Intro to AI · Part 3 of 3
What Is Possible
After this part you can
Tell what AI is worth trying on, what it cannot do, and where to go next in the course.
About this part
This is the last of three parts in the intro to AI module, which is optional and written for readers without much experience of AI. The aims of the course, the layout every part follows and suggested reading routes are in Course introduction. Reading time is about 15 minutes.
Part 1 said what AI is and part 2 how to use it safely. This part is a tour of what it can do in 2026, and what it still cannot, so that you know what is worth trying. It ends by pointing you to the right place in the rest of the course.
What part 3 gives you
Part 3 builds one idea: AI is strongest where the work is made of language and the result can be checked, and weakest where the work needs facts it does not have, judgement about your situation, or someone to be accountable. Most useful work is a mixture, which is why the best results come from a person and an AI working together, each doing the part they are better at.
1. What it does well
In plain terms
It is very good at anything made of words: writing, rewriting, summarising, explaining, translating, and answering questions about a document you give it. It is good at computer code, which is a kind of language with the advantage that you can run it to see if it works. And it is a tireless thinking partner.
Who should read it: everyone.
Working with text
- Drafting. Emails, reports, job descriptions, documentation, meeting agendas. It gives you something to react to, which is easier than a blank page.
- Rewriting. Shorter, plainer, more formal, friendlier, for a different audience. This is among its most reliable uses, because everything it needs is in front of it.
- Summarising. A long document, an email thread, a transcript. Give it the text and say who the summary is for.
- Explaining. A contract clause, an error message, a statistical term, a piece of someone else’s code, at whatever level you ask for. You can keep asking “why?” without embarrassment.
- Translating, between languages and between registers: legal to plain, technical to commercial.
- Pulling things out. The action points from the minutes, the dates from the contract, the complaints from two hundred reviews, sorted by theme.
Working with ideas
- Thinking something through. Describe a decision and ask for the considerations you have missed, the strongest argument against, or the questions a sceptical director would ask.
- Generating options. Twenty names, ten approaches, five ways to structure a talk. Most will be ordinary. Two may be good, and it took a minute.
- Preparing. Practise an interview, a negotiation or a difficult conversation, with the AI playing the other side.
- Learning. A patient tutor in almost any subject, which will explain it again a different way as often as you need. Check the facts that matter.
Working with code and data
Code is where AI has changed work most, and not only for programmers. It will write a spreadsheet formula, a script to rename five hundred files, or a query against a database, and explain what each line does. Many assistants can also run code themselves, which means they can analyse a spreadsheet you upload, draw the chart, and show their working. For software teams, coding agents now take on whole tasks, and part 5 of the practical AI module covers them.
2. From answering to doing
In plain terms
The first AI assistants could only talk. Current ones can be given tools: they can search the web, read your files, run programs, and work inside other software. An AI with tools and a goal, working through a task step by step, is called an agent. This is the biggest change of the last two years, and it is why the course spends so long on it.
Who should read it: everyone.
What tools add
| With this tool | It can |
|---|---|
| Web search | Answer from current sources and show you where each fact came from |
| File access | Read your documents, code or data, and work from what is actually there |
| Running code | Do arithmetic and analysis properly, and test the code it writes |
| Connections to other systems | Look up a ticket, check a calendar, query a database, draft a reply in your support tool |
| Control of a screen | Operate software that was never built to be automated, slowly |
Tools fix some of the weaknesses from part 1. An assistant that searches is no longer limited to what it remembered. One that runs code no longer guesses at sums.
They also raise the stakes. An assistant that can only talk can at worst say something wrong. An agent that can send, delete, buy or deploy can do something wrong. That is why part 7 and part 9 of the practical AI module are about deciding what an AI may reach and what it must ask permission for.
What agents are doing now
In software teams, agents fix bugs, write tests, review each other’s changes and carry out long, dull migrations, with a person setting the task and judging the result. Outside software, the same pattern is appearing in research, where an agent reads fifty sources and writes up what it found with references, and in support and operations, where an agent gathers the facts of a case before a person decides. The AI in the organisation module looks at where this pays off across a whole team.
3. Beyond text
In plain terms
The same family of technology can look at pictures, listen to recordings and watch video, and separate tools can create images, video, speech, music and 3D models from a description.
Who should read it: everyone. It is short.
- Taking media in. You can show an assistant a photograph of a whiteboard, a scanned invoice, a chart or a screenshot of an error, and ask about it. Some will transcribe and summarise a meeting recording. Part 6 of the practical AI module covers what they see well and what they miss.
- Making media. Image generators produce illustrations, product mock-ups and photographs of things that never existed. Video, voice, music and 3D tools are less mature and improving fast. The generative media module explains how they work, how to ask them for what you want, and the questions of rights and honesty they raise.
The caution from part 1 carries over. A generated image can look entirely real and be entirely invented, and the same is now true of voices and video.
4. What it cannot do
In plain terms
It cannot know what it was never told. It cannot be relied on for facts without a source, or for long tasks without checks. It cannot judge your situation, and it cannot be held responsible for anything. Those gaps are where you come in.
Who should read it: everyone.
| It struggles with | Because | So |
|---|---|---|
| Facts about your organisation, your customers, your code | It was never shown them | Give it the material, or connect it to where the material lives |
| Anything recent | Its knowledge stops at a date | Use an assistant that searches, and ask for sources |
| Exact facts, figures and references from memory | It writes what is plausible | Verify what matters |
| Long tasks with many steps | Small errors add up along the way | Break the work up, and check between steps |
| Knowing when it is wrong | It has no sense of its own uncertainty | You check. It will not warn you |
| Judgement about people and priorities | It knows nothing of the history, the politics or what is at stake for you | Use it to lay out the options. Make the call yourself |
| Accountability | It is software | A person owns every decision and every piece of work that goes out |
It also reflects what it was trained on. It can repeat common assumptions and biases as if they were neutral, and it leans towards the conventional answer. For anything that affects people, such as hiring, assessment or lending, that matters a great deal, and the law in many places says so.
What this adds up to
A useful way to think about any task is to ask two questions. Is the work mostly made of language, code or images? And can the result be checked, by reading it, running it or comparing it with a source? Where both answers are yes, AI will probably help a lot. Where the work depends on knowledge it does not have, or where nobody could tell a good result from a bad one, it will help less, and may mislead.
5. Where to go next
In plain terms
You now know enough to start the course proper. It begins with how to use AI well, then explains how it works, then explores further.
Who should read it: everyone who read this far.
| If you want to | Go to |
|---|---|
| Get better results from the AI tools you use, starting tomorrow | Practical AI, from part 1 |
| See where AI pays off across a team and an organisation, and make the case for it | AI in the organisation |
| Understand what is happening inside | Language models |
| Run a model on your own computer | Running AI locally |
| Make images, video, music or 3D | Generative media |
| See the kinds of AI that are not chat assistants | Other AI models |
The Course introduction has reading routes for different roles, including one that takes an hour.
The whiteboard version
AI is strongest where the work is made of words or code and the result can be checked. With tools it can search, read, calculate and act, which makes it far more useful and makes limits on it matter. It can read and make pictures, sound and video. It cannot know what it was not told, vouch for its own facts, judge your situation or answer for the outcome. Those are your part.
Say it two ways
| Idea | Technical version | Non-technical version |
|---|---|---|
| Tool use | The model requests an action, such as a search or a calculation, and the application carries it out and returns the result | It can ask other software to do things for it, and work with what comes back |
| Agent | A model with tools, run in a loop towards a goal, checking results as it goes | An AI that is given a job, not a question, and works through it step by step |
| Multi-modal | A model that accepts images, audio or video as well as text | You can show it things and play it things, not only type |
| Generative media | Models that produce images, video, audio or 3D from a description | Tools that make pictures, clips and sound from what you ask for |
| Checkable work | Tasks whose output can be verified automatically or quickly by a person | If we can tell a good result from a bad one, AI is safe to lean on. If we cannot, it is not |
Misconceptions to correct
“It can do anything now”
- True
- The range is startling, and it widens every few months.
- Misleading
- It is capable and unreliable at once. It will attempt almost anything and tell you it succeeded. What it can be trusted with is narrower than what it will try, and depends on whether the result can be checked.
- What to say
- “It will have a go at anything. We use it where we can check the result, and we check.”
“It is only a toy for writing emails”
- True
- Drafting and rewriting are where most people start, and they are modest gains.
- Misleading
- With tools and good instructions, the same technology carries out multi-step work: analysing data, researching with sources, changing software and testing it. Teams that stop at email are using a fraction of it.
- What to say
- “Email is the shallow end. The larger gains are in work it can carry out and we can check, and that is what the rest of this course is about.”
“It will replace the people who do this work”
- True
- It changes what the work consists of, sometimes a great deal, and some tasks disappear.
- Misleading
- It has no knowledge of your situation, no judgement about what matters and no accountability. What it does is shift people’s effort towards deciding what should be done and judging whether it was done well.
- What to say
- “It takes over the producing. The deciding and the answering for it stay with us, and those become more of the job.”
Glossary
Terms introduced in this part, in plain language and in alphabetical order. Earlier terms are defined in parts 1 and 2.
- Bias
- A lean in the model’s output that reflects patterns in what it was trained on
- Coding agent
- An agent with a programmer’s tools: it reads and changes files and runs tests
- Generated media
- Images, video, sound or 3D made by a model from a description
- Multi-modal
- Able to take in more than text: images, sound or video
- Tool
- Something an assistant can ask other software to do for it, such as a search or a calculation
- Verification
- Checking a result against something independent: a source, a test, a second method