Basic AI Tools for Everyday Work

Lesson 1

What AI tools are, and what they do well and badly

Who this course is for

People new to AI tools, who may have tried a chatbot once or twice and want to use one at work with confidence. Each module takes about 45 minutes, split into lessons of 5 to 10 minutes, and ends with a quiz.

Practice rule: every exercise uses made-up or public sample text. Never practice with real customer, employee or client data.

Usage and accuracy figures here come from surveys, benchmarks and secondary summaries, and the source list says which. Accuracy changes as tools change, so practice on the tools your employer approves, and follow your employer's own AI policy first.

What you'll learn

• Name the four main kinds of AI tool you are likely to meet at work.
• For each kind, say one task it does well and one it does badly.
• Explain why a fluent answer is not the same as a correct one.

Lesson 1.1: The main kinds of tool

• General chatbots (for example ChatGPT, Gemini, Copilot, Claude, Meta AI) write, rewrite, summarize text you paste in, brainstorm and explain jargon.
• AI answers in search results give a summary above or beside the list of links.
• Transcription and meeting-note tools turn speech into text, then summarize it and list actions.
• Assistants built into office software help with writing, email and spreadsheets inside the programs you already use.

Image and slide tools and spreadsheet formula helpers also exist. Their accuracy and rules are less well researched for this course, so check their output carefully and follow your company's approved list.

Lesson 1.2: What chatbots do well and badly

Good at: first drafts, rewriting for a tone, summarizing a text you paste in, brainstorming, and explaining a term.

Bad at: facts from memory, exact citations, recent events, and anything you cannot check.

How common is this? Pew Research Center's survey of 5,119 US adults (fielded 17 to 23 February 2026) found about half of US adults never use a chatbot and about a quarter use one daily [Source 1]. A secondary summary of the same survey reports that 44% of US adults have used ChatGPT [Source 1].

Lesson 1.3: Fluent is not the same as correct

A chatbot writes in confident, smooth sentences whether it is right or wrong. That is why every use needs a check. Module 3 covers the checks.

Worked example

A new team member asks a chatbot for the deadline in the EU rules on AI transparency. The chatbot gives a date with a confident tone. She does not paste it into her report. She checks the official EU source, finds the correct date, and notes the source she used. The chatbot helped her find the right place to look; it did not replace the check.

Key terms

• Chatbot: software you talk or type to, which writes replies.
• Prompt: the instruction or question you type to a tool.
• Large language model (LLM): the kind of software behind most chatbots. It predicts likely text from patterns in its training data.

Practice

Ask an approved chatbot to explain one work term you use but do not fully understand, using no real data. Then write down two things you would check before relying on the explanation.

Job aid: which tool for which job

• Draft or rewrite: chatbot.
• Summary of a meeting: transcription tool, then check.
• Fact or citation: check a primary source.
• Approved list: check before use.

Check what you learned

1. Which task is a chatbot generally good at?
   • Recalling exact citations from memory
   • Drafting and rewriting text you provide
   • Knowing today's news
   • Verifying a fact without a source
2. Pew's February 2026 survey found roughly what share of US adults never use a chatbot?
   • About a tenth
   • Everyone
   • About nine in ten
   • About half
3. Why does a confident-sounding answer still need checking?
   • It is always wrong
   • Tools write fluently whether or not the answer is correct
   • It is always out of date
   • Only lawyers can check it

Lesson 2

Writing and research with a chatbot

What you'll learn

• Write a clear prompt that states the task, the audience, the length and any limits.
• Give a chatbot the text it needs, rather than asking it to recall facts.
• Improve a draft in two or three rounds, then finish it yourself.

Lesson 2.1: A good prompt has four parts

• Task: what you want, in one sentence.
• Audience: who will read it.
• Length and format: for example "five bullet points" or "under 150 words".
• Limits: what to leave out, and what to say if it does not know.

Example: "Draft a short email to a supplier asking for a delivery date. Audience: a procurement manager who knows us. Under 120 words. Do not invent dates or prices; leave a [date] placeholder."

Lesson 2.2: Give the text, do not rely on memory

The most reliable use of a chatbot is working on text you already have. Paste the meeting notes and ask for a summary. Paste your draft and ask for a clearer version. This is much safer than asking the chatbot what a policy says from memory. Summarizing a document you supply still produces errors in some cases. Vectara's October 2026 benchmark puts that rate at about 7 to 14% for most flagship models [Source 2]. Check the summary against the original.

Lesson 2.3: Improve in rounds

Round one: ask for a first draft. Round two: ask for a specific change, such as "shorter, plainer, more formal". Round three: check the facts and finish by hand. Keep your own final edit; the output is a draft, not the final work.

Worked example

A coordinator needs a short note to staff about a new room booking rule. She writes: "Task: draft a note announcing the new booking rule. Audience: all staff. Under 100 words. Use the rule text below only; do not add anything. [rule text pasted]." The draft is clear. She changes one sentence to match the company's tone, checks the date against the policy, and sends it.

Key terms

• Prompt: the instruction you type. A good prompt states the task, audience, length and limits.
• Draft: a first version. It is not finished until a person has checked and edited it.

Practice

Write a four-part prompt to draft a 100-word note inviting colleagues to a sample team lunch. Use no real names. Then run it and make one improvement round.

Job aid: a prompt in four lines

• Task: ___
• Audience: ___
• Length and format: ___
• Limits: do not invent facts; mark gaps with [brackets].

Check what you learned

1. Which is the four-part prompt structure?
   • Task, audience, length and format, limits
   • Hope, wish, ask, pray
   • Name, date, place, cost
   • Question, answer, source, date
2. Which approach is most reliable?
   • Paste the policy text and ask for a summary, then check it
   • Ask what a policy says from memory
   • Ask for the answer without any text
   • Trust the first reply
3. What should you do after the first draft?
   • Send it
   • Ask for a longer version every time
   • Delete it
   • Ask for specific changes, check facts, and finish it yourself

Lesson 3

Checking AI output before you use it

What you'll learn

• Explain what a hallucination is and why error rates vary by task.
• Check names, numbers, dates and citations against a primary source.
• Decide when a result is safe to use, and when to stop and ask.

Lesson 3.1: What a hallucination is

A hallucination is a confident answer that is false. The tool is not lying; it produces likely-sounding text, and sometimes that text is wrong. Some experts prefer the term "factual error" [Source 3]. The error rate depends on the task. Summarizing a supplied document, answering from memory, legal questions, and citations each produce different rates, so you cannot assume one number applies to all your work [Source 2; Source 4].

Lesson 3.2: Real cases

Real cases show the cost. A court database lists 2,202 decisions where a party relied on hallucinated content, as of 19 October 2026 [Source 4]. In 2024, a Canadian tribunal held Air Canada responsible for wrong fare advice given by its chatbot [Source 4]. The lesson for office work is simple: whoever sends the output owns it.

Lesson 3.3: The four-point check

• Names and numbers: compare each against the original.
• Dates and deadlines: confirm them in the official source.
• Citations: open the source and confirm it says what the text claims. If you cannot find it, remove it.
• Tone and claims: remove any certain-sounding claim you cannot support.

Lesson 3.4: When to stop

Stop and ask a colleague, a manager or a specialist when the output involves legal, medical, financial or safety advice; a decision about a person; or a figure someone will rely on in public. A chatbot can help you prepare the question, but the answer must come from someone accountable for it.

Worked example

A chatbot writes a summary of a report and includes a percentage with a citation to "a 2025 industry survey". You look for the survey and cannot find it. You remove the figure and cite only the report you read. You note in your file that the chatbot's citation could not be verified.

Key terms

• Hallucination: a confident answer that is false [Source 3].
• Primary source: the original document, dataset or official record, not a summary of it.
• Grounding: when a tool looks up documents before answering. It reduces, but does not remove, errors [Source 5].

Practice

Take a short public article you already know. Ask a chatbot to summarize it, then mark every name, number and date in the summary. Check each against the article and write down how many were correct.

Job aid: before you use AI output

• Names and numbers checked?
• Dates confirmed?
• Citations opened?
• Claims you cannot support removed?
• Specialist needed?

Check what you learned

1. What is a hallucination?
   • A deliberate lie
   • A software update
   • A confident answer that is false
   • A type of spreadsheet
2. A chatbot cites a survey you cannot find. What should you do?
   • Keep the figure, since the chatbot said it
   • Remove the figure, or confirm it in the original
   • Change the number to a round one
   • Ask the chatbot again until it finds the survey
3. When should you stop and ask a specialist?
   • Never
   • For legal, medical, financial or safety advice, or a figure others will rely on
   • Only for long texts
   • Only on Fridays

Lesson 4

Meeting notes and transcripts

What you'll learn

• Explain how transcription and meeting-note tools work, and where they make mistakes.
• Set up a meeting-note tool so you use it with the right permission.
• Treat a transcript as a draft record and check the action items before sharing.

Lesson 4.1: What the tool does

A transcription tool turns speech into text. A meeting-note tool then summarizes the text and lists actions and decisions. Both steps can make mistakes. Word error rate (WER) is the share of words a transcript gets wrong, measured against a human reference [Source 6]. The Open ASR Leaderboard, dated 19 September 2026, reports a median word error rate of about 9.66% on AMI meeting audio, compared with 1.35% on read audiobooks [Source 6].

Lesson 4.2: Where it goes wrong

• Several people talking at once.
• Noisy rooms, poor microphones, and video calls with dropped audio.
• Names, product codes and jargon the tool has not heard before.
• Accents and fast speech.

A vendor test reports accuracy falling 15 to 25 percentage points when two or more people talk at once [Source 7]. Treat it as a direction, not a precise figure.

Lesson 4.3: Permission and records

Before you record or transcribe, tell the people in the meeting and ask whether they agree. Check your company's rule on recording. Do not use a tool your company has not approved for meetings with confidential content. A transcript is a draft; it is not the official record unless your company says so.

Lesson 4.4: Check the action list

The most useful part of the tool is often the action list, and also the most risky. Check each action: who owns it, what the task is, and the date. Correct any name the tool got wrong before the notes go to anyone.

Worked example

A transcription tool records a team meeting with four people. The summary says "Sam to send the revised budget by Friday". In the recording, Sam agreed to review it, not send it. The coordinator corrects the action before sharing the notes, and asks the team to confirm the owner.

Key terms

• Transcript: a text record of speech, usually with errors to check.
• Word error rate (WER): the share of words a transcript gets wrong [Source 6].

Practice

Record a short practice conversation between two people reading a made-up script. Run it through an approved tool. Count the errors in names and action items, and write one change you would make to your setup.

Job aid: meeting notes

• Permission asked?
• Approved tool?
• Actions checked for owner, task and date?
• Labeled as a draft?

Check what you learned

1. What is word error rate?
   • The number of typing mistakes you make
   • The price of a tool
   • The number of meetings
   • The share of words a transcript gets wrong, against a human reference
2. Which situation most often reduces transcription accuracy?
   • Several people talking at once
   • One clear speaker in a quiet room
   • Reading a prepared script
   • A short meeting
3. Before you share meeting notes, what must you check?
   • Each action's owner, task and date
   • The color of the summary
   • The font size
   • The number of emojis

Lesson 5

Privacy, copyright and disclosure at work

What you'll learn

• Decide which data can go into which tool, using your company's rules.
• Explain what copyright means for AI output and why you should keep notes on your changes.
• Say when and how to disclose AI use, including the EU transparency rule for chatbots.

Lesson 5.1: Personal data

Under the GDPR, personal data is information about an identified or identifiable living person. Pseudonymized data can still count as personal data [Source 9]. Do not paste customer, employee or client personal data into a public chatbot unless your company has approved that tool for that data. Vendor guidance commonly advises training staff on this rule [Source 10].

Before you use a tool, ask three questions: Is it approved for work? Is the data personal, confidential or customer data? Does the tool use my inputs to train its models? If you cannot answer, ask IT or your data protection lead first.

Lesson 5.2: Copyright and AI output

The US Copyright Office's Part 2 report (29 January 2025) says copyright protects only human authorship, and prompts alone do not make the person who wrote them the author of the output [Source 11]. Human contributions such as selection, arrangement and edits can be protected [Source 11]. In practice, AI output may not be yours to own or to stop others copying. Keep a note of what you changed, and credit the sources you read. Rules differ by country, so check your own.

Lesson 5.3: Disclosure

Follow your company's disclosure rule first. The EU AI Act's Article 50 transparency duties applied from 2 August 2026 [Source 12]. Chatbots must tell people they are dealing with AI [Source 13]. Organizations that publish AI-generated text on matters of public interest must disclose it, unless a person with editorial responsibility has reviewed it [Source 14]. These EU rules apply in the EU; check the rules in your own country. A simple disclosure line is often enough: "This draft was prepared with an AI tool and checked by [name]."

Lesson 5.4: Deepfakes

A deepfake is AI-made image, audio or video that resembles real people, places or events and would falsely look authentic [Source 14]. Do not create or share one of a real person. If you receive one, check with the person it shows through a known channel before you act on it.

Worked example

A staff member wants to paste a list of client names and contract values into a free chatbot to draft a renewal email. The company's policy says client data may only go into approved tools. She does not paste the list. She uses the approved tool with a made-up example to write the email template, then adds the real client details herself, in the approved system.

Key terms

• Personal data: information about an identified or identifiable living person (GDPR) [Source 9].
• Shadow AI: AI tools used at work without the employer's approval.
• Deepfake: realistic, falsely authentic AI-made media of a real person or event [Source 14].

Practice

Write your company's rule for three kinds of data: what can go into approved tools, what can never go into a public chatbot, and who to ask when unsure. If your company has no rule, write the question you would ask your manager.

Job aid: before you paste or publish

• Is the tool approved?
• Is there personal or confidential data?
• Do I know whether inputs train the tool?
• Have I changed the output and kept notes?
• Does my disclosure rule apply?

Check what you learned

1. A customer's name and contract value are in a file. Can you paste them into an unapproved public chatbot?
   • Yes, if you delete the surname
   • Yes, always
   • No, not under a rule that bars personal data in unapproved tools
   • Only on weekends
2. According to the US Copyright Office (January 2025), what does a prompt alone do?
   • Makes you the author of the output
   • Puts the output in the public domain automatically
   • Does not, by itself, make you the author
   • Transfers ownership to the tool
3. What is the best first step on disclosing AI use at work?
   • Hide it
   • Never write anything with AI
   • Claim the work was done by hand
   • Follow your company's disclosure rule, and add a short honest note if none exists

Lesson 6

Final check

Ten questions from all five modules

Answer all ten, then check your answers. Many teams start with a pass mark of 80% (8 of 10). If you miss one, go back to the module named in the question.

Check what you learned

1. Which task is a chatbot generally good at? (Module 1)
   • Recalling exact citations
   • Verifying facts without sources
   • Knowing today's news
   • Drafting and rewriting text you provide
2. What makes a fluent answer risky? (Module 1)
   • It is always wrong
   • Tools write fluently whether or not the answer is correct
   • It is always old
   • It is always long
3. What are the four parts of a good prompt? (Module 2)
   • Hope, wish, ask, pray
   • Name, date, place, cost
   • Task, audience, length and format, limits
   • Question, answer, source, date
4. What is the most reliable way to get a summary? (Module 2)
   • Paste the text you have, then check the summary
   • Ask from memory
   • Trust the first reply
   • Ask for a longer answer
5. What is a hallucination? (Module 3)
   • A deliberate lie
   • A software update
   • A confident answer that is false
   • A spreadsheet error only
6. A chatbot cites a source you cannot find. What do you do? (Module 3)
   • Remove the claim, or confirm it in the original
   • Keep it
   • Change the number
   • Ask the chatbot again until it finds one
7. What is word error rate? (Module 4)
   • Typing mistakes you make
   • Share of words a transcript gets wrong, against a human reference
   • Number of meetings
   • Cost of the tool
8. What must you check before sharing meeting notes? (Module 4)
   • Each action's owner, task and date
   • Colors
   • Font size
   • Emoji count
9. Does a prompt alone make you the author of AI output under the US Copyright Office's January 2025 report? (Module 5)
   • Yes
   • Only if you pay
   • Only for images
   • No
10. Before pasting client data into a tool, what must you check? (Module 5)
    • That the text is short
    • That the tool is popular
    • That the tool is approved for that data
    • That the font is clear

Lesson 7

Glossary and sources

Glossary

• Chatbot: software you type or talk to that writes replies.
• Deepfake: realistic, falsely authentic AI-made media of a real person or event.
• Draft: a first version that a person must check and finish.
• Grounding: a tool looking up documents before answering.
• Hallucination: a confident answer that is false.
• Large language model (LLM): software that predicts likely text from patterns in training data.
• Personal data: information about an identified or identifiable living person.
• Primary source: the original document or record, not a summary.
• Prompt: the instruction or question you type.
• Shadow AI: AI tools used at work without approval.
• Transcript: a text record of speech, to be checked.
• Word error rate (WER): share of words a transcript gets wrong.

Sources

Source numbers in the lessons match this list. Many figures come from secondary summaries or vendors, and the list says so; check the original before quoting them. The EU rules in Module 5 apply in the EU; check the rules where you work.

• 1. Pew Research Center, Americans and AI 2026 (survey of 5,119 US adults, fielded 17 to 23 February 2026). Usage by name via a secondary summary; check Pew.
• 2. Vectara hallucination leaderboard summaries (October 2026), via secondary summary (count.co); check the original.
• 3. Hallucination (artificial intelligence), general reference summary (2026); check the cited originals.
• 4. Secondary summaries of court-database and legal-hallucination counts (updated 19 October 2026), and the Air Canada tribunal case (2024); check originals.
• 5. Vectara and Suprmind summaries on grounding and retrieval (2026); secondary.
• 6. Open ASR Leaderboard (19 September 2026), via MeetWave summary; check the leaderboard.
• 7. MeetMemo accuracy comparison (2026); vendor.
• 8. Gallup, 2026 workforce survey (fielded 26 January to 24 March 2026); check Gallup directly for the 28% weekly-use figure.
• 9. AI Governance Courses, GDPR and AI overview (2026); secondary.
• 10. Vendor guidance on GDPR and AI (2026); directional.
• 11. US Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (29 January 2025).
• 12. European Commission, FAQ on transparency obligations under Article 50 of the AI Act (applies from 2 August 2026; machine-readable marking for systems already on the market until 2 December 2026).
• 13. EU AI Act, Article 50 transparency rules; artificialintelligenceact.eu (undated); check EUR-Lex.
• 14. EU AI Act, Article 50 text (Regulation (EU) 2024/1689); artificialintelligenceact.eu; check EUR-Lex.