Frequently Asked Questions
Answers to the questions people ask most in our generative AI trainings, about Microsoft Copilot, ChatGPT, Claude and Gemini at work.
111 questions
How AI Works
What is the difference between a reasoning Model and a Normal AI Model?
A reasoning model thinks through a problem step by step before it answers, while a fast everyday model replies straight away with the first likely answer. Reasoning models do not know more, they spend longer thinking. That makes them slower but better at complex tasks such as analysis, planning and maths.
Should I always use the most Advanced AI Model?
No. The most advanced models are slower and use up your limit faster, so they are not worth it for simple tasks. A big reasoning model can take a minute where a fast model takes seconds, and on a paid plan a few heavy questions can put you on a cooldown until your usage resets.
Can I make an AI Chatbot reason step by step with a Prompt?
Yes, partly. If your tool gives you no choice of model, ask it to solve the problem one step at a time and to wait for you to say continue before each next step. It then puts its full attention on one step instead of trying to handle the whole problem at once.
Why does ChatGPT give Different Answers to the same Prompt?
Because generative AI is probabilistic, so its answers vary by design. Every time you run a prompt, the model works through the whole task again and picks likely words, which leads to slightly different results. Ordinary code is deterministic and gives the same output every time, but language models cannot work that way.
Can a Chatbot read all the sub-pages of a Website from one link?
No. When you give a chatbot a link, it usually reads only the information on that page and does not crawl the sub-pages behind it. Cached or hidden data on a page can also lead to wrong answers, such as an outdated price. The safest way is to copy the text you need and paste it in.
When should I use Code instead of Generative AI?
Use code when the data is structured and the rules are fixed, and use generative AI for unstructured language that never arrives in the same shape, such as emails, transcripts or opinions. Code is deterministic, so the same input always gives the same result. Generative AI is probabilistic and is never 100 per cent right.
Does Copilot understand dialects and idioms?
Yes, mostly. Copilot handles major languages and common idioms well, because they appear often in the text it learned from. Regional dialects and rare local expressions are less reliable, so for anything important, ask a native speaker to check the output before you use it.
Why does AI-generated text all sound the same?
Because with no context, AI falls back on the most average wording it knows, so everything comes out alike. Give it your own examples, your house style and a description of who the reader is. Then do a human edit before anything goes live, because that edit is what makes the text sound like you.
Has AI Search got worse?
No, the problem is mostly the internet itself. There is more AI-generated content and misinformation online than ever, which makes it harder for any search tool to tell good sources from bad ones. At citing its sources, AI search is actually getting better, and results improve when you tell it which sources to use.
Choosing an AI Tool
Which AI Tool is Best for Research, writing and answering questions?
It depends on the job. As a rule of thumb, Perplexity is strongest for research because it cites its sources, ChatGPT is strong for drafting text, and Gemini works well for questions and answers over long material. For anything involving work data, use the tool your organisation has approved.
What do you get with paid ChatGPT, Claude or Gemini that the Free plan does not give you?
It depends on the chatbot, but mostly you get higher usage limits and extra features. Free plans hit a usage limit and then make you wait for it to reset, and Claude's free plan tends to run out fastest. Paid plans also give a longer context window, so one conversation can run much longer.
Which AI Chatbots have a Projects feature?
Most of the big ones do. ChatGPT, Claude, Gemini, Mistral, Copilot and Grok all let you create projects, which keep related chats, files and instructions together in one place. You create each project yourself in the tool's projects section, then add the files and instructions you want it to use.
Is there a limit on how much text I can paste into an AI Chatbot?
Yes, and it depends on the tool. Gemini accepts very long text and is close to unlimited for most everyday use, while Claude has a much stricter limit. If you run out of space, put the text into a Word document and upload it as a file instead of pasting it into the chat.
Is Perplexity Good for Translation?
Yes, very much, especially with its own Sonar models. Perplexity is an aggregator, so on the Pro plan you can switch between models such as GPT, Claude and Gemini underneath it. Check which model is switched on before you judge the result, because the translation quality depends on that model.
Is Perplexity Finance any Good?
Yes. Perplexity Finance tracks markets, crypto and company financials, with charts, balance sheets and history, and it answers with sources. Open it at perplexity.ai/finance, or type a company name or ticker followed by the word stock, for example NVDA stock. On the mobile app, type finance in the search bar.
Is NotebookLM better than ChatGPT for Research?
Yes, for research on public documents. NotebookLM, now called Gemini Notebooks, stays within the sources you give it, while an ordinary chatbot tends to wander off into its general knowledge. It runs on Gemini, which has one of the largest context windows of the public AI tools, so it handles many long documents.
Can AI replace Human translators?
It depends on the type of text. For technical and formal text, most organisations now let AI translate and a human check the result. Artistic or creative translation is different, because the creative idea still has to come from a person. So AI speeds up routine translation but does not remove the human from it.
How can I check an AI Translation in a Language I do not speak?
Translate it back. Open a new chat, paste the translation and ask for it in English, then compare that with your original. Small synonym changes are normal. A difference in meaning is what you need to fix, or send to a native speaker to check before the text is used.
Can Copilot translate as I type?
Partly. You can write in one language and have Copilot work in another: say so once at the start of the chat, or put it in your custom instructions so it applies every time. Live translation word by word as you type is not something Copilot does, that is a job for a tool like Google Translate.
Which AI Model runs behind a No-Code Tool like N8N, and who pays for it?
Any model you choose, as long as the no-code tool has access to it. You connect it through an API key, which is separate from your monthly chatbot subscription. In N8N you add your own key, for example from the OpenAI platform with prepaid credits, and you pay per run, usually fractions of a cent.
Can you run an AI Model locally on your own computer?
Yes. You can download a model and run it on your own computer, cut off from the internet, so nothing you type leaves your machine. Most laptops cannot run the newest large models, so expect less reasoning power than the big online chatbots. LM Studio is a good starting point.
What computer do I need to run a Local AI Model?
Check your RAM, or your graphics card memory, first. The model has to sit comfortably under that amount, never right at the limit. A model that fits on your disk can still be too heavy to run. LM Studio downloads models for you, but it does not test whether your computer can handle them.
Why did Mistral remove Agents?
Because AI tools are moving from agents to modular building blocks. An agent is just instructions, knowledge and tools put together, and the chatbot already has all three. In Mistral you now combine a Skill, a library and a connector yourself, which is messier than picking one ready-made agent.
Why can I not find Gems in Gemini?
Because Gems have been deprecated and the feature is no longer available in Gemini. AI tools are moving away from separate custom bots towards skills, so the idea of saved instructions now lives under a different setup. There is no point looking for the Gems menu.
How do I share a Claude Artifact?
Only on a paid Claude plan, as the free plan no longer has publishing. If you have the feature, you will see a publish or share button on the artifact, which gives you a hosted link anyone can open. If you do not see the button, you cannot share it.
Can I use work Mode in the Free version of ChatGPT?
No. Work mode is only available on paid ChatGPT plans. On the free plan you can still chat, upload files and use many everyday features, but the longer, more independent tasks that work mode handles need a paid subscription.
Does AI work Mode keep running if I close the window?
No. Close the window and the task stops, then it carries on when you reopen it. Even complex work mode tasks usually run for about 20 to 25 minutes, not hours. For long or scheduled work, tools like Claude Code or Codex go further, but they also stop when your computer is off.
Does regenerating an Answer use up my ChatGPT or Claude limit?
Yes. A regenerated answer is a whole new answer, so it counts against your usage again. In ChatGPT it counts as another message, and in Claude the full answer is generated, and counted, again. Editing your original prompt is the more economical option, because the old exchange is replaced rather than stacked on top.
Prompting Tips
Can AI write a better Prompt for me?
Yes, and it is the preferred way to work. Ask the chatbot to write the prompt and give it as much information as possible: the scope, what to focus on, whether to search the web and whether to cite sources. AI explains a task to itself better than most people do. Check the prompt before you run it.
What is one-shot Prompting?
One-shot prompting means giving the chatbot one example of what you want. Zero-shot is no example and few-shot is several. The example shows the format, tone and length you expect, so the output matches it. It has nothing to do with shorter answers, the example sets whatever style you want.
How do I summarise a long report with AI?
Start with a content map. Ask the chatbot to map out the report first, listing its sections and what each one covers, before it summarises anything. It then has the whole structure organised and can do the heavy lifting section by section. It is optional, but it makes long reports much easier.
Can I ask AI to add something it missed to a summary?
Absolutely. Highlight or quote the part that is missing, then add a note or a question about it. The chatbot then refers to that exact passage and adds the point to the summary. Being specific about what was missed works far better than asking it to try again from scratch.
How do I make AI follow my Template or house style?
Upload the template or an example document and tell it to follow it. AI does not know your house style, so describe it in detail: fonts, colours, logo, section order, tone and length. An example document works best, either as an attached file or as pasted text.
Can one AI Chatbot play several roles or personas?
Yes. A chatbot can switch roles on cue, for example from editor to critic to customer. You are effectively building several characters, so describe each one clearly and say how to switch between them. You can also ask the chatbot to write the role-switching prompt for you.
How do I build an AI Agent for a Multi-Step Process?
Keep it as one agent so it holds the context across all the steps, and add the steps one at a time, testing each before you move on. Look into chain-of-thought prompting, which makes the model work through a task in order. With public data, Mistral is a good place to start.
Why should I tell AI my gender?
It is about grammar, not personality. In gendered languages, AI defaults to the masculine form when it writes as you. Tell it your gender once in your custom instructions and the grammar is right in every chat. Writing from a particular point of view in one document is a prompt, not a personal setting.
Can I tell AI to use only one Source?
Yes, and naming the source in your prompt is one of the cheapest quality improvements there is. It narrows where the AI looks, but it does not guarantee the figure is right, so check it yourself. If the source is your own document, do not just name it, attach it.
Does telling ChatGPT not to hallucinate work?
Partly. It lowers the odds but guarantees nothing, so you still need to check. What the instruction really means is: say you do not know rather than guess, so write exactly that. Ask for proof as well, such as citations or the exact passage each answer came from.
Should I give AI a word limit?
Yes, always say how long you want the answer. Give a number, such as 250 words, 500 words or one page, depending on where the text is going. Length belongs in the expectations part of your prompt, at the end, together with the format and tone you want.
What is a Prompt Library?
A prompt library is any place where you keep your reusable prompts: a Word document, an Excel sheet or a notebook page all work. Save a prompt once it gives you good results, then swap the details the next time you use it. That saves rewriting the same instructions again and again.
What should I write in Custom Instructions?
Describe how you work and communicate, and let the AI draft the instructions for you, because it explains things to itself better. In Copilot, a strong approach is to ask it to analyse your sent emails and write instructions that match your style. Then paste the result into the custom or draft instructions box.
What is the difference between Custom Instructions and Memory in AI?
Custom instructions hold your stable preferences, such as language, tone and role, and are sent with every message. Memory holds things that change over time. If the two contradict each other, the newest memory usually wins, but a memory is never created without you, so you stay in control.
Is there a limit on how many Memories AI can save?
It depends on the tool. In paid Copilot you are unlikely to hit a limit on memories, while tools like Claude and Mistral do cap them. Custom instructions are limited to about 8,000 characters because they are sent in full with every message. Memories are separate items, and only the relevant ones are pulled in.
Why can I not see a Memory I just saved?
Sometimes it is a sync delay and the list has not caught up yet, so close it and open it again. Sometimes it is a genuine glitch and the memory was never saved, so create it again. To test it, save a second, different memory and see whether that one appears.
Does Copilot remember my Previous chats?
Only if you set it up. You can turn on chat history in Personalisation, but it rarely pulls anything useful out of old chats, so do not rely on it. What reliably carries over is what you save yourself as a memory or in custom instructions. If you use none of these, each chat starts fresh.
Can I share my Custom Instructions with colleagues?
No, custom instructions stay with you and cannot be shared. What you can share is an agent, and only within your own organisation. When someone uses your agent, both sets of instructions apply: their own custom instructions and the instructions built into the agent.
Can I paste text into an AI Chat instead of uploading the file?
Yes, it is completely fine, but put the instruction first and the text after it. Pasting works perfectly for plain text such as a transcript or notes. For large files it is slower, and you lose images, charts and table structure, so uploading the file is better for those.
Accuracy and Checking
How do I check AI output quickly?
Ask the AI to quote the exact source text under every statement, then use Ctrl+F to find each quote in the original. That turns a full re-read into a quick spot check. You still read everything, because AI removes the drafting work, not the responsibility for checking it.
Does AI always give its Sources?
No. You get clickable sources when the AI searches the web, but when it answers from your document or from what it learned in training, it shows no source unless you ask. Ask for the exact snippet, word for word, then Ctrl+F it in the document to confirm it is really there.
Can AI update old Documents to reflect New Regulations?
Yes, but treat the result as a draft, never a finished version. Ask the AI for a word-for-word list of every change it made, then find each change with Ctrl+F and verify it yourself against the new regulation. Regulations are exactly where a confident but wrong edit costs the most.
How Accurate are AI meeting Summaries?
Not accurate enough to use unchecked. Accents, people talking over each other, poor microphones and connection drops all cause errors. An AI recap saves you writing the minutes, never proofreading them, and only someone who attended the meeting can check it properly. One attendee reviewing it is usually enough.
Should AI meeting minutes be saved as Word or PDF?
It depends, as each format has pros and cons. A Word file lets you open, check and change the minutes yourself afterwards, which matters because AI minutes always need a human review. A PDF can recreate the look of your template more closely, but it is much harder to edit.
Does Copilot use my own Data or outside Sources?
It prefers your own data and only goes outside when it has to. In Excel, for example, if it can work out a figure from your columns, it builds a formula. If the ingredients are missing, it searches for them. You can override this either way, for example by telling it to take the figure from the World Bank.
Why does my Copilot notebook show fewer references than I uploaded?
Because it is still loading. The notebook only counts the references it has finished indexing, so wait until everything has loaded and the count catches up. Separately, there is a limit to how much it searches in one go, so ask targeted questions and ask it to quote the sections it used.
Data Privacy and Safety
How do I stop ChatGPT from using my Data for Training?
In ChatGPT, go to Settings, then Data controls, and switch off Improve the model for everyone. The setting is backed by the provider's terms and conditions, not just a button. Many free AI tools have training switched on by default, so check the setting in every tool you use.
Does Microsoft Copilot learn from my Data?
Not with a work account. For business and enterprise users, training on your data is switched off by design, and Copilot only keeps the context of the current chat plus anything you deliberately save, such as custom instructions or memories. The consumer version of Copilot, used without a work licence, can learn from your data.
Are Enterprise AI Tools Safer than Free Chatbots?
Yes. Enterprise tools are built with security in mind. Free and personal versions of ChatGPT, Copilot, Claude and Mistral can train on your chats by default unless you opt out, while enterprise versions promise in their contracts not to train on your data. They also encrypt data in transit and can keep it in the EU.
Is Copilot on the Web Safe for work Data?
It depends on which account you sign in with. With a personal account, treat Copilot as a public third-party tool. With your work credentials, it runs inside your organisation's Microsoft 365 environment with enterprise data protection. Look for the green shield at the top, which shows enterprise data protection is switched on.
Is it Safe to connect ChatGPT to my Google account?
It is about as safe as the Google account itself. Connecting through an official connector does not change the security of your Google data, and both companies hold recognised security certifications. Perfect safety never exists, though, and a personal tool is never protected like a managed work environment, so for personal use the decision is yours.
Can I use AI to create Presentation slides?
Yes, AI presentation tools are fine for building a draft deck from public information. Let the tool create the structure and content from public material, then take the deck into your own tools and template and add any work-specific or confidential data there yourself. That keeps sensitive data out of public tools.
Microsoft Copilot
Can Copilot fill in a Word Template?
Yes. With a Microsoft 365 Copilot licence, open the template in Word, reference your source document in the Copilot side panel and ask it to populate the template with that data. Word can even grow a section to fit. With Copilot Chat alone or another chatbot it works, but it is clunkier.
Can Copilot use my company Template in PowerPoint?
Yes. Copilot in PowerPoint asks whether you want your organisation's style and builds the deck on your template. It is getting close, but layouts, logos and fonts are not always picked up perfectly, so check every slide before the deck goes out.
Can Copilot create a PowerPoint file?
Yes, Copilot creates an actual presentation file you can save, just as it does with Word documents. Asking from the chat window gives a weaker result than working inside PowerPoint itself, so for anything that matters, open PowerPoint and use Copilot there.
Can Copilot write speaking points from my Documents?
Yes. In Word, reference the files in the Copilot side panel and it drafts straight into the open document. It can combine several sources, for example four or five project documents into one combined report or one set of speaking points.
Can Copilot change the font in Word?
Yes, directly in Word. Ask in the Copilot pane inside Word, with the document open in edit mode, and it changes the formatting there. If you ask from Copilot Chat instead, it creates a new Word document, which is much clunkier for formatting changes.
Can Copilot write VBA macros in Excel?
Partly. Copilot cannot create or run macros inside your workbook, but it can write the VBA code for you in the chat. You then paste that code into the VBA editor in Excel yourself and run it from there.
How do I analyse a Large Excel Spreadsheet with AI?
First ask: could this be a formula? If so, ask the AI to write the formula, and it runs over 40,000 rows instantly. Only split the data into chunks when every row needs its own reasoning, roughly ten thousand rows at a time. If the cells stop visibly changing, it has frozen, so tell it to continue.
Can Copilot build a growing Excel Database from reports?
Yes, with agents in Excel. The workbook becomes the database, with one sheet holding all the rows and a sheet called .rules holding the instructions on which fields to pull and how to format them. Bring it new reports and it appends the data. Correct it in chat and the rules improve.
Can Copilot use Prompts saved in an Excel file?
Yes, with a little setup. Either write in your custom instructions that your prompt library lives in that Excel file, or type / in the chat and bring the sheet in directly. Without one of these, the request is too vague and Copilot may not find the file.
Why can I not see Work IQ in Copilot?
There are several possible reasons. Your organisation may have switched it off, you may be signed in with a personal account instead of your work one, or your Microsoft 365 Copilot licence may not be applied yet. Work IQ comes with the full licence, so check which licence your account shows.
What is the Copilot sidebar in Word, Excel and PowerPoint Online?
It is Copilot sitting beside your document in the online app. With a Microsoft 365 Copilot licence it has full access to the app and can work on the open document. With basic Copilot Chat you use it as a protected chat, but it cannot work inside the document.
How do I open Copilot in the Edge sidebar?
Open the Microsoft Edge browser and click the Copilot icon in the top right corner, just under the close button. The side panel opens and can read the page you are on. On private or work sites it asks for permission first, and you can say no.
How do I get Copilot in Outlook?
Use the new Outlook, because classic Outlook does not have Copilot. The browser version is always the new Outlook, so open it there if you are unsure. Your licence then decides how much you get. A Copilot panel that looks different is usually an older build, since Microsoft rolls out updates to organisations at different times.
Where are Copilot draft Instructions in Outlook?
In the new Outlook, open Settings, scroll to Copilot at the bottom and choose draft instructions. They are not in classic Outlook, so open Outlook in the browser if you cannot find them. If you only see Preferences, Outlook still uses your Copilot custom instructions, and draft instructions are an extra layer on top.
Can Copilot auto-accept meeting invitations?
Yes, for invitations from a specific sender. The rule has to name an individual email address, not a team or group. Create it in the Copilot side panel in Outlook rather than on the settings page, then check it was saved under view instructions in the Copilot settings.
Can Copilot create calendar events?
Yes, any Copilot chat can create calendar events, not only the one inside Outlook, because the feature is tied to your profile rather than the window you are in. At the time of writing it reliably creates only one event at a time, so ask for several events separately.
Can Copilot make a priority list from my Emails and Teams messages?
Yes. With Work IQ, one prompt reaches Outlook and Teams together, so Copilot can build a single priority list from both. Commitments made only in a Teams message still surface as open items, so anything you agreed to in a chat will come back to find you.
Can Copilot summarise a Teams channel?
Yes, but not from inside the channel. There is no Copilot panel within a team or channel, but the main Copilot chat can reach your channels, so ask it there to look at a specific channel. This uses your work data, so it needs Work IQ and the full Microsoft 365 Copilot licence.
Can Copilot transcribe a recording?
No. Turning an existing video or audio file into text is a separate capability that some tools have, but Copilot does not. For Teams meetings, transcription has to be switched on at the time of the meeting. For an existing recording, use a dedicated transcription tool instead.
Can I use Copilot in Power BI?
Only with the right licence. Copilot in Power BI needs its own licensing, on top of a standard Microsoft 365 Copilot licence, so most organisations do not have it by default. Ask your IT team whether it is included in your setup before you plan around it.
Can I schedule Prompts in Copilot?
Yes, with the full Microsoft 365 Copilot licence. Scheduled prompts are not on basic Copilot Chat. Each scheduled prompt is tied to the chat it was created in. If you delete that chat, the prompt goes inactive, and it cannot be saved in the Library, so pin the chat to protect it.
Do Copilot tasks show up in Microsoft To Do?
No. Copilot tasks live inside Copilot only, not in To Do or your Microsoft task list. The feature is narrower than its name suggests and is not a task manager. It works well for recurring jobs such as a news digest, an email and task overview, or a weekly spreadsheet check.
What does a Copilot Connector do?
A connector links Copilot to an external service, such as Gmail, so your chat or an agent can pull data out of it or push data in. Think of connectors as tools. You find them under Settings, then Sources, and set each one up once by signing in.
Do Free AI plans have Connectors?
Yes, connectors are now available almost everywhere, free plans included. Before you rely on one, open the connector's list of functions, because it can only do the actions it was built for. Connectors change without notice, so test them live before you build anything on them.
Why can I not attach a file in Copilot?
Usually because of where the file came from. Files shared from another organisation often refuse to attach. Save the file to your own OneDrive and attach it through cloud files instead. Or copy and paste the text into the chat, with your instruction first and the data after it.
Can Copilot read files on my OneDrive or network drive?
Usually yes for OneDrive and SharePoint, with Work IQ on a full Microsoft 365 Copilot licence, but always test it, because it varies between organisations. Pick a document and ask Copilot whether it can see it and its contents. A network drive is not the same as OneDrive, so test that separately.
Do the Settings in the Copilot Chat menu apply only to that Chat?
No. Settings in the three dots at the top of Copilot Chat open the same global settings as the menu at the bottom of the screen. Any change you make there applies to all your chats. The three dots only show this option on a new, empty chat.
Can I tell Copilot how to name my chats and Pages?
Chats yes, Pages no. Put a naming rule in your custom instructions or save it as a memory, and new chats then follow it in your chat history. Pages are different: Copilot names them in a premade way before it reads your instructions, so rename each page yourself into whatever structure you want.
What is the Library in Copilot?
The Library, in Copilot's left-hand bar, is where Copilot stores what it creates: your Pages, canvases, generated images and other creations. It is storage, not a working feature, and it does not run prompts. Pages are saved there automatically, and a page survives even if you delete the chat it came from.
Can my colleagues or IT see my Microsoft Copilot chats?
Your colleagues cannot see your Copilot chats. Microsoft 365 Copilot only works inside your own permissions, so nobody can ask it about your mail or files. Your organisation can still keep and search Copilot conversations for compliance and audit through Microsoft Purview, the same way it can with email, so treat work chats as work records.
Does Microsoft Copilot warn me if I paste Personal Data?
Not reliably, so do not depend on it. Some organisations set up data loss prevention policies that can block or flag sensitive information, but there is no general alert that catches every name or personal detail you paste in. Checking what you share with Copilot is always your own responsibility.
Can I refuse to be recorded or transcribed in a Teams meeting?
Only by leaving the meeting. When the organiser starts recording or transcription, Teams shows a consent message. You can accept and stay, or decline and be disconnected, but you cannot opt out while staying in the call. The same applies to people from inside and outside your organisation.
Can I make Copilot forget Sensitive Information?
Not cleanly. You can delete chats and saved memories, but you cannot make Copilot forget one item from a meeting transcript or a document it has already processed. The safest approach is prevention: do not start transcription or share files when sensitive information could come up, and be open about it if it happens.
What is a Microsoft 365 tenant?
A Microsoft 365 tenant is your organisation's own private space in the Microsoft cloud. Your email, Teams, files and Microsoft 365 Copilot all run inside it, and your data stays in that space instead of being pooled with other customers. That is why the work version of Copilot is safer than a consumer chatbot.
Is Copilot Chat-Only Mode Safer for Sensitive Documents?
No, the mode makes no difference. Whether you use chat only or agent mode, Copilot reads the document once you give it access. The real decision is whether that document should go into Copilot at all, so check your organisation's rules on which data classifications are allowed before you open it with AI.
Notebooks, Agents and Automation
What is the difference between a Copilot notebook and an Agent?
Very little: notebooks and agents are very similar, and both can still reach everything your Copilot can. The real question is how you want to work. A notebook keeps all its chats and sources together in one place, while agent chats scatter through your chat history.
How do I create a notebook in Copilot?
In Copilot Chat, open Notebooks in the left-hand bar and choose to create a new one. Give it any name and press create. You can skip adding references and start with an empty notebook, then add files, links and instructions later.
Can I change a Copilot notebook's Instructions?
Yes. Open the notebook and rewrite its instructions yourself whenever you want. You do not need a new notebook, and everything else stays as it is. Save the old instructions as a Page first, so you can go back if needed.
Can I add Website links to a Copilot notebook?
Yes, notebook references now accept links to websites as well as files. That is useful for anything online that changes, such as academic or legal material, because the notebook works from the current published version rather than an old download.
Can a Copilot notebook reference a Recurring meeting?
No, a reference points at one specific past meeting, even in a recurring series. The workaround is to write in the notebook's custom instructions that it should summarise that meeting since last time. Then you just open the notebook and ask it to run.
How do I share a Copilot notebook or Agent?
Both share in one click. A notebook has a share button at the top, and an agent shares from its three dots menu. You can only share with people inside your own organisation, not with anyone outside it.
How can I get AI to work with lots of PDFs at once?
Stop uploading and start pointing. A chat caps uploads because it recalculates everything on every message, while a Copilot notebook processes documents once, so it holds far more. In testing, a notebook took 52 PDFs without complaint. You can also build an agent that reads a OneDrive folder.
Can a Copilot Page work only from a set of Documents?
No, Pages are for working on a document with Copilot, not for holding sources. To ground Copilot in a fixed set of documents, build an agent, which you can even create from a page, or use a notebook, which is the hub for all the sources of one piece of work.
Can AI connect two Apps automatically?
Yes, but you often do not need generative AI for it. For personal apps, no-code tools like Zapier, Make or N8N connect two apps without code, and at work you use what your organisation approves, usually Power Automate. Check first for a native integration, because many apps already sync with each other.
Can I automate payment reminders with AI?
Yes, but do it without generative AI. Use an automation tool such as Power Automate connected to your email with if-then logic, so there is no risk of hallucinations. Map the whole process on paper first, then automate only the parts that are truly repetitive.
Can I build an Always-Listening AI Assistant?
Yes, but it is a build project, not a setting. You need something always running, such as a mini PC or cloud server, that records continuously and keeps only a rolling buffer, for example the last five minutes. A wake phrase then triggers a speech-to-text model like Whisper to transcribe that buffer.
Can Copilot automate searching for Data and emailing people for it?
Partly, and it is a good candidate for an agent. The task splits into a search step and an email step. Until you can build an agent, a notebook with the steps written as its instructions gets you close, or save the prompt in your prompt library and swap the details each time. Always keep a human check.
Ethics, Copyright and the EU AI Act
Can I use AI-generated Images for work?
Yes, for internal use: testing ideas, drafts and internal material are fine. Do not use AI images in anything public-facing or customer-facing, because you cannot be sure an image is free of copyright or other rights. The tool cannot guarantee where the elements in an image came from.
How are Deepfakes made?
It is not a difficult process, which is what makes deepfakes worrying. Generative AI needs enough photos of a person to recreate their face convincingly, and for a video where they speak, recordings of their voice as well. Many tools refuse photos of real people, but often allow it if the user claims to be that person.
Can AI choose the Best job candidate from CVs?
No. AI can filter out unqualified applicants, but it should never make the final pick. Beyond your stated criteria, it invents its own from its training data, and that is where bias shows, for example filtering out women or people from poorer areas. If you have full criteria, a form or code applies them correctly every time.
What Rules apply to using Generative AI at work in the EU?
The main framework is the EU AI Act, with GDPR covering any personal data you put into AI tools. Many organisations add their own rules on top, especially about which tools are approved for work data. The headline principles are human oversight, treating every output as a draft, and transparency about AI use.
Nothing matches that.