Your First AI Setup: The Stack That Actually Works in 2026

The Problem With Most AI Advice

Most "AI stack" articles are written by people who've been immersed in this world for three years and assume you have been too. They reference API endpoints, token windows, and model benchmarks the way a car enthusiast describes engine displacement — fluently, and for an audience that already knows what it means. This article isn't that.

The premise here is different: you want to get real work done with AI. You have a limited budget. You don't want to spend your evenings experimenting with tools that sound impressive at a conference but don't actually change what you produce on a Tuesday morning. This guide is built around a single principle — start simple, go deep before you go wide, and measure what actually changes. Here's where to start.

Layer 1: The Free Foundation

The free tiers of the major AI platforms are genuinely capable. In 2026, "free" no longer means a hobbled, watermarked trial — it means access to models that would have seemed extraordinary three years ago. Here's what each offers:

• ChatGPT free tier: Solid general-purpose capability. Best for writing, brainstorming, summarisation, and answering broad questions. The most familiar interface for most people, which matters for adoption.
• Claude.ai free tier: Noticeably stronger for long documents, careful reasoning, and tasks that require holding a lot of context at once. If you regularly work with reports, legal documents, or research, start here.
• Google Gemini free tier: The obvious choice if your work lives in Google Workspace. Native integration with Docs, Sheets, and Gmail means it can read your actual work rather than requiring you to copy-paste everything.

The most important instruction at this stage: pick one, not all three. The temptation is to try everything simultaneously, which results in shallow familiarity with all of them and genuine proficiency with none. Decision guide: if your daily work runs through Google Workspace, start with Gemini. If you write a lot — emails, reports, proposals — start with Claude. If you need a capable generalist for varied tasks, start with ChatGPT. Commit to that choice for at least two weeks before evaluating anything else.

Layer 2: The First Paid Upgrade (£15–20/month)

You'll know you're ready for a paid tier when one of three things happens: responses start feeling noticeably slower during peak hours; you hit context limits mid-task and have to restart; or you're spending more than two hours a week on tasks that AI could handle but keeps truncating or missing. Any of these signals mean the free tier is now a bottleneck, not a feature.

The first upgrade is straightforward: ChatGPT Plus or Claude Pro, both around $20/month (approximately £16). Both unlock access to the most capable models, significantly longer context windows — meaning you can paste in an entire 50-page document and ask questions about it — and much faster responses during busy periods. The math is simple: if the tool saves you two hours a week and your time is worth anything, the subscription pays for itself in the first week.

What you're not paying for at this tier is automation, integrations, or team features. That comes later, and only if you need it. The paid tier upgrade is purely about getting the most capable version of the tool you've already learned to use.

Layer 3: BYOK — Bring Your Own Key

There's a smarter alternative to subscription bloat that most people haven't heard of. BYOK — Bring Your Own Key — means connecting your own API key directly to a platform instead of paying for a subscription. You pay only for what you actually use, billed by the token rather than by the month. For moderate users who might go weeks without heavy AI use, this typically costs 60–80% less than a subscription.

Getting an API key is simpler than it sounds: visit platform.openai.com, navigate to the API section, and click Create Key. The whole process takes five minutes. Once you have the key, platforms like Aether Intel's Playground let you plug it in directly and use the same underlying models without the monthly fee. The catch is minor: you need to manage your own usage and keep an eye on your balance. For anyone who uses AI intermittently — heavy some weeks, barely at all others — BYOK often makes more financial sense than a fixed subscription that keeps charging during quiet months.

The Automation Layer

Once you have a base model working well, automation is what multiplies its value. A standalone AI tool that you have to visit manually is useful. An AI tool wired into the tools you already use is transformative. The good news: you don't need to write any code to get there.

Zapier or Make (both have usable free tiers) let you connect AI models to almost any app you use. A practical example that takes thirty minutes to set up: any email that arrives from a specific sender gets automatically summarised by AI and dropped as a message in Slack. No inbox triage required. Similar workflows work for processing form submissions, summarising meeting transcripts, or drafting responses to common customer enquiries.

Notion AI is worth considering if you already use Notion for notes or project management. Writing, summarising, and reformatting content inside your existing knowledge base removes the friction of switching between tools. Browser extensions like Sider add a persistent AI sidebar to any webpage: right-click any article, contract, or product page to summarise it, translate it, or ask specific questions without opening a separate tab. For research-heavy work, this alone saves meaningful time every day.

Your First Week Checklist

Theory is useful. A specific plan is more useful. Here's exactly what to do in your first seven days:

1. Pick one model — ChatGPT, Claude, or Gemini — and use only that for the full seven days. No hedging, no switching.
2. Identify the three tasks you spend the most time on that involve writing or research. Not everything, just the top three.
3. Spend thirty minutes on day one prompt-testing each of those three tasks. Write three or four different versions of each prompt and compare the outputs. Find the phrasing that works.
4. Save your best prompts in a plain notes document. This is the beginning of your personal prompt library, and it compounds in value over time.
5. On day seven: measure actual time saved. If you've saved less than one hour across the week, change how you're prompting — be more specific, provide more context, define the format you want. Don't change tools. The issue is almost never the tool.