TRAINING
Practise with workplace tasks
Build skills in giving instructions, using source material and reviewing results. Give staff a shared way to judge AI-generated work.
AI AT WORK · A PRACTICAL INTRODUCTION
From rough notes
to a useful draft.
What can generative AI do for your team—and what still needs a person’s judgement?
Generative AI produces material such as text, images and audio using patterns learned from examples. Here we focus on text, following one everyday task: turning workshop notes into an email.
A teaching simulation. These are prepared examples, not answers from a live AI service. The interactions make a few ideas visible; they do not reproduce a full AI model.
You have the facts. You need an email.
AI can organise rough information into a draft. It still needs the right instructions and someone to check the result.
THE MATERIAL YOU PROVIDE
INTERNAL WORKSHOP NOTE · FICTIONAL EXAMPLE
- Team workshop: 18 October
- Start: 10:00
- Venue: Room B
- Bring a laptop
- Reply to the organiser by 12 October
Your instruction (the “prompt”)Turn these notes into a short email to colleagues. Use a friendly, professional tone. Include only the supplied facts. Do not invent missing details.
What else could I ask it to do?
For this task, you could ask for a shorter version, a subject line or a more formal tone. These are changes to how the supplied information is presented. They do not supply new facts.
THE POSSIBLE OUTPUT
Choose “Show an example draft” to see how the notes can become an email.
The date, time, room, laptop instruction and reply deadline come from the note. The greeting and sentence structure are added to make an email.
The value: a starting draft you can review and adapt.
Formatting and phrasing are useful work. Whether the email is accurate and suitable remains your decision.
Step 1 of 4
An answer is built a piece at a time.
Before you use a language model, training adjusts its internal numerical settings using many examples of text. When you give it a prompt, it uses those learned patterns and the available text to calculate possible next pieces.
A SMALL PART OF THE WORKSHOP EMAIL
Here we slow down a sentence. Choose an opening and generate one piece at a time.
Please
Start with the choices shown on the right.
What is a token?
A token is a unit of text a model processes. It may be a word, part of a word, a space or punctuation. This demo treats whole words and punctuation as tokens for readability.
For example, a tokenizer might split “workshop” into workshop. This is an illustration; actual token boundaries depend on the model.
ILLUSTRATIVE NEXT-TOKEN CHOICES
These percentages are chances of choosing a token. They are not measures of whether a statement is true.
Watch the choices change after each selected word.
What this simulation leaves out
The choices here come from a small, hand-written branching model. A real language model calculates scores using learned numerical settings and the text available to it. It does not use this fixed sentence tree. The training process is described here, not simulated.
The text a model can use at once is limited by its context window. Long documents may need to be handled in sections, with relevant information carried forward.
The words already present affect what comes next.
Compare “Please join…” with “Please bring…”. The possible continuations change. In a real model, this happens across a much larger set of possible tokens.
There can be several plausible continuations.
The system does not always choose the most likely token. Some tools offer a setting called temperature that changes how strongly the system favours likely choices. It controls variation, not truth or intelligence.
THE SAME OPENING EACH TIME
Please…Which word will be chosen next?
More predictableMore varied
At zero, this demo always chooses the highest-scoring option.
Try zero, then move the slider to 1.5 and try again. A more varied setting can still repeat the same choice.
SIX INDEPENDENT ATTEMPTS
Choose “Try six choices”. Every attempt starts again at “Please”.
Varied wording is not evidence of better accuracy.
This is one token from a short sentence, not a comparison of six complete emails. The chances are calculated from the same hand-written scores as the previous step.
What should I do in a real AI tool?
If a temperature control is available, a lower setting can suit consistent factual drafting; a higher setting can suit exploring ideas. Neither removes the need to check the result. Many chat interfaces do not expose this setting.
Different wording can be useful. Changing facts is a problem.
Acceptable variations in an email might change the greeting or phrasing. They should not quietly change the workshop’s date, room or reply deadline.
A convincing sentence still needs evidence.
Your colleague asks, “What time does the workshop finish?” The supplied note gives a start time, but no finish time. Decide what the answer should say.
CHECK THE SOURCE
THE SAME INTERNAL WORKSHOP NOTE
- 18 October at 10:00 in Room B
- Bring a laptop
- Reply by 12 October
No finishing time is provided.
Which response is supported?
Why might AI invent a detail?
A plausible continuation can include details that are not supported by the source. People often call this a hallucination. It is a possible failure, not an inevitable response: a model can also say it does not know.
Giving sources makes checking possible, but does not guarantee that the model will read or use them correctly. A retrieval system can bring relevant documents into the prompt; people still need to check the resulting claims.
NOW REVIEW A DRAFT
Hi team,
Please join our workshop on 18 October at 10:00 in Room B. Bring a laptop and reply by 12 October. The workshop will finish at 12:00.
Thank you.
Select every detail supported by the note:
Use a clear instruction, approved information and human review.
Specify the task, audience and output. Ask the model to flag missing information. Check its claims against the source and confirm that you are allowed to share the material with the tool.
PUT THIS INTO PRACTICE
Bring us a task your team needs to tackle.
You have seen how AI can turn notes into a draft, why wording can vary and where a person’s judgement matters. Coutts Media offers training and advice for applying those ideas to your organisation’s work.
ADVISORY
Agree how AI should be used
Work through suitable tasks, information handling and human approval. Make responsibilities clear before adopting a new workflow.
CONTENT WORKFLOWS
Improve a real process
Examine a writing, editing or communication task. Decide where AI contributes and what good output and human review should look like.
Discuss AI training for your team
Tell Lloyd what your team does and where you would like to use AI. Start with one task or concern.
Prefer webmail? Write to [email protected]. The button opens your email app; you review and send the message yourself.