My article about M365 Copilot from a couple of weeks ago keeps being forwarded and read. Every time I check the stats, it’s the one article consistently growing in popularity.

I must say this is one of the biggest surprises here so far. Copilot? Really??

But then again, it works great for me as a Substack writer.

After all, it’s the only AI tool I am allowed to use at work! And because all my work sits in Microsoft ecosystem, it’s natural that Copilot has become my true partner in crime. I honestly grew to like it and I am delighted, daily - which I know might be a shocker to a lot of people.

It might be a good moment to re-introduce myself: My name is Justyna and I am a Client Service Leader (in some companies known as Customer Success Manager or Client Delivery Manager).

My company provides payroll & HR solutions. My clients are major enterprise players. We all use - primarily - Microsoft tools.

What I actually do most days: read what a client asked for, work out what they meant, and turn it into something operations can act on. Then do it backwards: take a technical answer from a specialist and explain it to stakeholders who need to decide by Thursday. Somewhere in between, I collect complaints and try to spot which ones are the same complaint wearing different clothes — and come up with tangible improvements.

I have nagged have showed enough interest in AI adoption & enablement to have been involved in the Federated Developers program. I was given the Copilot Premium license a while back and it’s been a game-changer for me. And no, it’s not about writing smooth e-mails — and I hope most of my readers are at a different level now.

So in this article I want to cover three things, really.

1/ what is the difference between Copilot Chat vs. M365 Copilot vs. M365 Copilot Premium? (btw, forget Copilot Pro - that doesn’t exist anymore as a standalone solution).

2/ what can you get done with M365 Copilot Basic version?

3/ when & how to ask for an upgrade to Premium.

What this article won’t cover is Copilot Studio - I’ll put together a separate post about it because it deserves it.

A couple of caveats before we start, just like in the previous Copilot article:

1. It’s September 2026 and this article has an expiry date on it - Microsoft is changing things fast.

2. What you can see and do in Copilot will depend on your license AND on your company settings. Your IT can disable some options so my Premium plan might be different to yours. If you aren’t a Premium user, the options are limited. It’s always worth asking for an upgrade! I did, and was given one.

Microsoft has confused the hell out of people with naming their every other AI product “Copilot-something”.

Apparently, there are now 78 products with Copilot int the name. Consistent? - sure! Helpful - not so much.

Let’s start with the three most common ones.

1. Copilot Chat vs. M365 Copilot vs. Premium

The easiest way to hold this in your head: one knows the internet, one knows your company, one sits inside your apps and works alongside you + creates agents.

Copilot Chat is the one at copilot.microsoft.com that you can use without logging in with your work account. It knows the web. It knows nothing about you, your clients or your files. This is the version most people tried once, shrugged at, and quickly went to ChatGPT or Claude.

M365 Copilot (Basic) is what you get automatically with a work Microsoft 365 licence. No add-on, no extra cost, and most people have it without realising. The big difference is the boundary: you’re signed in with your work account, so what you type stays inside your company’s tenant instead of wandering off. You can hand it a file and it will actually work on it. What it won’t do is go looking through your mailbox, your Teams chats or SharePoint on its own.

M365 Copilot Premium is the paid add-on your company buys per person. It appears inside Word, Excel, PowerPoint, Sharepoint, Outlook and Teams, suggests changes, take my free text command and act on it. It can search across everything I already have access to. That’s the part that turned it into a partner in crime for me: I can ask “what did we agree with this client about the reporting cycle back in spring?” and it goes, looks through e-mails, chats, transcripts and finds the thread(s). I cannot tell you what a blessing this has been.

And then there are the agents you can use and create for yourself. More info on that in my previous post.

Want to know which one you have? Go to m365.cloud.microsoft and look at the label under your name.

2. What you can actually do with Basic

Quite a lot, honestly - although I would cry ugly tears if I were to go back to Basic.

Tell it who you are, once. Copilot has a memory. Use it. Instead of re-explaining your job every Monday morning, sit down and tell it properly: "I manage a portfolio of enterprise payroll clients. Here are my three biggest accounts, my weekly reporting rhythm, and the two escalation processes I follow. Remember this." Everything you ask afterwards lands in context instead of starting from zero.

Give it the file instead of hoping it finds one. Upload the document, then ask. “This is a 40-page product presentation. Prepare an executive-level summary.” Basic won’t go hunting through SharePoint, but hand it the thing directly and it works properly.

Use it as your thinking partner before a difficult conversation. Paste in the escalation email you received. “I need to respond to this. Ask me 3 questions about the context before you draft anything.”

Translate, summarise, restructure. Boring, and it’s still where most of my saved time comes from.

Build your own memory in Loop. Bonus tip! Copilot generates something useful, so save it into a Loop page. Keep adding your own notes, actions, decisions. Then next time you need an updated version of that status report, you upload the Loop page and say "regenerate this with the current information." You've effectively given yourself the continuity that Premium gets for free, by hand. It's not elegant - but it works.

What Basic won’t do: no sidebar in Word or Excel, no meeting recap in Teams, no searching your own mailbox. If someone demos those to you and you can’t find them, you’re not doing it wrong. You just have a different licence.

3. When and how to ask for Premium

Do not ask for a licence. Ask for a task to be solved.

Spend two weeks writing down what you do by hand. Actual minutes, actual tasks. “Reading back through six months of Teams threads to reconstruct what we promised and therefore saving x amount of time” is a business case. “I’d like to try Copilot” is not.

Pick a use case Basic genuinely can’t cover. This matters more than anything. If your ask can be answered with “just use the free one”, it will be. Cross-mailbox retrieval, meeting recaps and in-app work are the arguments that survive.

Write the email so your manager can forward it. Executives everywhere are under pressure to look AI-forward and quietly have no idea what that means. Someone with a concrete use case and a number attached is a gift to them.

Ask for a pilot, not a permanent seat. Thirty days, one team, one measured outcome. Much easier to approve. Also check whether your company already has unassigned licences sitting there, because it often does. In the corporate world, people often as enthusiastic about change and novelty.

Thank you for reading this far. I want my Substack to be as helpful and actionable and relatable as possible. I know there’s a lot of us out there - busy, corporate folks with households to run - and I hope AI can give you back a bit of what those days eat first: a clear head, an hour that's actually yours, and the feeling that you're on top of things rather than underneath them.

~ Move forward, keep it steady. Justyna

psst..read till the end, a useful prompt to copy/paste added!

Yesterday, I submitted a use case at work.

It proposes an AI agent that helps our client leaders read, do a search and interpret client contracts. I think it’s a good one. It says who has the problem, how much time they lose to it now, and what we’d get back if it works.

I want to build this agent - I think it can help a lot more people than just me. Which is exactly why I wrote the case ready to answer every doubt that could potentially question its relevance.

A month or two earlier I’d started a different case, for a tool to prioritise my inbox. I submitted it for review, but a couple of weeks later, when Copilot updated, what I needed was suddenly just there in Outlook. No agent needed, no case needed - so I dropped it.

Looking back, it was the process working exactly as expected: the “boring” tool caught up, and I kept my time for a problem that was actually still a problem.

And I get it. There is a lot of pressure right now to be the person who “brought AI in.” Your feed is full of people who automated something and made a post about it. So you spot a problem, and the instinct is to yell “Here! I found it! Now give us AI to fix it!”

But your manager doesn’t fund tools. Your company isn’t a charity. And sometimes, as you might recall from my previous articles, AI is not necessarily the answer.

So this is really an article about how to build a case that’s honest enough to survive contact with a sceptical manager, because a sceptical manager is exactly who you’re pitching to.

What a manager is really asking

You might have done a version of this before. This is a summary of the most bullet-proof elements of a good use case proposal:

What problem is this, and what does it cost us right now?

Open with the problem and a number. Not “I’d like to propose an agent,” but “our client leaders spend real hours a week going through pages and pages of complex contracts, and it’s the bottleneck before we can get back to a client.” If you can’t put a rough figure on the current pain, you don’t have a case yet.

Who actually has this problem?

Size it by the task, not the headcount. Again, my own example: the contract agent doesn’t scale with how big the company is. It scales with how many people read contracts. That might be forty people in a division of thousands, so forty is my number. This step might be harder than you think in big companies and might require you to reach out for help to get the right figure.

How do you know it isn’t just you?

Now, the group might be 40 people, but how do I know they will actually find it useful? This is where I did something very low-effort that worked well. I asked a question on our team channel of about 100 people and watched who replied. A dozen or so said “yes, I have this too,” another dozen liked the message, and then some more provided rough time estimates, as well as some extra problems I haven’t thought about.

But be careful what that proves. It proves the problem is real and shared. It does not prove the size, because most people never reply to a channel post, so the true number is most likely higher - and messier than the replies.

But by all means, talk to your users.

Show what you already ruled out

Before you propose anything, write down what you already ruled out. The cheaper fix, the process change, the feature that might already exist - you get my point.

This step is what makes the whole case credible, and most people leave it out (I know I did in my first proposals). It’s also the real point of this whole newsletter - but why AI? And maybe not AI?

When you show your manager “I checked whether a simpler fix would do, and here’s why it won’t,” you stop looking like someone chasing the shiny thing and start looking like someone solving a problem. My dead email case belongs right here as proof: sometimes you do the check and the boring option wins, and that is a good outcome, not a wasted afternoon.

Be honest about the payoff

One trap to avoid when describing the ROI: time saved is not necessarily money saved. If the agent frees each client leader two hours a week, that’s freed capacity, not cash in the bank, unless the headcount actually changes. So frame the benefit as: “Reduced time spent searching for specific clauses” or “Faster and more consistent responses to the client” or “Reduced risk of overlooking commercial opportunities”.

“Saves us £40k a year” invites someone to prove you wrong - but if you are certain of a financial benefit, that’s of course your strongest argument.

There’s a second honesty tax with anything AI, and on a contract-reading agent it matters more than anywhere: someone still has to check what it produces. The agent gives your client leader a faster first draft of the interpretation - it does not sign off on a clause and it does not carry the liability. Say that out loud in your case. Far from weakening the pitch, that line tells a sensible manager you’ve actually thought about the caveats.

A Skill you can steal

And now, a small gift from me to you!

Paste this into Copilot or any LLM when you’re building your own case, and save it however your tool lets you keep prompts. It interrogates you, refuses to write anything until you’ve got a real number, and is perfectly willing to tell you to drop the idea.

You are a sceptical business-case reviewer. Your job is not to help me sell an idea, it’s to pressure-test whether it holds up. Interview me one question at a time, and don’t move on until each answer is specific. Don’t invent numbers. Work through, in order: (1) What is the problem, and what does it cost us now in time, delay or errors? If I can’t put a number on it, tell me I don’t have a case yet. (2) Who has this problem, described by the task they do, not their job title? Help me size it by how many people do that task. (3) How do I know it isn’t just my problem, and is there real system data that beats my guess? (4) What cheaper or simpler fixes have I already ruled out, including doing nothing? If I haven’t checked, send me to check. (5) Only now: what am I proposing, and why does it beat those alternatives? (6) What’s the honest benefit, in capacity or turnaround, not cash? Remind me a human still checks the output. (7) What am I actually asking for: help me lower it to a small pilot, not a full yes. At the end, help me write a short plain message to a senior person. Don’t use these seven points as headings. If the honest answer is that the boring tool already does this, or the problem is only mine, say so and help me drop it.

That’s the whole method. Nothing clever about it: it just does on purpose the boring work most people skip when they’re excited about a tool.

I’m doing this live at work right now, so I’ll tell you how the contract one lands, whether it gets a yes, a pilot, or a “no, and here’s why.”

What about you? Have you ever built a case for a tool? What was the outcome? Tell me in the comments, I read all of them.

~ Move forward, keep it steady. Justyna

Before I jump in, a bit of “the making of”. The first version of this article was written within 30 minutes. I just had the best flow going, but then I decided it didn’t belong here on my Substack. It felt too…whimsical? So I put it on my LinkedIn instead. Next thing I know, the post is getting hundreds of views and likes. Generating some great discussions. I decide to expand on it and post it on Substack after all, but first I go to my brilliant new friend here, , and I ask him “I wrote this little thing, care to read it and see what it sparks in you?”. And boy, does he deliver. I do want to call out the validation and inspiration Lee generously gave me. You will see a couple of quotes from him below, too.

Once a thought has a name, we stop questioning whether it earned one.

AI has a weird way of validating actions and ideas. And I don’t mean the sycophancy, the agreeing bit, but rather how everything gets labeled immediately upon creation. A two-liner becomes a Skill. I have eight active Projects in Claude. A quick Copilot chat is saved under Unresolved tensions in AI adoption philosophy (not kidding).

It’s…odd. The act of naming is symbolic; it officially brings creations to life. Call the stray cat Mr Socks and suddenly you feel like a responsible owner. Give a volunteer a fancy sounding job title, and they become more relevant. Categorising is a tool that helps us think; naming a concept makes it more real - project managers know it well, which must be why they come up with those names like “Project Sequoia”.

We all trust AI outputs more than we should, it’s already been researched and confirmed.

I read this somewhere and smirked: “AI is the equivalent of that confident guy at work who can make up something that just sounds right”. Even the name for AI itself - Artificial Intelligence - confirms the very thesis of this article. Did you know the term was coined in 1955 by the mathematician John McCarthy? He must have thought that “Cybernetics” or “Automata Theory” (used previously) didn’t sound quite as powerful.

At its core though, whatever you produced with AI might still very well be complete rubbish not worth taking a second look at. Somehow though it becomes harder to dismiss. Yes, we all know LLMs are just prediction machines, but at the same time it’s been produced by Fable, come on! Or the visual is completely off but it’s now sitting in your Downloads as The Momentum Framework.webp and it’s fancier than anything you have ever created before - how dare you delete it?

There’s a risk we will treat AI outputs more seriously not because they weren’t created by a human, but because they are better formed, elegantly looking, titled accordingly. That dopamine hit we get when we see something completed is real, and it’s addictive.

: I think AI can sometimes give us the emotional reward of progress before the progress has actually happened. The idea has a name. The project has a structure. The document looks professional. Your brain says “excellent, we did a something.” Except sometimes all we’ve actually done is put a very nice label on an empty box.

Another risk I’ve noticed is naming something too early and potentially risking cutting off its growth, “freezing” the thinking. It’s tempting to call an idea “The X Framework” (by the way, are you seeing frameworks everywhere now too? Frameworks and workflows, popping like mushrooms) and then get your team stuck with it, with all its restrictions.

And it’s a fine balance, because on the other spectrum lies the Cemetery of Half-Baked Ideas (a moment of silence here).

Well, AI doesn’t let you forget so quickly.

The tab on the left serves as a reminder of all the stuff you’ve worked on. Every time I open Notion (notebook name: Spanish Vocab Hunter), I remember I should practise español, and the But Why? AI Claude Project reminds me of the LinkedIn carousels I painstakingly created (waste of time and effort, don’t go there).

It’s not a bad perk, not at all. Another cool thing I quickly noticed is how working with AI helps me fight my internal critic, the I-thought-you-were-gone-now-I’m-over-40 imposter syndrome. The moment someone helps me name something I’ve been doing (it doesn’t have to be AI btw), I feel a wave of validation and relief. A small thing I thought was random or possibly irrelevant, is now The Thing Called Something. Call it silly, but it gives me motivation to continue - and the courage to share with others.

The third benefit goes out to my neurodivergent friends, and to anyone whose brain works against them some days: your neurones create a wonderful galaxy of fireworks, convoluted labyrinths, detours and patchworks, but there are also days when the thread is gone, when starting is the hard part, when linear thinking and single focus is a no-go. If any of that is you, you either love AI already - or you will soon.

Lee writes about building his “second brain” a lot, exactly for that purpose. He lives with Parkinson’s and it’s fascinating to observe and learn from what he’s been so cleverly constructing:

So, yeah, rather unsurprisingly the moral of the story is: AI can be breathtakingly innovative for our thinking and working - it can also be a trap. Personally, I still catch myself instinctively trusting the AI-generated results sitting neatly in front of my eyes. I have to stay vigilant to keep coming back to the messy, unfinished, half-ready, MINE.

Copilot is a bit like your boomer dad. Not cool in the slightest, definitely not sexy, but trying really hard to keep up with his flashy, hyper-caffeinated kids; safe, surprisingly resourceful and always THERE.

If you work in a Microsoft environment, Copilot lives there, so you're sort of doomed to use it. But it's grown up a lot, and it's improving by the day. challenged me to write about it (hello Sophie!) and she had a point: for most people a lot of it still unknown terrain.

A couple of caveats before we start:

1. It’s August 2026 and this article has an expiry date on it - Microsoft is changing things fast.

2. What you can see and do in Copilot will depend on your license AND on your company settings. Your IT can disable some options so my Premium plan might be different to yours. If you aren’t a Premium user, the options are limited. It’s always worth asking for an upgrade! I did, and was given one.

Ok, now over to practical stuff.

Notice two important toggles before your start working:

1. Copilot sits on a layer Microsoft calls Work IQ, which can draw on your internal emails, files, meetings and chats. If you turn it off, it goes to look up general internet knowledge.

1. You can now choose which AI model answers you. Copilot runs several, including Anthropic’s Claude and OpenAI’s models. If you would rather not think about it, leave it on default and Copilot picks something sensible for the task.

   Note: Claude is switched off by default in the EU, EFTA and UK, because Anthropic's processing sits outside the EU Data Boundary.

   

Five layers from chatbot to Cowork.

The foundation is naturally the chatbot - I'll skip it as it’s no different to any other you know.

Layer one: Copilot inside the apps you already use

This is the small rainbow icon sitting in Word, Excel, PowerPoint, Outlook and Teams. For its first couple of years it was fairly basic but now, it’s been a game-changer for me.

What that can look like in practice:

Word: draft a whole document from a brief and then refine it in place, structure, tone, tables and images included. Thanks to Work IQ it can pull relevant material from your own files and emails, and it will offer a summary when you reopen something complicated after a break.

Excel: point it at messy data and ask it, in plain English, to clean the blank rows and inconsistent formats, then find the trends and build a summary sheet with charts. No formulas required, though it will still explain a formula in plain English if you ask. Cool thing: it notices inconsistencies and suggest fixes itself (for example, one row has US & Canada and the other USA & Canada).

PowerPoint: build a deck from a topic or an existing document, grounded in your actual project rather than generic filler, then reorder the story or restyle the slides just by asking.

Outlook: summarise a long email chain in one click, get help lengthening or shortening a reply, and have it suggest meeting times and an agenda from a thread. Some people say the editor is weak, but I’ve found it quite good with some of my e-mail rants professional responses.

Bonus: the Premium license will give access to search your OneDrive and SharePoint sites, so the searches are so so so much faster.

Layer two: Pre-installed Agents

M365 Copilot agent is a specialist assistant pointed at one kind of job. Two of them (or more, depending on your set-up) are already sitting in Copilot, under All Agents, then Built by Microsoft. You do not build these, you just open them and go:

Pre-Installed and Pre-Pinned Agents

• Researcher: Deep-dives into internal data sources and web context for multi-page briefs.

• Analyst: Processes structured data sets, runs calculations, and visualizes trends.

Layer three: Notebooks

A Notebook is a workspace you fill with the files, policies, meeting notes and email threads for a single project, and then talk to as one searchable thing. If you have used Google’s NotebookLM or Claude’s Project, this is that idea living inside Microsoft. You can give it standing instructions such as “always consider our limited budget,” move useful chats into it, and search across everything at once.

I will be honest, I only learnt about this function when researching for this article and I am actually childishly excited to get back to work and try it!

Layer four: Building your own agent

Ok so rung two was using an agent someone else built. This is a whole other game: building your own, for a task you do again and again: the same briefing you keep retyping, a process your team needs to follow the same way each time, or knowledge you want to capture before someone leaves.

The steps, roughly:

1. Plan it first. Write down what it should do, who will use it, and what it needs to know.

2. Go to Agents, then Create an Agent, then Configure.

3. Give it a clear name.

4. Write its instructions as if you were briefing a new team member. This becomes its permanent brief, so be specific and generous with detail.

5. Add its knowledge sources: the files and documents it should draw on.

6. Bonus: Set a few starter prompts so other people know how to begin.

Once it exists, everyone who uses it gets the same considered output, without you needing to be in the room. Before sharing your agent though you need to follow certain governance steps, which should exist in your company. If they don’t, that’s a whole other problem I probably need to write a post about.

Layer five: Cowork

This is the top of the ladder and the newest part. The earlier rungs help you do a task, whereas Cowork does the task for you. You describe the outcome you want, it grounds itself in your relevant emails, meetings and files, builds a plan, shows it to you, then works through it across your Microsoft 365 apps, pausing for your approval before anything consequential like sending an email.

It can draft and send emails, set up meetings, create documents and decks, post in Teams, tidy your files, run deep research, and repeat a task on a schedule so a recurring job simply happens. It runs in the cloud and on mobile, so a task keeps progressing after you close your laptop.

Now the caveat: Cowork needs a Microsoft 365 Copilot licence, might or might not be enabled for you, and beyond that it is billed by usage (in what Microsoft calls Copilot Credits), priced per task. If there is an IT or finance team who own that budget, you will likely need their approval first.

A few things that don’t sit neatly on the chatbot-to-cowork ladder, but are worth a look:

Pages is a shared canvas. When Copilot gives you an answer worth keeping, you can send it to a Page and build it into a proper document you and your colleagues edit together, instead of watching it disappear up the chat scroll (similar to Artifacts in Claude).

Vision lets you show Copilot what is on your screen, or point your phone camera at something, and then ask about it out loud. Stuck in a fiddly settings menu, or looking at a chart you did not build, you can just ask “what am I looking at” rather than trying to describe it.

Create is the image tool (worth knowing the name, since the old Designer app folded into it). Describe what you want and it makes images or a simple infographic, and it will also clean up a photo, drop out a background or recolour something to your brand. Not Canva, but closer than you would expect, and fine for a quick header or an internal slide.

What about you: how many of these functionalities were you familiar with? And if you have built an agent or a Cowork flow that saved you real time, tell me what it does in the comments, because those are the ones the rest of us want to steal.

~ Move forward, keep it steady. Justyna

Right, welcome to Part 3 of my little Substack trilogy! You can read it on its own, but you’re welcome to start with the two previous articles.

So you’ve identified the repetitive task where you are the bottleneck.

You’ve checked what’s available to you to automate this task outside of AI. You’ve decided that it’s indeed AI you need. And you already know what level of AI you want: chatbot, repeatable skill, automated workflow, or full agent.

Nice! - you’re further than 90% of the population (my guesstimate only, don’t quote me on it).

Ok, but what’s next?

Let me start by saying: if anyone’s telling you a specific LLM, AI tool, or agent is a must-have, they’re lying. Sure, some people out there understand the AI solutions landscape really well. But most of us are barely keeping up. Yes, it’s fun to experiment but there isn’t enough time in the world to try everything. Yesterday’s hype is today’s old news.

When I first started branching out from ChatGPT and Copilot to other options, I was fairly intimidated. LinkedIn was screaming “10 AI tools you have to try.” Tech bros were like “You’re already behind”.

Soon enough though, everyone seemed to decide Claude was the best thing out there. Around the same time, I won the battle at work for an upgrade to Microsoft 365 Copilot Premium. I also got Gemini Pro through a course I took. Slowly but surely, I found uses for different tools, and tools for different needs.

I won’t give you ready recipes here. Everyone has to try what works for them, and it’s likely your company has rules you need to follow, too. But I’ll share what I’ve learned, from people wiser than me and from my own trial & error - that continues every day.

Scenario 1: Use what’s already in your toolkit

Before you go looking for something new, use what you already have. Tell Copilot, Gemini, Claude, or ChatGPT what you’re trying to achieve, and ask how it can help. As simple as that. If you’ve been treating your LLM as a chatbot only, you might be surprised what else it can do.

Look at this example - all I did was ask Copilot Agent Builder a simple question and it came up with all of the rest within 20 seconds:

Very often, nothing is required from you in terms of actual building. You might need to click a couple of “approve” buttons, test the setup, and give feedback. But I promise, it’s simpler than it sounds.

And don’t get too hung up on whether it’s called a Skill, an Agent, a Workflow, or something else entirely. Ultimately, it needs to work the way you need it to work, and label is just a label.

Did you get where you needed to go? Or did you go all in, the workflow broke, and you got confused and discouraged? Start small. Test easy improvements, iterate, have fun with it.

Scenario 2: I want a new toy!

Ok, you want to expand - fair enough, you’re my kind of a person.

Consider one of two routes:

Route 1: upgrade what you already have.

For me, that meant Claude Pro, for the Coworking capabilities, Artifacts, and Connectors. This alone will let you build a ton of AI workflows that save you real time, without adding a new tool to learn. I haven’t tried Coding (yet) but that’s there too.

ChatGPT is currently the strongest all-rounder for image generation, its GPT Image 2 model tops the quality rankings for realism and getting text right inside an image, so if you need a decent visual and already have the subscription, don’t go shopping elsewhere first. Gemini’s current edge is scale: a context window large enough to hand it an entire report, a stack of PDFs, or an hour of video in one go, which makes it the one to reach for when the task is “read all of this and tell me what matters,” especially if your company already runs on Google Workspace.

Last but not least, if you’re in Microsoft environment at work, and on the unpaid version of Copilot - consider coming up with 3-4 ideas how the upgraded version would save you (and your company) time, and show it to your manager. Feel free to use my prompt and the decision matrix I outlined in my last article Chatbot, Skill. Agent - or Maybe Nothing At All?. Most of the corporations have the budget for that, they just want to be sure the ROI is there.

Route 2: bring in a dedicated tool.

The below is by no means an exhaustive list, and I have tried half of these tools roughly myself, but they keep coming up in the recommendations online.

Category: Notes and thinking.

Notion and Obsidian. Use Notion when you need a database, not a document: something with statuses that move, a content pipeline (Idea -> Draft -> Published), a client tracker, a reading list. Set the properties up once, and it becomes something an AI tool like Claude can update for you by reading a sentence, instead of you clicking through fields by hand (this is exactly how my Kanban board updates itself, from Mix and Match below). Use Obsidian when you need somewhere to think that nobody else will ever tidy up: meeting scribbles, half-formed ideas, the ugly first draft of this very article. Turn on Daily Notes, dump everything into today's note without worrying about where it "belongs," and let the backlinks surface the connections on their own later.

Category: Automation.

Make and Zapier both connect the apps you already use so they talk to each other and automate certain tasks. An easy beginner example for an automation is sending a notification in Slack whenever you receive a specific new email in Gmail.

Zapier’s free plan gives you 100 tasks a month, one step at a time. Make’s free plan gives you 1,000 operations a month. Zapier is the easier one apparently, but some will say Make is the better tool once you actually know what you’re automating.

Category: Building your own app or website.

Yes, really, no coding background required (I almost didn't believe it either). Lovable lets you describe the app you want in plain English and get a working, full-stack product back, database and all. I used it to create a beautiful, fully functioning landing page - it took me less than hour, on the free version. Take a look here: https://butwhyai.lovable.app

Category: Research.

Perplexity and Gemini Notebook (ex-NotebookLM) both call themselves research assistants, but they’re answering different questions. Perplexity searches the live web and cites its sources, unlimited basic searches for free, a handful of deeper “Pro Searches” a day before it asks you to pay. Gemini Notebook stays anchored to documents you upload, and it’s still completely free. Use Perplexity when you need to know what’s true out there right now. Use Notebook when you need to know what’s true in the 40-page report sitting on your desk, or when you want to make a podcast out of it. I have a whole article on fun uses of NotebookLM here.

Category: Voice and audio.

A little Polish pride here: ElevenLabs, the voice-cloning and text-to-speech company now valued at $11 billion, was founded by two guys from Warsaw who grew up frustrated by badly dubbed films. The free plan gives you 10,000 characters a month, enough to try it before you trust it with anything real.

Category: Design.

Canva remains the easiest tool for anyone without design training: Magic Studio’s free tier gets you a handful of Magic Design layouts and Dream Lab image generations a month before Pro unlocks the rest. If you actually need commercially safe, high-fidelity image generation and you already live in Creative Cloud, Adobe Firefly is the more serious tool, it’s trained on licensed content, which matters the moment a client asks where an image came from.

Category: Dictation.

Wispr Flow Just talk to any app where you would normally type - both on your laptop and phone. Wispr Flow turns your speech into clean, formatted text (removes repetitions, ummms, correct errors) and it’s how I now write half my Claude and What’s App messages. Free plan: 2,000 words a week on desktop, 1,000 on iPhone. It’s enough for me, and the great thing is that it transcribes Polish really well, too (as opposed to Claude voice recognition). Great for dumping your thoughts.

Category: Music.

For the “just for fun” category: Suno and Udio both turn a text prompt into a full song, vocals included. Suno’s free plan is the more generous one right now: roughly 10 songs a day, though nothing you make on the free tier comes with commercial rights. Udio gives you 10 songs a month free, but downloads are currently switched off while it sorts out licensing.

Mix and match.

None of this has to live in its own silo. The benefits compound once you connect these tools to each other, so they stop being separate apps you switch between and start being one system that does what you say.

On my own setup, Claude cleans up my MacBook desktop, organises my photos, and adds reminders straight to my Google Calendar. I connected it to Notion too, and once I added Wispr Flow into the mix, I can just talk (out loud, to my laptop) and Claude updates my Kanban boards or dashboards in Notion on its own. None of these tools is doing anything clever on its own here - the magic is entirely in the connection.

A few honest caveats before you run off and install six new apps. Free-tier limits change constantly - everything above is accurate as of writing this, but check the pricing before you build a workflow. And don’t switch tools just because a LinkedIn or Substack post told you to. Going deeper on one will most probably give you better results than jumping from one to another.

If you got here, thank you so much for reading, and do leave even the tiniest comment. Was this useful? What would you add? What’s actually in your toolkit right now?

~ Move forward, keep it steady. Justyna

Last week I argued you should not use AI unless you absolutely have to. This is part two of three - and it's the how: once you've decided a task is worth it, what form should the help actually take?

Today, I want to give you a prompt and a framework to map your repetitive work and match it with an appropriate solution.

This practice started when I did the intense 6-week AI course run by Google and SGH Warsaw School of Economics. I got inspired by one of the classes and used the prompt and all the advice, and then re-purposed it slightly for myself. It served me a number of times already and I shared it with a couple of my friends & family too.

But first, this visual shows the layers of automation and “AI-ness” we are working with today:

0 - Leave it human. The base of the pyramid, and the biggest layer on purpose. This is work that runs on your judgment, your voice, or a real relationship: deciding what to write, solving someone’s specific problem, the actual conversation. You can hand bits of it to AI. You mostly shouldn’t. If a task is different every single time and the quality depends on you, it stays here.

1 - LLM support. The lightest touch. “LLM” is just the chatbot itself - ChatGPT, Claude, Copilot, Gemini - and “support” means you open it when you’re stuck and close it when you’re done. Good for one-off thinking: untangling a messy paragraph, getting unstuck on a tricky email, asking “what am I missing here?”.

2 - Skill or Gem (or similar). A Skill (in Claude) or a Gem (in Gemini) - ChatGPT calls it a Custom GPT - is a little reusable assistant you set up once with your instructions, your style and your context. After that you just open it instead of re-explaining yourself every time. It makes sense to create it when a task looks the same week after week but the content changes: a newsletter, a client summary, a certain kind of reply. Just tell Claude: “save this into a Skill”. ***A quick warning on words: Microsoft calls its reusable Copilot helpers "Agents", which is the same word people use for the fully autonomous kind at the top of the ladder - but in fact, they are closer to this second layer.***

3 - Non-AI automation. Notice this rung has no AI in it at all, and that’s the whole point. If a task is identical every single time - copy these numbers into that sheet, send this reminder every Monday - you don’t want a clever model that might “improvise”. You want a boring, reliable rule: a spreadsheet formula, a scheduled export, a simple automation (Zapier, Make, a Power Automate flow at work). Cheaper, faster, and it never fabricates. Reaching for AI here is using a chainsaw to butter toast.

4 - AI agent. The loudest, shiniest layer, and the one you need least often. An “agent” is AI that doesn’t just answer, it does: it takes several steps on its own, uses tools, and carries a task from start to finish with limited hand-holding. It sits at the top because it’s the narrowest slice of your real work, not the widest.

Now, the promised framework.

You will need an hour of your focus. Be prepared to list all the tasks you’re doing on a regular basis. One tip: run it on a single role, not your whole life at once. You'll get a far cleaner picture.

Here are the steps:

1. Open your LLM of choice - I did this exercise with Claude Opus 4.8 set to High.

2. Copy and paste this prompt:

   You are helping me decide, task by task, whether to use AI in my work
   and how. Interview me, ONE question at a time, waiting for each answer.
   
   1. Context. Ask which single role we're mapping (if I say "everything",
      make me pick the one where repetitive work costs the most hours) and
      what it's there to deliver.
   
   2. Brainstorm. Ask me to list tasks I do repeatedly, prompting with four
      types: things I make, analyse, organise, communicate. If an answer is
      vague ("reporting"), split it into concrete tasks and confirm.
   
   3. Score each task. Ask three things: frequency (daily / weekly / ad hoc),
      how long one takes, and repeatability 1-5 (1 = different every time,
      full judgment; 3 = same format, new content; 5 = identical, mechanical).
      Don't score for me.
   
   4. Show the list as a table, with weekly hours (frequency x duration).
   
   5. Route each task to ONE rung, with a reason:
      0  Leave human - judgment, voice, relationships. Don't automate.
      1  LLM support - ad-hoc chat, no setup.
      2  Skill / Gem - reusable asset holding my process and voice.
      3  Non-AI automation - a script, for tasks identical every time.
      4  AI agent - autonomous multi-step work; only for chained repeatable steps.
   
   Rules: no specific tools or products. Push back if a repeatability score
   contradicts how I described the task. If tasks are steps of one process,
   tell me to build ONE Skill, not several. Never route judgment or personal-
   voice work above rung 1, even if it looks repeatable.

3. Now, have fun with the interview.

4. Once you decide which tasks you do want to automate, continue chatting and experimenting to find the best tool for the job.

And when you’re done, I’d love to know: what surprised you?

Which task turned out to be a fifteen-minute distraction, which did you automate - and how, and which is the core of your individual contribution?

For me, it was the Skills that turned out to be a game-changer. I now make a point to remember to save each good prompt into a repeatable Skill, also, I save important information know I'll need to come back to as a separate Skill:

Tell me about your discoveries in the comments, I am so curious, and I read and respond to every single one.

~Move forward, keep it steady. Justyna

I know you’re overwhelmed. I know I am. You understand AI more and more, and you see it’s not black and white. There are trade-offs, there are ethical and security concerns, there is serious money being pushed by the tech overlords, and you refuse to be part of the problem.

I hope I can help make it all a little clearer and more actionable for you.

This is the first of three posts series that will tackle the questions:

1. How do I make my workday easier and more productive - with or without AI

2. What to check before you decide if AI is even needed

3. What other tools are available

4. Should I get something ready-made, or build it myself

5. What to do if I cannot build or buy but have an idea at work.

I’ll start with my story of how overcomplicating a task led me to a simple automation that was there all along.

Every month, I send the same email to the same group of people: a governance deck template, and a reminder to fill it in by the deadline. And every month, until some months ago, I did it exactly the same way: Remember the governance call is next week → Download the template. → Copy the email body from wherever I’d saved it last time. → Track down the list of recipients. → Paste, check, send. Same time, next month.

It’s not a hard task. It’s just tedious in the way where you know, every single time, that there has to be a better way, and you never quite get round to finding it. So when I finally decided to fix it - I went to AI first (no surprises there).

I then spent a good chunk of an afternoon going back and forth with Copilot, trying to get it to somehow take this off my plate. I wasn’t a great prompter back then, either. Ask, get a suggestion, try it, hit a wall, ask again - we never got to a reasonable solution.

Finally, I remembered Power Automate, a not-so-sexy tool sitting quietly inside Office 365 the entire time. I had tried it in the past but didn’t get far. Within an hour, with Copilot guiding me, I built a super simple flow that grabbed the template, filled in the email, and sent it to the same list, once a month, without me touching it again. Oh the joy..!

Does this sound familiar? Maybe you’ve been trying to find “the right AI tool” for something but you just don’t know where to start. Maybe your company keeps talking about AI but hasn’t actually shown you how to use it for real work. Or maybe you’re just not that keen on AI right now, full stop, and you’d rather stick with tools you already trust.

So, useless, this AI business?

Not at all, but there’s a real difference between what a rule-based automation does and what AI does, and it’s worth understanding before you reach for either.

A tool like Power Automate is like a vending machine. You press a button, you get the same snack, every single time. It does exactly what you told it to do, in the order you told it to do it, forever, until you change it.

AI works nothing like that. Ask it the exact same thing twice and you can get two different answers, because it isn’t looking up a fixed result, it’s guessing the most useful response each time from how you phrased the request. It’s closer to a sharp new intern than a machine: brilliant when the job needs judgement, wording, or working out what you actually meant. Useless, and honestly a bit risky, for “run these same five steps every month, in this exact order.”

Before you go looking for an AI tool for something repetitive, it’s worth checking three places first:

(Note: platforms like Make are free up to a certain point but most likely won’t be allowed by your company. They’re still worth experimenting with for your personal use, the workflows are super intuitive to set up).

One honest caveat, because I never publish a post without one: automation like this is only as good as the setup behind it. Say the governance deck template changes format, or the recipient list changes structure: the flow doesn’t notice, it just keeps doing exactly what you told it, even if that’s now the wrong thing. Check your flows every so often. And get the manual version of the task completely nailed down before you automate it. If you don’t know the steps cold yourself, the automation won’t either.

And if you’ve genuinely looked, and nothing available to you does what you need? That’s when it’s worth putting together an actual business case, not an AI experiment done on your own time.

→ Write down the problem

→ the steps involved,

→ how long it currently takes,

→ and what you’d want the outcome to look like & how much time it’d save you.

Take that to your manager, your IT team, or whoever handles this at your company. Let them tell you what’s possible. It doesn’t have to end in an “agent” - honestly, it rarely should. A simple automation, the boring kind, solves most of these problems just fine.

None of this means AI is never the answer. Sometimes it genuinely is, and telling the two situations apart is a skill on its own, not a gut feeling. So that’s next on my list: one post on how to actually tell when a task needs AI, and after that, one on whether to download something ready-made or build it yourself.

For today though, if you take one thing from this: check the boring tool first and tell me in the comments, I’d love to steal it.

~ Move forward, keep it steady. Justyna

I recently wrote about the safety do’s and don’t of using AI at work In it, I mention the risk of so called “shadow AI”, and as it often happens, soon after the theory and practice met: I have since started working on our company’s Federated Developers framework, where the main goal is to empower our AI builders while at the same time enforcing governance, security and auditability.

First, what shadow AI actually is

Shadow AI is simply using AI tools at work without approval or oversight. Pasting a client email into free ChatGPT because Copilot was being slow. Recording a meeting on your phone app or running meeting notes through a transcription app you found yourself. Or you’re simply curious (or is it just me..?). A 2026 PagerDuty survey found that 66% of office professionals have used AI at work despite believing it was against company policy. Verizon’s 2026 breach report saw shadow AI detections quadruple in a single year.

And I get it! Nobody does this to be reckless. You do it because there’s a deadline, the approved tool doesn’t quite do the thing, and the unapproved one is one browser tab away.

But the classic version of shadow AI, pasting text into a LLM chat window, is a one-off event. The data leaves once, you get your answer, it’s over. Bad enough, but contained.

Agents change that completely.

Chatbot vs agent: the one distinction worth understanding

A chatbot answers you. An agent acts for you. It can read your inbox, move files, send messages, update spreadsheets, kick off other automations, all without you watching each step.

Think of it this way: asking a chatbot a question is like asking a clever colleague for advice. Building an agent is like hiring a personal intern, handing them your badge and your inbox password, and telling them to handle things while you’re in meetings.

Now imagine everyone in the office quietly hiring their own intern. Nobody keeps a list of who hired whom. The interns have their employers’ access rights. Some of them keep working after their employer changes teams, goes on maternity leave, or leaves the company entirely. That’s agent sprawl.

And it’s not hypothetical. Microsoft’s own 2026 data showed active agents in the M365 ecosystem grew 15x year over year, wildly outpacing the governance built for the chatbot era. Meanwhile Verizon's breach investigators found that 67% of employees accessing AI on corporate devices do it through personal accounts, completely invisible to their companies.

Three things make agents riskier than chat, and the third one raises the biggest alarm bells for me:

1. They stay. You built it on a Tuesday, it runs every Tuesday forever. Including after you’ve forgotten it exists.

2. They inherit your permissions. An agent built by someone with access to HR data has access to HR data. It doesn’t know which files were the sensitive ones.

3. They compound. One agent’s output becomes another automation’s input. When something goes wrong three steps down the chain, good luck tracing where the wrong number came from. And IBM calculated what these breaches can cost: incidents involving shadow AI cost on average $670,000 more, largely because nobody can reconstruct what touched what.

What I learned co-creating a federated developers framework

These numbers are truly shocking and if you’re already hesitant about AI, I can see how those risks might make you go “Nah, thanks, nope”. But my Substack is ultimately about safe AI adoption, not AI fear-mongering, so stay with me please.

Agents - used cleverly - can do amazing things, save you time and uncover new possibilities. So instead of pretending only IT builds things, you accept - or even encourage - that motivated people in every department will build agents anyway (they will, the stats above prove it). And you give them a process: a log of who builds what, training on data and permissions, guardrails agreed upfront, and a named human owner for every agent. The builders get legitimacy and support and the organisation gets visibility. Nobody has to hide their clever automation like beer in a brown bag.

What you can do, without waiting for your company

If you are tempted to build agents, here are a few habits worth keeping:

• Bring your use case to your manager/IT/Product approvers first. Perhaps there’s an automation for it already, perhaps someone else built an agent or is building one.

• Tell someone. If your company has any kind of AI channel, register what you’ve built. If it doesn’t, tell your manager anyway. The sentence *”I’ve automated X, it has access to Y, here’s how to switch it off”* is the cheapest insurance you will ever buy.

• Find a buddy. Do a 4-eye-check. Pressure test the agent with someone.

• Keep your own inventory. One list, anywhere: every agent, automation or scheduled AI job you’ve set up, what it touches, and when you last checked it. Keep it updated.

• Know what your agents can access. Before building anything, ask yourself: if this agent went wrong, what’s the worst file it could read or send? If the answer includes client data or HR information, that’s not a solo project.

Where to be careful

My usual reminder, louder this time. An agent produces work while you’re not looking, which means your verification habit matters more, not less. An LLM fabrication (I choose to use this word instead of “hallucination” inspired by this article by and ) sits in front of you waiting to be caught. An agent’s fabrication gets actioned before you’ve seen it.

So: spot-check your agents’ output regularly, especially anything involving numbers or client-facing text. Document what the agent is supposed to do, so you can notice when it drifts. And think before you chain automations just because you can because every link you add is a link you might one day have to debug (or your company will need to do it).

Your judgement is the key component, something that keeps you both safe AND relevant. Keep it in the loop deliberately.

So now I’m curious: have you built any agents or automations at work, official or otherwise? Does anyone besides you know they exist? Or if your company introduces them, do you know how to request a new agent or how to try and build one?

~ Move forward, keep it steady. Justyna

---

Sources:

• PagerDuty, 2026 Shadow AI Survey (Wakefield Research, 1,250 office professionals; 66% used AI against policy), 2026 - https://www.pagerduty.com/blog/ai/shadow-ai-workplace-survey-2026/

• Verizon, 2026 Data Breach Investigations Report (shadow AI detections up fourfold; 45% regular AI users, up from 15%; 67% via non-corporate accounts), 2026 - https://www.verizon.com/business/resources/T1ae/reports/2026-dbir-data-breach-investigations-report.pdf

• Microsoft, 2026 Work Trend Index (active agents in M365 up 15x year over year, 18x in large enterprises), 2026 - https://www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization

• IBM, Cost of a Data Breach Report (breaches involving high levels of shadow AI cost ~$670K more on average), 2025 - https://www.ibm.com/reports/data-breach

Over the past weeks, I have shared a fair amount of warnings and advice, some of it fairly technical. Contrary to what my LinkedIn selection shows, I have to remember that a lot of us are still in the very early days of using AI. Some recent conversations I had with my friends made me realise that the overwhelm is real, and many people haven’t started at all, or never moved outside of asking Copilot for advice.

And I get it! Every week there's a new AI tool you're supposed to master. A new workflow, a new agent, a new thing your LinkedIn feed insists will make you obsolete if you don't adopt it by Friday. Then there are doom & gloom articles about AI destroying the job market, or worse. Hell, maybe even this very newsletter makes you roll your eyes.

So here's a suggestion that isn't about productivity or getting ahead. It's about play.

The tool is NotebookLM, Google’s free research notebook, and the feature is Audio Overviews: it turns whatever you feed it into a podcast, complete with an AI host talking through your material like it’s been genuinely curious about it for years. You don’t need a prompt engineering course. You don’t need to understand large language models. You upload your own stuff, and it makes you a show about it.

I want to show you three ways to use it that have nothing to do with work efficiency and everything to do with enjoying yourself:

1. coaching yourself through a career question,

2. giving your kids a personalised travel guide,

3. and learning new information fast.

How this actually works, in short

Before the three stories, let me give you the mechanics, because you’ll need them for all three:

1. Go to notebooklm.google.com and sign in with your Google account. It’s free, you don’t need to download anything either: the free plan gives you 100 notebooks, up to 50 sources each, and audio overviews included, so you won’t hit a paywall doing any of this.

2. Create a new notebook, then add sources. This is the part people don’t realize is so flexible: you can paste text directly, upload files (PDF, Word, plain text, PowerPoint, CSV), paste in website links (even a whole list of them at once), or drop in YouTube links and audio files. NotebookLM only knows what you give it. It won’t wander off and pull random things from the internet, which is exactly what makes it feel safe to play with - unless you ask it to research stuff on its own, which it can also do.

3. Once your sources are processed, open the Studio panel and select Audio Overview. Don’t just hit Generate. Click Customize first.

3. Once you have all the options you want, hit Generate, wait a few minutes while it writes and voices an actual script, then press play. If Interactive mode is available for your notebook, you can even raise your hand mid episode, ask a question out loud or by typing, and the hosts will stop, answer using your sources, and pick the conversation back up - nuts!

4. You can listen in NotebookLM or download to listen on your walk or in the car later on.

That’s it. Now, some ideas to hopefully inspire you:

Use case 1: A career coaching podcast, starring you

Earlier this year I was part of “Umiejętności Jutra,” a course run by Google and SGH, and one of our assignments was to build a podcast. I loaded NotebookLM with my CV, my background, my interests, and what I wanted my future to look like, then asked it to help me envision a career as an AI consultant.

What came out was two “experts,” one framed as a career coach and one as a tech founder, discussing my actual path, my actual skills, and my actual goals, like they’d read my file and had opinions about it. Hearing two voices talk specifically to me, not about career advice in general, was the weirdest, but fun, feeling. This very Substack exists partly because of what came out of that episode. It shaped how I started thinking about what I wanted next.

To try this yourself: paste in your CV, a note about what you’re good at, what you enjoy, and where you feel stuck. Add anything else relevant, an old performance review, a job posting you’re curious about, notes from a mentor. Then in that “focus” box, be specific: tell it to frame the hosts as a career coach and a founder in the field you’re considering, and to speak directly to you using your own material rather than giving generic advice. The more honest and detailed your sources, the more useful and less generic the conversation.

Use case 2: A tour guide for your kids that never runs out of patience

This one is inspired by a friend of mine who recently did something absolutely brilliant: before a trip to Rome with his 8 year old son, he built a notebook full of information about the sites they’d actually visit, the Colosseum, the Forum - the kind of stories that make ancient history feel less like homework and more like a tale - and generated an audio overview out of it.

They started listening on the plane. His son was hooked before they’d even landed. Once they were actually walking through the old city, the boy kept listening, and apparently spent two hours in the Colosseum just absorbing it, fascinated in a way my friend said he could never have pulled off just narrating facts himself.

To build one of these: gather your sources first, a couple of solid articles or guidebook pages about the specific place you’re visiting, maybe a Wikipedia page or two, anything reliable and not too dense. Paste them in or add the links. In the focus box, tell it explicitly who the audience is: an eight year old, or whatever age fits, curious, easily bored by dry facts, wants stories and surprising details, not dates and statistics. Ask for a shorter length so it holds a kid’s attention, and consider The Brief format if you want something tighter and more energetic than the default Deep Dive. Don’t forget to add the name! It makes it so much cooler when the kid hears the podcast talking to THEM specficially.

Then download it on a phone before you leave, because once you’re at the actual site, a kid staring at a screen for the source material isn’t the point. Listening while looking at the real thing is.

Use case 3: Digesting a pile of new knowledge fast

The third use is the most personal for me, because it’s genuinely how I sometimes keep up when I need to learn something quickly and can’t spend a week reading around it. My method is simple: I add websites I already trust, I copy and paste text, articles, or files I already have sitting around (a great way of storing them all in one place, too).

The point isn’t that NotebookLM checks facts for you. It doesn’t, and you shouldn’t treat it like it does. The point is that if you’ve already done the work of gathering good sources, it will turn that pile into a podcast that helps the ideas actually stick.

To do this: collect four or five sources you already trust on the topic, articles you’ve saved, a report, a couple of pages from a site you know is credible, and add them all to one notebook. In the focus box, tell it what level of expertise to assume you have, and what specifically you’re trying to understand or decide. Choose Deep Dive if you want real depth, or The Brief if you just need the headline version before a meeting. Then listen on a walk, or when you drive. It’s a different way to learn than reading, and for me it often works better.

None of this requires you to become fluent in AI.

You’re not learning a new skill so much as handing your own material to something that’s good at turning it into a conversation. There’s no risk of getting it “wrong,” because the worst case is you generate a mediocre podcast and try again with better instructions.

Do let me know what you end up creating!

~ Move forward, keep it steady. Justyna

Some people I work with won’t use the company’s AI tools for anything else than e-mail writing. Not because they’re resistant to technology in general, but because every time they’re about to paste something into Copilot, they panic about the PII data. Or the unreleased product spec, the salary figures, the patient note, the half-finished code, the contract a partner sent you under NDA.

I understand that instinct. Pausing before you share data with a system you don’t fully understand is good judgment, not technophobia. But paralysis has its own cost, and our customers - as much as they appreciate our diligence - expect more every day. So here is what I wish someone had explained to us earlier.

"Approved by IT" doesn't mean "paste anything"

Most of us aren’t secretly running personal AI tool on our work laptop (“shadow AI” is a topic for another post). We’re using Microsoft Copilot inside Word and Teams, or Google Gemini inside Workspace - tools our IT department contracted, approved, and paid for. Surely that settles it? Partly. The enterprise contract solves one specific problem: your data isn’t used to train external AI models. But it leaves other questions open. What gets stored, and for how long? Who in your organisation can see your prompt logs? What is safe to share or what is an absolute no-no?

Even in a properly contracted enterprise tool, you are still the data controller under GDPR. Your company is legally responsible for what happens to your customers’ personal data. Your customers agreed to you storing their information in your CRM. They almost certainly did not agree to it being processed by an AI tool.

There is also data minimization: a separate GDPR obligation that says you should process only the minimum data necessary for a given purpose. An HR analyst spotting a retention trend doesn't need 2,000 employee records - they need the aggregated rate by team. A developer debugging an error doesn't need to paste the production log with live customer IDs in it - they need the stack trace.

Which is worth defining properly, because the first time I heard this term I wasn’t sure what was expected of me here. Aggregated data shows patterns across many people, not facts about any one person. “23% of customers in this cohort churned in Q1” is aggregated. “John Smith churned after three missed payments” is not. The test: could someone use this to identify or make a decision about a specific real person? If yes - aggregate first, or don’t paste it.

So how do you get from a big dataset to an aggregated one?

1. If the dataset lives in a tool you’re already using, the answer is: don’t export it at all. Copilot in Excel, Gemini in Google Sheets, Einstein in Salesforce, Copilot in Power BI - all of these let you ask questions in plain English directly inside the platform. “What’s the average contract value by region?” The tool runs the calculation within your company's secure cloud tenant. The data is sent to the enterprise AI server, but it never leaves your organization's walled garden.

2. If those aren’t available to you, aggregate first inside the platform. Build the pivot table, run the report, export the summary. What you paste into a chat window should be the output - percentages, patterns, totals - not the source records. And if you’re staring at a 100,000-row export and none of this is set up, that’s a conversation with IT, not a workaround problem (louder for those at the back, am I right?).

3. When you do need to work with real data (a specific contract, a case study, a draft email about a client) there are techniques that actually work. Pseudonymization replaces real names with consistent codes: Client A, Employee 001. The AI can still reason about relationships between them. Generalization turns “Sarah, 34, from Manchester” into “a mid-career professional in the UK” — you lose precision but keep the pattern you need. For contracts, fictional substitution works well: swap the real company name and figures for plausible fakes. “Acme Corp, €500k, renewal March 2025” becomes “Company X, €500k, renewal next quarter.” The legal structure you’re asking AI to help you analyse doesn’t change.

What your IT team should have verified

Training on your data is off by default in properly contracted enterprise tools -Microsoft, Google, Anthropic and OpenAI all state this explicitly in their enterprise terms. But there are other things worth asking IT to confirm if you want to be extra-safe:

• how long conversation logs are retained,

• who in your organisation can access them,

• which third-party plugins or integrations are active (each has its own data terms, sometimes negotiated separately),

• For EU organisations using Copilot or Gemini, data residency settings are worth checking - you can often pin processing to European servers.

When you’re not sure, here’s who to ask

Many people don’t know where to go, feeling they should somehow know what the right thing to do is. Do not assume. We are all new to this.

First, your company will have created an AI Policy by now, so start by browsing your intranet. Then, your Data Protection/Data Privacy team is the right call for anything involving customer, patient, or employee data. If you don’t know who yours is, HR or Legal will. IT and InfoSec handle questions about approved tools, active integrations, and - if you’ve accidentally pasted something sensitive - audit logs and actual exposure. Legal and Compliance own anything touching contracts, IP, or sector-specific rules (think clinical, financial, or education records:HIPAA, FCA, FERPA).

Here’s a template you can use when contacting a relevant department - tried and tested:

I want to use [Copilot/Gemini] for [specific task — e.g. analysing customer churn patterns / drafting contract summaries]. Before I do, I wanted to run it past you.

What I’m planning to use: [describe the data or documents involved]

What’s the outcome I am trying to achieve: [describe the end product/output]

What I’ve already done to reduce risk: [e.g. I’ll aggregate the data first and remove names / I’ll use anonymised examples only / the documents don’t contain customer PII]

Where I’m not sure: [e.g. the report includes contract values / I’m not sure if the plugin we use has been reviewed / is XYZ considered PII data].

What’s the thing you’ve been hesitating to use AI for at work? Is it actually a data risk, or just uncertainty?

~ Move forward, keep it steady. Justyna

A while ago I created an AI news monitoring dashboard in Google Sheets (following ’s presentation, thanks!). I checked it today, and more items are landing in Society & Ethics category than in Technology & Tools - that was a surprise. This suggests that it’s not just my bubble then, maybe the conversation has in fact moved from chasing the next wonder tool to facing uncomfortable consequences of using them.

I won't be sharing easy tips or productivity hacks today. I want to go back up a little and tell you a couple of uncomfortable truths about how I was using technology: at the beginning of the AI hype, I outsourced more than I care to admit. I asked for the outline before I’d decided what I thought. I read the summary instead of the thing. None of it felt like a cost - it felt like a superpower!

Then I started reading the cognitive science behind it all, and the landing was brutal.

Here is the optimistic story, the one I told myself first: we have always offloaded thinking. Writing was supposed to wreck our memory (Socrates said so). Calculators were supposed to kill arithmetic. They didn’t. And don’t even get me started on the Internet. Every generation panics about the new tool and the species survives, AI is just the next calculator - relax.

I wanted that story to hold. But it doesn’t, not quite. Not all offloading is the same, and the science is getting specific about the difference.

Start with GPS, because it is the cleanest case we have. A 2020 study in Scientific Reports tracked regular drivers and found that the more people relied on GPS, the worse their spatial memory got over time (hi, hello, yes). The mechanism is not mysterious: building your own mental map engages your hippocampus. London taxi drivers, who memorize the whole city (famously, The Knowledge), literally grow that part of the brain! The rest of us, following the arrow… we let it shrink.

Now the AI version. An MIT Media Lab team wired people up to EEG headsets and had them write essays. Three groups: one used ChatGPT, one used a search engine, one used nothing. The ChatGPT group showed the weakest brain connectivity. By the last session they were copy pasting. When asked to quote a sentence from the essay they had just written, many couldn’t. They didn’t feel ownership of it, because in a real sense they hadn’t done it.

But the most useful finding wasn’t the scary one. It was about order. The people who wrote unaided first and got the AI later stayed sharp and used it well. The people who started with the AI and then had it taken away struggled. The difference was whether they had built the thing themselves before they handed it over.

A third study, from Microsoft and Carnegie Mellon, surveyed 319 knowledge workers. The more someone trusted the AI, the less critical thinking they did. And we know we tend to trust smooth, confident writing more - and that’s exactly what LLMs are trained to produce. And the worst part: people skipped the thinking most on exactly the tasks they were least equipped to judge. The less you know, the more you defer. The more you defer, the less you ever learn enough to stop deferring.

So here is where I’ve landed, and it isn’t comforting: the calculator was safe because I learned arithmetic first and kept the number sense. The GPS kind of offloading is different. It does the map-building for me, until I can't build the map on my own and now I know I'm not the only one who can feel their spatial neural pathways gathering cobwebs.

So, what needed to change?

• I now start with a blank page, whether it's an email, a LinkedIn post or a Substack article. Because I know I can always use AI later, I don't get stuck chasing perfect. I just keep writing. And because I loaded Claude and Copilot at work with my own instructions, style and examples, it keeps me, me. The “devil's advocate” and “challenger” orders are in there too, so I never have it too easy (find out how in my old post here:)

• I still use AI summaries for long e-mail threads or lengthy articles but I am cautious. I will double-check the crucial parts for work, and if the article is important to me, I will spend time poring over the details. If anything, digital cognition traps teach me to be extra-cautious these days.

• I embrace the beneficial offloading: using AI to format paragraphs, suggest a better English equivalent, clean up messy data, or fix syntax errors. This frees up my memory.

• but I am more critical now, deliberately looking for friction. I actively research data and papers on cognitive science in the age of AI & tech. I follow fascinating people who shine a light on what is happening to your brain, psyche and relationships. They don’t have all the answers, some results are contradictory (my favourite dispute these days: should you be nice to AI or not?), but it’s been the most intellectually challenging period of my life post-university.

What are your experiences? What rules have you set up for yourself? And if you are deliberately not using AI, do you see a future in which you might start doing it?

~ Move forward, keep it steady. Justyna

Sources:

- Dahmani & Bohbot, “Habitual use of GPS negatively impacts spatial memory during self-guided navigation,” Scientific Reports (2020). https://www.nature.com/articles/s41598-020-62877-0

- Kosmyna et al., “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,” arXiv preprint (2025). https://arxiv.org/abs/2506.08872

- Lee et al., “The Impact of Generative AI on Critical Thinking,” CHI 2025, Microsoft Research and Carnegie Mellon. https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/

I was comparing two insurance policies a couple of weeks ago: nothing complicated, just lots of boring legal jargon with two offers. In the spirit of automating what I absolutely hate and procrastinate with, I asked AI to help me figure out which one was better value over time.

It got it wrong. Twice.

First, it told me one policy was 2.5x more expensive than the other. When I pushed back and asked it to show its working, it turned out the opposite was true: the policy it flagged as expensive was actually cheaper by about 10 PLN a year. It had the comparison completely backwards.

The second mistake was more fundamental. One policy was priced for a 3-year term, the other annually. The AI compared them as if they were the same thing. It never flagged the difference, never adjusted for it, it just... proceeded, confidently, with the wrong assumption baked in.

Gritting my teeth, I started grilling Claude to understand the reason of its incompetence. To my surprise (and frustration) AI said something I thought was worth sharing: “I failed because extracting the correct variables from a wall of legal text is a complex language processing task that tripped me up before I even got to the arithmetic.”

Which is a strange thing to hear from a tool people increasingly trust with financial decisions (and yet another proof AI is not taking our jobs just yet).

Alright, so I needed to understand: what is actually happening here?

LLMs - large language models - are exactly what the name says. Their core job is processing and generating text, not performing arithmetic.

Here, it helps to know one basic thing about how AI reads text. It doesn't process words and numbers the way we do, but it breaks everything down into "tokens," which are small chunks of text, sometimes whole words, sometimes fragments. Numbers get tokenized too, which means AI isn't reading "178.54" as a value — it's reading it as a sequence of characters. [if you’re curious, this Substack post by Dr Teodora Szasz goes deep]. This matters because when numbers are scattered across a long, context-heavy conversation, the model has to do significant language work just to figure out which numbers are relevant before it can attempt anything resembling arithmetic.

When you ask an AI a math question, it doesn’t reach for a calculator the way you might. It predicts what a plausible-looking answer should be, based on patterns in the enormous amount of text it was trained on. For very clean, isolated problems, let’s say “what is 1449 times 4000”, this works reasonably well because modern AI systems can hand that off to a computation layer that gets it right.

But the moment math gets buried inside context - multiple numbers floating around, different timeframes, conditions that affect which figures actually matter - the model has to first understand the language before it can attempt the numbers. And if it misses the point, like it did with my policy questions, it just makes stu

So - useless? Absolutely not.

Here’s where you will find it reliable:

• Structuring a comparison. If you give AI two offers and ask it to lay out what’s included, what’s excluded, and what questions you should be asking, it does that well. That’s language work.

• Spotting what you might have missed. In the same insurance conversation, the AI correctly identified differences in coverage scope and flagged terms I hadn’t noticed. That analysis was useful and accurate.

• Simple, isolated calculations with no context to confuse it. “If I pay X now versus Y per year over 3 years, which costs more?” - give it just that, and it handles it fine.

Where to be careful:

Hopefully you've already noticed that despite my enthusiasm for AI experiments, I remain critical and will continue to remind you of the constraints and guardrails you need. In this case, I urge you to use your judgement and verify the numbers. It might sound counter-productive to have to check AI’s logic and numbers yourself, but until the models get better, it is still worth the effort.

Of course, prompting is key:

• Chain of Thought: Tell the AI to “think step-by-step.”

• Extract First, Calculate Second: Instruct the AI to first extract all the relevant numbers into a clean table, verify that table with you, and only then perform the calculations.

• Show the Formula: Ask the AI to provide the exact mathematical formula or logic it is using before it gives the final answer. This would have instantly revealed the 3-year vs. 1-year discrepancy you encountered.

I will say that the errors I got were on Claude’s Sonnet 4.6. When I tested on Opus 4.8 (paid subscription), the result was exponentially better. This post might very well become obsolete soon what with the advances made in this field, but until then - stay safe my friends.

I had some people telling me they benefited from my advice recently and it absolutely made my day. If you ever find any posts (not just mine!) helpful or inspirational in any way, make sure to drop a quick note in the comments - you will absolutely make someone grin all day.

~ Move forward, keep it steady. Justyna

The most time I ever wasted with AI? Getting it to analyse a big dataset.

I came in with zero context, a one-liner instruction, and absolutely no idea what I was doing. I asked it to “combine those two reports and analyse the main ticket trends". Oh the fool… What followed was enough blood, sweat and tears to almost put me off using LLMs for anything beyond basic chat.

What actually changed in 2026

AI didn’t just get smarter. It changed what it does.

It used to predict the next word. Now, when you upload a spreadsheet, it writes and runs code (Python, formulas, the lot) to generate charts, run calculations, and reshape your file. You are no longer just making charts. You are having a conversation with your data.

No, you don’t need to know Python or SQL. But you do need to know how to communicate and how to use your judgement throughout.

Treat it like a brilliant new hire (who can hallucinate)

This is the mindset shift that fixed everything for me: imagine handing a stack of spreadsheets to someone on their first day and saying “analyse this.” They’d stare at you, or try their best guesses. AI does the same, but it doesn’t tell you it’s confused - it produces something that looks completely confident and polished.

So brief it properly.

Tell it what the data is. One or two sentences: where it came from, what the columns mean, the time period. Something like: “This is a raw export from our ticket tool covering [Month]. Each row = one support ticket. Columns: Ticket #, Team, Tower, Service Process.”

Tell it the business problem. Don’t say “analyse this.” Say: “We noticed a spike in payroll tickets last month. Find which teams or service processes are driving this, and flag any anomalies.”

Tell it how to present the output. If you need three bullet points for a board slide, say so. If you want a step-by-step breakdown, ask for it. AI defaults to thoroughness. You have to impose structure.

Two techniques worth stealing

Orient → Plan → Produce. Don’t cram everything into one prompt. First, ask the model to orient itself: “What columns are in this file? Are there any anomalies?” Then plan: “What analyses would be most useful here?” Then produce, iterate, try again. Slower, but the outputs are dramatically better.

Make it ask you questions first. Before it touches your data, paste it in and say: “Before giving me any insights, ask me 3–4 questions about my business - like seasonality or margins - to make sure your analysis is accurate.” This one step prevents so many obvious errors.

And now, RISKS. Read this section carefully if you work with datasets, because in 2026, AI data errors don’t look like obvious hallucinations. They look like beautiful, polished mistakes. I remember asking Copilot completely incredulously “Where did you take these conclusions from? I am not seeing this in the graphs?” and it responding: “You’re right, I made those up”. Argh! I almost pasted them into a client MBR.

• The consistency problem nobody warns you about. One word in your prompt can change the outcome. Claude and Gemini will give you different answers on the same data. Run the same prompt a month later and you might get something different again. This isn’t a bug you can fix. It’s the nature of the tool. Stay alert. Stay in control.

• Don’t start with a specific question. If you go straight to “what were the escalation volumes by the region?” the model answers that, and says nothing about the giant anomaly sitting two columns over. Start open: “Tell me about this dataset.” Discover what you didn’t know to ask.

• Watch for double-counting - it’s the most believable mistake AI makes. Say your export has rows for London, Manchester, Birmingham, and then a “UK Total” row your reporting tool added automatically. Ask AI for “revenue by country” and it will include that UK Total row as if it were raw data - then add it again when it sums the cities. Your UK number is now roughly double what it should be.

  Fix: before any aggregation, ask “Does this dataset contain any summary, subtotal, or rollup rows?” Then explicitly tell it which rows to exclude. Never assume it will spot the hierarchy on its own.

• Don’t trust complex maths blindly. Ask AI for basic addition and it’s fine. Ask it to calculate your renewal forecast - ARR at risk, weighted by health score and days to renewal - and it might apply the weighting inconsistently across segments, or treat a blank health score as neutral rather than unknown. Your forecast is off in a direction that only becomes obvious after the quarter closes.

  Fix: never ask for the answer in one shot. Ask it to “calculate this step by step, showing the formula and the actual numbers used at each stage.” A wrong answer you can trace is fixable. A wrong answer that looks right isn’t.

The final habit: defend your numbers

Think of reviewing AI analysis the way a CFO reviews a report from a new accountant. You don’t accept it on faith - you check it. Before you take any output seriously, ask: “List the assumptions this analysis depends on.” What the model surfaces will often surprise you.

AI gets you to insight faster. It does not replace the human who decides what to do with it. Your judgement is crucial. Document your prompts, document the assumptions, and make sure you can walk someone through the logic in a meeting three weeks from now.

If you have any experiences worth sharing, tips or questions - please drop them below or just respond to the e-mail. I cherish every comment.

~ Move forward, keep it steady. Justyna

If you started experimenting with ChatGPT over the last two years, you probably asked some random questions first before moving to more specific tasks: “analyse this report”, “write an e-mail to the client about xyz”, “find best Chinese restaurants in town”. From there, you either developed more curiosity and learnt to prompt better, or gave up because “AI gave me a stupid answer”.

We already know we need to write longer prompts: add more context. Define the role for our LLM. Add examples if possible, specify the constraints. But what happens when you’re faced with a bigger challenge, you’re not an expert, and you want AI to be a true thinking partner, not just a vending machine? You meta-prompt. And I cannot stress it enough, it’s really crucial that you start using it ASAP.

I’ll give you an example: I recently experimented with analysing a big dataset of Reddit posts to understand the types of Personas (or rather, Jobs To Be Done) in a specific AI-skeptic thread. I know better than to ask “analyse this dataset”, and so for the sake of this exercise, I made it slightly more specific:

And so it continued - I cut the rest, but the six points was what I got. It’s not bad, right? But if you’re after a solid deep dive, with more “meat” to it, and you want to understand AI’s logic (possibly refine it, too), you really want to ask AI to tell you what to tell it. You’re basically asking:

how do I prompt you so that I get the best results, and my goal is xyz?

Or: what do you need from me to provide the most comprehensive answer?

Or: ask me clarifying questions before getting to work.

In my example, this was my actual meta-prompt:

And here’s Claude’s response - the prompt it created for me to copy/paste:

Ok, so you can immediately see this isn’t something I’d come up with myself, not in my wildest dreams. Once I ran this prompt, my results got 100 x better. For brevity’s sake, I will only paste the JTBD cards, but there was a much more in-depth analysis, too.

(By the way, what you see is Claude’s free version output, set to Sonnet 4.6 Adaptive. It’ really neat visually and that’s a bonus if you want some course or workshops materials, or social media content.)

From my experience, I will tell you that LLMs will surprise you with the level of complexity and accuracy once you specify what your context is and ask for an efficient prompt. Combine that with what I posted about last week - personalising your voice and giving Claude instructions how to think & respond - and you’re really very close to 100% correctness.

Next week I'm going deeper on dataset prompting - the risks and the tips. Worth following if that's part of your work.

~ Move forward, keep it steady. Justyna

Let’s talk about something we all grew to notice and dislike, very quickly: generic AI slop.

You know it when you see it online, you cringe when you see it as a response to your question. Technically fine - perfectly formatted - completely forgettable. It has that particular flavor of confidence with nothing behind it, like a LinkedIn post written by a committee of people who’ve never had a bad day.

People tell me the world is getting dumber, that they don’t want to use AI in fear of losing their voice and sounding like a robot. Yes, AND: when you give an LLM zero context about you, it doesn’t guess. It averages. It pulls from everything it’s ever been trained on and produces the mathematical middle of the internet. Which, as you may have noticed, is not great. So if your AI outputs sound like everyone else’s AI outputs, it’s just doing its job.

If you want the practical fix now, scroll down to the bottom where I added a quick graphic guide for Claude, Copilot, Gemini and ChatGPT.

The thing nobody tells you about “just using AI”

There’s a concept I keep reading about: the atrophied self. When you hit “generate” without injecting your actual voice, your judgment, your taste - your hard-earned perspective starts to erode. You stop exercising the muscle. That’s scary in itself, but also dangerous in terms of your career. If there’s nothing individual about your outputs, how are you going to defend at least some parts of your job?

Remember the goldfish?

In my last post I explained that AI is essentially a highly sophisticated auto-complete with amnesia. Every new chat, it starts completely from scratch. But there are ways to modify that and use to it to your advantage, and that’s exactly what the configuration tools to add personalised context/preferences exist for.

Every LLM has those “anchors” (well, apart from Copilot, but there’s a workaround). In my quick guide I focus on the four most common LLMs: Claude has Instructions for Claude and a Style feature, ChatGPT has Personal Context, Gemini Personal Instructions AND Gems*. These aren’t gimmicks or nice-to-haves. Try them out now: they are the only way to stop the goldfish from defaulting to the average of the entire internet every single time you open a new chat.

15 minutes of setup, that’s the 10% that will give you the result: your e-mails/ decks/ presentations/ posts stops sounding like everyone else’s.

~ Move forward, keep it steady. Justyna

* message me if you want to learn more about Gems, I’ve been using them a lot and they can be a lot of fun!

Once the LLM is trained, the model is static. It learns nothing new about you permanently. It doesn’t build up your personal wikipedia. It doesn’t carry information from one chat to the another. What really happens is:

Every time you type a new message and hit send, the app behind the scenes grabs your new message, attaches it to the entire history of your current conversation, and sends the whole giant transcript back to the AI. The AI reads the entire chat from scratch, predicts the next logical response, and sends it back. It feels like a continuous memory to you, but to the AI, it is reading a script for the very first time, every single time.

The risks? For one, the infamous Context Window. Imagine storing all your information on your desk, notebooks piling on top of one another. You can only fit so many books on a desk before they fall off. For LLMs, the space on this desk is measured in “tokens” (which are basically syllables or chunks of words).

If your conversation gets so long that the total word count exceeds the context window (the desk fills up), the system will start silently deleting the oldest messages from the top of the transcript to make room for your new messages. When this happens, the AI suddenly “forgets” how the conversation started.

The second big problem is the massive carbon footprint of generative AI.

Alright, so here are some tangible tips to start using today:

• Start fresh often: Don’t keep one mega-chat going for days about ten different topics. The longer the chat, the more “noisy” the context window gets. If you change topics, open a new chat.

• Give it a role and rules upfront: Because the AI reads the whole transcript every time, putting strict rules at the very beginning of the chat (e.g., “Act as an expert editor. Only give me feedback in bullet points.”) anchors the AI’s behavior for the rest of the session.

• Summarize instead of dumping: Don't copy-paste a 50-page PDF into an AI if you only need answers about chapter two. In fact, if possible do not paste PDFs at all - copying and pasting the text directly is much greener and more efficient.

• Be direct: Cut the pleasantries. Saying "Please," "Thank you," and "Could you kindly..." adds unnecessary tokens that servers have to process. The AI doesn't have feelings, so save the electricity and just give it the direct command (no, being polite to Claude won’t save you in case of a robot uprising).

Bottom line: treat your LLM as an incredibly smart colleague with amnesia. And never ever forget that you’re the one in charge.

~ Move forward, keep it steady. Justyna

* Claude Projects, Gemini Gems, NotebookLM and others are in fact second-brain-like. We will talk about those later but for the purpose of this blog entry, we’re focusing on simple chats.

Here is the truth: the 10% effort that does 90% of the work is actually very simple:

If your prompt is shorter than 21 words (yes, weirdly specific but that’s the number that seems to does the trick), you aren't prompting - you’re just using a very expensive, energy-hungry version of Google. Because LLMs are mirrors, a shallow 5-word question can only reflect back a shallow, generic answer.

The “But Why?” Logic: Before you hit enter, check your minimum. Your prompt needs a Role, Context, and a Goal. If you haven’t hit at least two sentences, or 21 words, you will be discouraged by the result.

The Example:

• ❌ The Google Search: “Make this email sound more professional.”

  • Result: Generic fluff that needs more prompts to refine. Bad for the planet, zero human judgment.

• ✅ The Logic Filter: “I am a consultant writing to a long-term client who is late on a payment. Keep the tone warm but firm about the late fee. Prioritize the relationship over the jargon.”

  • Result: An immediately more effective response.

The Bottom Line: Don’t treat AI like a magic wand. Treat it like an intern who needs a clear briefing. Keep your voice, use your discernment & your experience.

I would love to hear if your LLM adventures reflect that? Please comment with your best prompt, or your worst/best results. This is a community we’re building here, and your feedback will be super important to me.

~ Move forward, keep it steady. Justyna

#ButWhyAI #HumanCentricAI #AILogic

• Add more tools.

• Add longer prompts.

• Add more “AI-powered” workflows.

But at But Why? AI, we believe in addition by subtraction. Before you can master the 10% of logic that drives results, you have to stop the 90% of habits that cause overwhelm.

Here is your “Stop Doing” list, filtered through our three pillars of Common Sense:

1. Stop “Wishful Thinking” (Logic)

The Mistake: Using “Magic Wand” prompts (e.g., “Write a marketing strategy”) and being disappointed when the output is generic junk.
WHY? AI is an intern, not a mind-reader.
The Fix: Stop asking for final outputs. Start asking for drafts. Give the AI the “Who, What, and How” before you hit enter. If the logic is missing in your head, it will be missing in the AI’s response.

2. Stop “Tool-Hopping” (Convenience)

The Mistake: Signing up for every “Top 10 AI Tools” list you see on LinkedIn. Or the oppostite - giving up as soon as you hear a new name.
WHY? Choice-paralysis is the enemy of productivity. New tech and advanced AI looks scary before you try it.
The Fix: Pick one high-quality LLM and master the logic of how to talk to it. Ignore the rest. Ask But why? if the results are sub-par.

3. Stop “The Data Dump” (Safety)

The Mistake: Treating an LLM chat window like a private, secure company vault.
WHY? Most free-tier tools use your inputs to train their models. If you put a secret in, it’s no longer a secret - it’s training data.
The Fix: Scrub the “Proper Nouns.” Replace client names with “Client A” and specific numbers with “Variable X.” The AI doesn’t need the secret to understand the logic; it just needs the pattern.

The Bottom Line:
You don’t need a degree in Computer Science. You need a filter.

Next week, we’re diving into the first Logic Map: How to audit your own prompts to find exactly where they are breaking.

~ Move forward, keep it steady. Justyna