“AI” isn’t just for people who write code

In the first article, I was speaking to programmers. I talked about terminals, CLAUDE.md, SDLC workflows. And the closing message ended up sounding like “this is only for technical people.”

And yes, that’s true, in the context of that article. But it’s also exactly the problem I want to tackle now.

Because the most important part of that article was NOT Claude Code. It was the idea underneath it, which has nothing technical about it and works just as well for a wide range of professions: from lawyers, accountants, people managers, and doctors to roles that feel completely different, like personal trainers or marketing and advertising professionals. This article is about that idea, and about the tools that make it practical for anyone who is never going to open a terminal in their life.

I’ll start with the conclusion, because it’s the single most important thing to keep in mind:

The difference between people who get value from AI and those who only receive generic answers is not technical skill. It’s whether they treat their own documents and their own context as the raw material.

Look at the pattern: In the first article, CLAUDE.md worked because it gave the agent the accumulated context of your project, the architecture, the conventions, the lessons learned. It wasn’t a blank chat every time. It was hand?selected memory, with curation.

This is not a programming idea, it’s the core idea. Because we will always have two groups of AI tool users:

• Those who paste a vague question into ChatGPT, cross their fingers, and get an average answer pulled from the average of the internet
• Those who feed the tool their contracts, their reports, their meeting notes, their processes and get answers anchored in their own reality.

The first person is using AI as an oracle. The second is using it as an extension of their own work, and this article is meant to turn you into someone in that second group (if you’re still in the first).

Almost all the tools I’m going to show you next share exactly this trait: they are built around your material, not around a genius prompt. Keep that in mind as you read; it has to be the guiding thread.

One note before we go on: none of this replaces your judgment. These tools speed you up, but the responsibility for the result is still yours. That’s true for code, and it’s doubly true for a legal opinion, a set of accounts, or a clinical decision.

Before jumping into tools for each profession, it’s worth closing the loop with the previous article, because Anthropic has literally built that bridge.

It’s called Claude Cowork. The most honest way to describe it is: “It’s what IT people have in Claude Code, but for the rest of your work.” The same agent architecture I showed in the first article, the ability to read your files, work on top of them, and return a finished result — but with no terminal and not a single line of code. It lives inside Claude’s desktop app: you point Claude at a folder with your documents, give it a goal instead of a question, and it executes the steps.

Anthropic itself says the audience that needs this most isn’t programmers: it’s analysts, operations teams, legal and finance professionals, people who work with documents and files every day and would rather spend their time on the big decisions instead of assembly work. There are even vertical agents built by Anthropic for legal work (contract review, case law research, citation checking) and finance (results analysis, reconciliation, reporting).

Notice this is exactly the same principle again: Cowork is only useful because it works on top of your local files. It’s CLAUDE.md without the .md. One honest safety note before you start: because the agent touches your files directly, make a backup before you let it loose on an important folder, and review what it did. Autonomy is the advantage — and it’s also the risk.

If your job is essentially reading a lot, synthesizing, and producing reliable text from it, there’s a tool that was literally designed for you: Google’s NotebookLM.

Its premise is the opposite of a generic chatbot. NotebookLM only answers based on the sources you give it. You upload PDFs, Google Docs, web pages, even YouTube videos, and from there it answers, summarizes, and cites, always grounded in your material, not in some average of the internet. For anyone working with sensitive information, the game?changer is this: Google states, that the data you upload is not used to train its models.

What this unlocks in practice:

Human Resources. ou load your internal policies, employee handbook, and applicable labor law. Suddenly you can ask “How many days of parental leave apply in this specific case?” and get an answer with the exact citation of the source. You can generate an audio summary of a new policy to share with the team, because NotebookLM can turn documents into a dialogue between two hosts, perfect for people who prefer listening to reading.

Legal. Here the use case is almost obvious: you upload an 80?page contract or a case file with multiple documents and use NotebookLM to locate clauses, compare versions, and do the first pass of issue spotting. The critical point, and I’m putting this in bold on purpose: The output is your STARTING point, never your final product. The tool can misinterpret a nuance, and in a legal document that’s expensive. It saves you the mining work; it does not excuse you from reading.

Accounting and finance. Here another type of tool comes in. Skywork.ai positions itself as a kind of “Office with AI”: specialized agents that generate documents, presentations and spreadsheets, always with research and cited sources. Like NotebookLM, it builds on a knowledge base you upload, and the desktop version even processes files locally, which helps address privacy concerns. One honest warning, because I don’t want to sell you smoke: some reviews mention aggressive billing and pricey premium plans (here’s the comparison table between plans and what each one offers). Try the free plan calmly before entering your card details, I’ve been using the tool occasionally and the free tier has been fine for me. For a lot of finance work, the Claude Cowork setup from the previous section with the finance plugin does the same kind of tasks (reconciliations, variance analysis, reporting) directly on your own files.

And now, zooming out: where does the raw material come from? Very often your context is born in a meeting, and nobody enjoys being the one stuck taking notes instead of participating. That’s where Fireflies.ai comes in: it connects to your Zoom, Teams, or Google Meet calls (those are the platforms I’ve tested), records, transcribes, and summarizes automatically, with speaker detection and an action list at the end. The result, again, is your own raw material: a transcript you can then feed into NotebookLM or give to Cowork. From a privacy perspective, Fireflies says it follows major standards (SOC 2, GDPR with EU hosting, HIPAA with agreement, and a policy not to use your data to train models). But here’s a serious warning, especially if you’re in legal or HR: the tool joins the meeting as a recording participant, and in many jurisdictions recording people, especially capturing voice characteristics, requires informed consent from everyone present. There are active lawsuits around this exact point. The rule is simple: inform and obtain consent before you switch it on, especially in candidate interviews or sensitive meetings. The technology is excellent; using it responsibly is on you.

The common denominator in all these cases is the same as in CLAUDE.md: curated, persistent context will always beat a blank chat. You’re not asking “What do you know about this?”, you’re asking “What do my documents say about this?”.

There is a second group of work where AI is already very strong: turning an idea into something visual and concrete. In this space there are two tools I want to position carefully, because they are often recommended for the wrong things.

Google Stitch. Be demanding with this one. Stitch does one thing very well: designing application interfaces, app screens, dashboards, web pages, with code export. It does not do presentations, marketing materials, or social media graphics. If you are a product manager, founder, UX/UI designer, or simply have an idea for an internal tool and want to see how it would look, you describe the flow in natural language and it generates several screens that are coherent with each other. If you are in marketing and want a beautiful newsletter, Stitch is the wrong tool; do not use it just because everyone is talking about it.

Claude Design.From Anthropic, this is the right partner for what Stitch does not do, and this is where marketing and advertising professionals get the most value. It is built for broader visual work: prototypes, slides, one?pagers, brand materials, landing pages, campaign visuals. You describe what you need, it creates a first version on a canvas next to the chat, and you refine it in conversation. It exports to PDF, PowerPoint and Canva. The detail that ties it back to the core thread of this article is that during setup it reads your brand files (or your code, if you have a design system) and starts applying your colours, typography, and components automatically. Again, it builds on your own material. A marketer can generate on?brand campaign variations in minutes and then hand them to a designer for polishing; a founder can do the same with an investor deck.

One technical note that avoids frustration: Claude does not natively generate images, audio, or video. Claude Design handles layout, structure, and design using code and your assets, not photos invented from scratch. For visual finishing and collaborative editing, the natural bridge is Canva, which it exports to..

The practical rule for this group is: use these tools for the first draft and for exploring variations quickly, not for the final client delivery. They are excellent at killing the blank page. They are not a replacement for a professional’s eye on the final finish.

This is the group where I will be most cautious, and I am asking you to read that caution as part of the advice, not as a footnote.

Medicine, and where to draw the line. et me be very clear about what you should not do before talking about what you can do: you do not paste identifiable patient data into consumer tools, and you do not use any of these tools for diagnosis. Full stop. That said, there are legitimate, valuable uses around the edges of clinical work. A doctor can use NotebookLM to stay on top of the literature: you load papers from a given field, generate an audio summary, and listen in the car on the way to work. It can help prepare patient education materials in accessible language, or digest long guidelines. Administrative work such as letters, internal process summaries, and team training is where the gain is big and the risk is low.

Training and education. Here NotebookLM shines without asterisks. It turns your source material into quizzes, flashcards, mind maps and, on paid plans, into explainer videos with narration and visuals. A trainer can take a dense manual and generate several versions: one in audio for people who prefer listening, a mind map for visual thinkers, a quiz for consolidation. Because everything rests on the same sources, the information stays consistent across formats.

Personal trainers and nutritionists. The pattern is the same, applied in a new context: you load your protocols, your methodology, your reference sources, and use AI to generate client materials (explanatory plans, educational content, answers to frequently asked questions) that sound like you and follow your approach, instead of generic internet advice. The same caution as in the other cases applies: the tool drafts; you validate.

All the tools so far follow the same principle: they work from your own material. The next one does not, and it is only fair to say so. But it solves a very real friction many people feel with AI, so it deserves a mention.

It is called Superwhisper. It is not a content tool; it is voice dictation that runs on your Mac or iPhone and transcribes what you say, with good quality, inside any app, including the text box of any AI. Why does this matter? Because one of the reasons people give short, weak prompts to AI is that typing a long, detailed prompt is tiring. When you speak, you give much richer context in far less time. You describe the whole problem out loud, with nuance, and you get an answer that matches, instead of the usual telegraphic request that produces a telegraphic reply. It is one of those tools that seem small until you use them for a week and find you cannot go back.

In the first article there was a sentence I would not repeat today: the idea that these things are free. The reality in 2026 is more nuanced.

NotebookLM has a genuinely useful free plan: you get 100 notebooks, 50 sources each, and a daily allowance of questions and audio summaries that is more than enough to seriously test whether the tool fits your work. But the flashier features, like explainer videos, sit behind paid plans (Plus is included in Google One AI Premium, at around 20 dollars per month). Claude Design and Claude Cowork are included in Claude’s paid plans at no extra fee, but they draw from the same usage limit as everything else, and agent tasks burn through that faster than a normal chat. Skywork, as I already flagged, has expensive premium tiers.

The recommendation is straightforward: start with the free plan for any of these tools, use it on a real problem from your day-to-day, and only pay once you already know the value is there. Do not subscribe to anything based purely on a demo video.

Notice that this article did not give you a checklist of tools to memorize. It gave you a principle, and used the tools as proof of that principle.

If you walk away with a single sentence, let it be the same one from the first article, just translated out of the world of code:

Stop treating AI as an oracle you ask questions to. Start treating it as something that works on top of your own material.

The developer does this with a CLAUDE.md file. The lawyer does it with their contracts. The people manager with their policies. The accountant with their data. The marketer with their brand assets. The trainer with their handbook. The personal trainer with their methodology. The tool changes; the habit is the same. And it is that habit, not technical fluency, that will separate those who, in 2026, use AI to their advantage from those who keep pasting things into a chat and hoping for the best. The noise around AI is not going to slow down, but you now have a filter to cut through it: does this work from my context, or does it only give me an average of the internet? If it is the first, it is worth your time. If it is the second, probably not.

This is the second of two articles. The first one, focused on developers and Claude Code, can be read here.

This article was written by our alumni Anderson Leite; you can find the original article here!