Forget the productivity-revolution headlines. Here is what AI is actually good at right now for someone running a small business — and what it is still bad at.
Almost every conversation about AI right now is either "this will change everything overnight" or "this is all overhyped". Both miss the actual situation. AI in 2026 is genuinely useful for specific things, genuinely overrated for others, and the gap between the two is wider than the headlines admit.
This post is the practical version: what AI is actually doing in our day-to-day, and what we still do the old way.
What it is good at
Drafting
Anthropic's Claude, OpenAI's ChatGPT, Google's Gemini — all of them produce a plausible first draft of almost anything in under thirty seconds. Email replies, blog post outlines, contract clauses, project briefs, landing-page copy. The quality of the first draft is mediocre; the time savings are huge. You spend ten minutes editing instead of forty minutes staring at a blank page.
Translation and summarisation
Going from a Finnish source document to a Swedish-and-English version of the same content used to be a real cost line. With AI assistance plus a human reviewer who speaks both languages, it's now a fraction of the cost. We use this on multilingual sites — see our accessibility page for why we still have a human in the loop.
Code
For routine code (config files, glue scripts, small refactors), AI assistants like GitHub Copilot or Cursor genuinely save time. For new features in a complex codebase, they're useful as a junior pair-programmer who hallucinates occasionally — useful, but not autonomous.
Bulk customer service triage
For a small business with a steady stream of customer questions, an AI tool can categorise incoming emails, suggest draft replies, and flag the genuinely complicated ones for a human. Tools like Intercom Fin and Zendesk's AI agents target this directly. The honest version is "first draft, human reviews and sends".
Data analysis on small datasets
Drag a CSV of last year's bookings into Claude or ChatGPT, ask "what's our seasonality?", get a workable analysis in under a minute. Five years ago that was a 3-hour task in Excel. Today it's a one-line prompt.
What it is still bad at
Anything with a long-term memory
LLMs forget your context the moment you start a new chat. Tools like Claude Projects and ChatGPT custom GPTs help, but they're not the same as a coworker who remembers what you decided three months ago.
Strategic judgement
AI can generate ten possible business strategies in five seconds. It cannot tell you which one fits your specific situation, your specific customers, or your specific risk tolerance. That's still you, and the tools that help you think aren't AI tools — they're the customer-discovery work we wrote about earlier.
Anything where being wrong is expensive
Don't generate legal contracts. Don't generate medical advice. Don't generate financial recommendations. The model will produce something confident and sometimes plausibly wrong, and the cost of that being wrong is enormous. This is the Anthropic Responsible Scaling Policy in plain English: capability outpaced reliability for a while; verify everything that matters.
Original creative voice
The drafts AI produces sound like AI. Recognisably. It's a bland average of what's been written about a topic. Real voice — the way someone actually thinks and talks — is still a human thing, and probably will be for a long time.
How we use it at The Long Way
Concretely:
- Code completion in the editor (Copilot-style).
- First drafts of long-form content like this post — then heavily rewritten by a human.
- Translation drafts, then reviewed by a human bilingual editor.
- Email triage on inbox volume.
- Quick data exploration on customer questionnaire submissions.
Things we don't use it for:
- The first conversation with a new client. That's a human call.
- Strategic recommendations to clients. That's our judgement.
- Final copy that goes on a client's site without their review. That's a recipe for embarrassment.
What to read
- The Pragmatic Programmer's AI section — practical, not hype.
- Simon Willison's blog — the most reliable running commentary on what these tools can and can't actually do.
- Anthropic Cookbook — if you want to start scripting against the API directly.
For a follow-up on which AI tools to use for which business role, see our companion post. And if you want to put any of this into a website that runs your business, start with the questionnaire.