, source code, credentials, and anything covered by a regulation. Consumer tools are the free and individual plans anyone can sign up for. Business and enterprise plans typically carry contractual data protections, and approved tools have been reviewed for this purpose.

Rule 3: Understand the training question. Some consumer plans may use your conversations to improve future models unless you turn that off in settings; business and enterprise plans generally do not. Defaults change, so check the vendor's current data-usage documentation and turn off training where you can.

Rule 4: Strip what does not need to be there. Most tasks work just as well with names replaced by roles, account numbers removed, dollar figures rounded, and dates shifted. Redact before pasting, not after. Once text is sent, you cannot recall it.

Rule 5: Assume the transcript could be read by someone else. Chat histories can surface in legal discovery, security reviews, or a breach. Do not type anything you would not want to see with your name attached in a court filing.

If you work in a regulated role, the rules tighten. Healthcare: protected health information under HIPAA never goes into an unapproved tool. Legal: attorney-client privileged material and client confidences. Finance and accounting: material nonpublic information and client financial data under GLBA and SOX. HR: personnel records, medical accommodations, anything touching EEOC-protected characteristics. Education: student records under FERPA, especially for minors. Insurance: policyholder PII and claims data. In all of these, "my organization approved this tool for this data" is the only green light.

A worked example

Suppose an HR manager wants help preparing for a performance conversation with an employee who keeps missing deadlines. The wrong way: paste the employee's name, the last two reviews, and an email thread into a free chat tool. The right way: "I manage a mid-level analyst who has missed four deadlines in two months after a strong first year. I want a supportive but direct conversation. Give me an outline and six questions to ask." No name, no records, no thread. Same quality of advice, and nothing sensitive left the building.

Exercise

Build your do-not-paste list and keep it within reach.

You are a pragmatic compliance advisor who understands how people actually use AI at work. Help me build a one-page checklist of what I can and cannot put into AI tools.

My role: [YOUR JOB TITLE]
My industry and any regulations that apply: [E.G. HEALTHCARE / HIPAA, LEGAL / PRIVILEGE, FINANCE / GLBA, OR "NONE I KNOW OF"]
What my organization has approved: [E.G. MICROSOFT 365 COPILOT ONLY, NOTHING WRITTEN DOWN YET]
Types of information I handle daily: [LIST THEM]

Ask me up to three questions if you need more detail. Then produce three headed lists: "Never paste", "Paste only in approved tools", and "Fine after light redaction" (with the redaction step for each), plus three questions to ask my IT or compliance lead this week. Under 350 words. Practical workplace rules only, not legal advice.

Lesson 1.5: The habit loop for daily use

Knowing how AI works does nothing for you if you forget to open it. People who get the most from these tools attach them to specific moments in the day. A habit has three parts: a cue, a routine, and a reward.

Cues. Pick three moments that already happen every day: before writing anything longer than a paragraph, ask for an outline or rough draft; when stuck for more than five minutes, explain the problem to the assistant as if to a colleague; after reading something long, ask for a summary and check it against your own understanding.

Routine. Keep the assistant in a pinned tab or a desktop app with a keyboard shortcut. If it takes four clicks to reach, you will not reach for it.

Reward. At the end of each day, spend two minutes noting one thing AI did well and one thing it got wrong. After two weeks you will have a personal map of where the tool helps you.

The weekly layer. Once a week, move any prompt you used more than once into a personal prompt library (see Module 2), and try one new thing: a spreadsheet upload, voice mode, a critique of your work.

The first-week plan. Choose three recurring tasks from your Lesson 1.2 table that landed in "delegate and skim" or "draft and verify". Use AI for those three, and nothing else, for five working days; on Friday, decide which ones stay.

A worked example

Suppose an account manager picks weekly client status emails, CRM call-note summaries, and questions before renewal calls. Monday, the status emails take half the usual time. Wednesday, a call-note summary misses a pricing detail, which goes in the log. Friday, the renewal questions surface an angle the manager had not considered. All three stay, plus one new step: checking every number in the summaries.

Exercise

Have the assistant build your first week with you.

You are a productivity coach who helps busy professionals build new work habits. Design my first week of using an AI assistant on real tasks.

My role: [YOUR JOB TITLE]
My three chosen tasks: [TASK 1], [TASK 2], [TASK 3]
When each usually happens: [E.G. STATUS EMAILS MONDAY MORNING, CALL NOTES AFTER EVERY CALL]
The tool I will use: [CHATGPT / CLAUDE / GEMINI / COPILOT]
My biggest obstacle: [E.G. I FORGET, I DO NOT TRUST THE OUTPUT, I HAVE NO TIME TO LEARN]

Ask me two clarifying questions first. Then give me a five-day plan (under 60 words per day, naming the cue that triggers each task), a two-line end-of-day log template, and a note on how to decide on Friday which habits stay. Plain language, no motivational filler.

Module wrap-up

A language model predicts text: fluent, fast, and sometimes confidently wrong. It is strongest on work that is expensive to produce and cheap to verify. Three tools cover most needs, five privacy rules keep you out of trouble, and a cue-based habit keeps you using it.

Before moving to Module 2, check off each item:

• I can explain in one sentence what a language model does and why it hallucinates.
• I have sorted at least ten of my real tasks into the three buckets.
• I have chosen my primary assistant and pasted a profile into its custom instructions.
• I know which work-suite AI my organization provides, or I have asked.
• I have tried Perplexity or NotebookLM on a real question.
• I have read my organization's AI policy or asked for one in writing.
• My personal do-not-paste list is written and within reach.
• I have picked three recurring tasks and three cues for my first week, and a place to keep the daily log.

Related reading on the site: How large language models work, AI hallucinations explained, and AI privacy at work.

Next up: Module 2 turns vague requests into prompts that produce usable work on the first or second try.