Myth one: AI can't handle rands, local payment terms, or Takealot's style
The belief: AI models are trained on dollars and American English, so they'll write "shipping" instead of "delivery," quote prices in USD, and ignore that South Africans pay by EFT or SnapScan.
What actually happens: Every major model in 2025 understands currency context if you tell it once. I've written hundreds of Takealot product descriptions using Claude Sonnet 4.5 and Gemini Flash, and not one has defaulted to dollars when the prompt says "price in rands."
The trick is specificity. Don't write "describe this product." Write: "Write a Takealot product description for a 500ml stainless steel water bottle. Price: R149. Highlight free delivery over R450. Mention EFT and card payments. Keep it under 150 words."
Claude and Gemini both nail it. DeepSeek V3 does too, and it's faster. I tested this with a Johannesburg Shopify seller who runs a homeware store—she was rewriting every AI draft because she thought the models "didn't get South Africa." We added three lines to her prompt template: currency, payment methods, delivery threshold. Her rewrite time dropped from twelve minutes per product to two.
The models don't need expensive fine-tuning or a South African agency to "localize" them. They need clear instructions. If you're still paying someone R800 per product description because "AI doesn't understand Takealot," you're solving a problem that hasn't existed since 2023.
Myth two: auto-routing wastes tokens by testing multiple models
The belief: Auto-routing sends your prompt to several models at once to pick the best one, so you burn through tokens and rack up costs.
Reality: Auto-routing sends your prompt to one model. It just picks the right one based on your task.
I've used Kryotta's auto-routing for three months on everything from Takealot bulk uploads to WhatsApp customer replies during load-shedding. It routes a short product FAQ to Gemini Flash Lite (cheap, fast) and a detailed brand story to Claude Sonnet 4.5 (more expensive, better at tone). I don't choose. The system does.
A Cape Town seller I know was manually switching models for every task—Gemini for SEO descriptions, Claude for email campaigns, Llama for simple replies. She spent more time deciding which model to use than actually writing. Auto-routing cut her decision time to zero and her monthly AI cost by 30%, because she stopped defaulting to the most expensive model for jobs that didn't need it.
The myth probably comes from Compare mode, which does run your prompt through multiple models so you can see the outputs side by side. That's a different feature. Compare is for testing when you don't know which model fits your style. Auto-routing is for production work when you just want the best result without thinking.
If you're avoiding auto-routing because you think it's wasteful, you're wasting time instead.
Myth three: you need an agency to set up AI workflows for Shopify
The belief: Integrating AI with Shopify or Takealot requires a developer, a Zapier expert, or a R15,000 agency retainer.
What it actually takes: A Gmail or Shopify connector and fifteen minutes.
I've watched small online sellers in Johannesburg pay agencies to "build an AI content pipeline" when all they needed was a tool that plugs into their existing Shopify admin. Kryotta's Shopify connector pulls product data directly into the workspace—you write one prompt template, generate descriptions for twenty products, and push them back to your store. No code. No Zapier. No monthly retainer.
A Takealot seller I spoke to was paying a freelancer R500 per batch of ten product descriptions. She switched to Gemini Pro with a saved prompt template and now generates fifty descriptions in one sitting for under R20 in token costs. She uses Compare mode once a month to test whether Claude or Gemini writes better for her category (outdoor gear), then sticks with the winner.
The "you need an agency" myth survives because agencies benefit from it. If you're running a Shopify or Takealot store and you're still outsourcing product descriptions, customer email templates, or image alt text, try doing it yourself for one week. You'll either realize you don't need help, or you'll discover the specific task you actually need a human for (usually brand strategy, not execution).
Myth four: AI image tools can't handle product photography edits
The belief: AI is fine for generating abstract art, but if you need to remove a background from a product photo, fix lighting, or resize for Takealot's specs, you still need Photoshop or a designer.
Reality: Stable Image Core and SD 3.5 Large handle background removal, lighting adjustments, and resizing faster than most freelancers, and they don't charge per image.
I tested this with a Johannesburg jewellery seller who was paying a designer R150 per product photo to remove backgrounds and adjust white balance. She had 80 new products to shoot. That's R12,000. We ran the same photos through Stable Image Core in Kryotta's image editor—background removal took four seconds per image, and the lighting fix was one slider adjustment. Total cost: under R100 in credits.
The myth persists because people think "AI image tools" means Midjourney-style generation, not editing. But the same models that generate images also manipulate them. If you're still outsourcing basic edits—background removal, colour correction, cropping to platform specs—you're paying for something you can do in the time it takes to upload the file.
One caveat: if you need complex compositing (putting a product in a lifestyle scene that doesn't exist), a human still does it better. But for the 90% of edits that are "make this white background whiter" or "remove the shadow," AI is faster and cheaper.
Myth five: AI stops working when load-shedding hits
The belief: AI tools are cloud-based, so when Eskom cuts power and your router dies, you lose access and your workflow stops.
What actually happens: Your phone still has signal. WhatsApp Business still works. And if you've saved your prompts, you can generate product descriptions, customer replies, and social captions from your phone while your laptop is off.
I know a Shopify seller in Cape Town who runs her entire store from her phone during Stage 4. She uses Kryotta's mobile interface to generate Takealot descriptions, reply to customer questions, and edit product images. She's not waiting for power to come back. She's working around it.
The trick is preparation. Before load-shedding, she saves her five most-used prompt templates: product descriptions, WhatsApp replies for "Is this in stock?", "Do you deliver to Durban?", "What's your return policy?", Instagram captions, and email subject lines. When the lights go out, she opens the app on her phone, picks a template, fills in the product details, and generates. She's not typing long prompts on a phone keyboard. She's tapping three fields and hitting go.
The myth survives because people assume "AI workflow" means sitting at a desk with a laptop and a second monitor. It doesn't. If you're still telling yourself you'll "do the AI stuff later when the power's back," you're losing four hours a day that you could spend selling.
Questions people ask
Can AI really write product descriptions that rank on Takealot?
Yes, if you give it the keywords. Takealot's search works like any e-commerce platform—it looks for specific terms in your title and description. Tell the model "include 'stainless steel water bottle', 'BPA-free', and 'insulated' in the first 50 words," and it will. I've seen AI-written descriptions rank in the top three results for competitive categories.
Which model is cheapest for bulk product descriptions?
Gemini Flash Lite or DeepSeek V3. Flash Lite is faster; DeepSeek is better at following complex instructions. For 100 Takealot descriptions, you're looking at under R50 in token costs with either model. Claude Sonnet costs more but writes better brand voice if that matters for your category.
Does auto-routing work offline or during load-shedding?
No. Auto-routing is cloud-based, so you need internet. But if you're working from your phone on mobile data, it still works. Save your templates before load-shedding, and you can generate from your phone while your laptop is off.
Can I use AI for WhatsApp customer service replies?
Yes, and it's one of the best uses. Create templates for your ten most common questions ("Do you deliver to Pretoria?", "Is this still in stock?"), save them, and generate replies in seconds. You're not copy-pasting the same answer 40 times a day. You're tweaking an AI draft that's already 90% correct.
If you're still outsourcing product descriptions, paying agencies for "AI strategy," or waiting for load-shedding to end before you write anything, you're spending time and rands on problems that don't exist anymore. The tools work. The models understand rands and Takealot and EFT. You just have to use them. Try Kryotta's Compare and auto-routing features and see which model actually fits your store—it takes one prompt, not a consultant.
