Artificial Intelligence in Practice

Today's generative tools predict likely text, images or code from patterns in large amounts of data. They do not know facts the way a person who checked a source does.

• Predictive AI estimates a number or label (for example, whether a message is spam). Generative AI makes new content from instructions.
• Strengths: drafting, summarising, rewording, brainstorming, sorting, translating a first draft, finding patterns in tables.
• Weaknesses: made-up facts and sources (hallucination), uneven maths, old information, unseen bias, no accountability.
• The same instruction can give different answers; results need checking every time.
• AI is a tool inside a process. The person using it is responsible for the result.

Activity. Test one task three times, compare answers and write what changed and what was wrong.

Standard to meet. Build a comparison table with at least three differences, and list the errors you found.

Good first tasks are frequent, low risk and easy to check. Poor first tasks are rare, high risk or hard to check.

• Score each candidate task from 1 to 5 on frequency, time cost, risk if wrong and ease of checking.
• Start with tasks scoring high on frequency and ease of checking and low on risk.
• Examples of good starts: first drafts of routine emails, meeting summaries from your own notes, tidying tables, FAQ drafts.
• Examples to avoid at first: decisions about people, medical or legal advice, anything with secret data.
• Write the human who will check and what they will look for.

Activity. Score ten tasks from your own work and pick one to build.

Standard to meet. Score a table of ten tasks, and name a checker for the chosen task.

A good instruction gives the tool what a new colleague would need.

• Use five parts: role (who you want it to act as), context (the situation), task (the exact job), format (length, layout, tone) and examples (one good sample).
• Put facts in the instruction instead of expecting the tool to know them.
• Ask for steps or a checklist when the task is complex, and ask it to say when it does not know.
• Change one part at a time and keep a log of which version works best.
• Keep a reusable template for each routine task.

Activity. Write three versions of the instruction for your chosen task and compare the outputs.

Standard to meet. Keep a version log that shows what changed and why the best version won, and save the final template.

What you type may be stored, used or seen by others. What the tool produces may resemble existing work.

• Classify data: public, internal, confidential, personal. Only public or made-up data goes into public tools.
• Replace real names and numbers with placeholders (Customer A, 1,000).
• Read your workplace rule and the tool's data policy; ask whether inputs are used to improve the tool.
• Copyright and licensing differ by place. Do not present copied or closely mimicked work as your own; check licences for any images, text or code you publish.
• Disclose AI help where required, and keep a record of what the tool contributed.

Activity. Write a one-page data rule for your task: what may and may not be entered, with five examples each.

Standard to meet. Write a rule that names the data classes, gives five allowed and five forbidden examples, and says who to ask.

Trust is earned by testing. Build a small test set and a checklist.

• Test set: 20 or more realistic inputs with the answers you would accept. Include tricky and unusual cases.
• Checklist: facts correct, numbers recomputed, sources real, tone right, nothing sensitive, nothing unfair to a group.
• Verify any claim that matters in an original source. If a source is cited, open it.
• Bias: ask who could be treated unfairly by the output and test with varied names, places and cases.
• Record the error rate: errors divided by test cases.

Activity. Run your 20 test cases, score them, improve the instruction and run them again.

Standard to meet. Compare two score tables, compute the error rate correctly, and explain the improvement.

Design who does what: the tool drafts, the person checks and decides.

• Map the steps: input, tool draft, human check, send or publish, record.
• Set stop rules: when to skip the tool and do it by hand.
• Log each use: date, task, tool, who checked, result.
• Keep the final decision and signature with a person.
• Write a short guide so a colleague can follow it.

Activity. Draw your workflow and run it on 20 real (masked) or sample cases.

Standard to meet. Diagram has a human check and a stop rule; the log shows all 20 cases.

Value is time saved, quality kept and risk avoided. Measure before and after.

• Baseline: time per task by hand, error rate, cost per hour.
• After: time with the tool including checking, error rate.
• Time saved per task x tasks per month x hourly cost = monthly value. Subtract tool cost.
• Include the learning time and any checking time honestly.
• Report quality too: customer satisfaction, rework, complaints.

Activity. Compute baseline, after, and net value for your workflow.

Standard to meet. Calculate the arithmetic correctly with units and currency, report a quality measure, and list the assumptions.

Use AI to give people better work, not to hide mistakes or replace judgement.

• Be open: tell customers and colleagues when AI helped, if required.
• Keep humans in charge of decisions about people, safety, health, law and money.
• Review regularly: tools change, rules change, and errors may appear later.
• Skills that grow in value: asking good questions, checking, domain knowledge, ethics.
• Jobs: AI-assisted operations, content and support assistant, data clean-up, workflow designer, trainer for colleagues.

Activity. Write a one-page team policy and present your whole project in 10 minutes.

Standard to meet. Write a policy with five rules, an owner and a review date, and present the results, limits and one thing you chose not to use AI for.

On the slide: L1 Basic · What AI tools do

Notes: Explain in plain words what an AI tool does and one thing it cannot be trusted to do. Notes: show one hallucination example; learners say what a person should check.

On the slide: L2 Awareness · Safe first tasks

Notes: Choose a low-risk task and know what data must never go into a public tool. Notes: sort ten tasks by risk and ease of checking; only public or made-up data is used.

On the slide: L3 Foundation · Write a clear instruction

Notes: Write an instruction with role, context, task, format and examples. Notes: compare a vague and a clear instruction on the same task and score both.

On the slide: L4 Practitioner · A checked AI workflow

Notes: Build a workflow with a human check and measure its error rate. Notes: practice case is 5 errors in 20 tests, 25 per cent, falling to 10 per cent after a rule is added; learners rerun it.

On the slide: L5 Specialist · Value, bias and limits

Notes: Measure time saved and test for bias and invented facts. Notes: 4 minutes saved on 150 tasks is 10 hours a month; learners present their own saving with the checking time included.

On the slide: L6 Advanced · Lead team rules

Notes: Write team rules for logging use, disclosure and who signs off. Notes: review three logged uses; decisions on hiring, health, legal matters, safety and large money stay with a person.

On the slide: L7 Super expert · Govern AI use across a group

Notes: Design a governance plan with risk levels, audit and training. Notes: check privacy and copyright steps; present a 12-month plan with named owners.