The Tuesday morning I tried to explain "tokens" and watched five faces glaze over
We were sitting in the back room of my Lekki store—three sales assistants, one logistics coordinator, one social media person—and I'd just told them we were going to start using AI to write product descriptions for the 200 items we were listing on Jumia and our Instagram shop. I'd opened Kryotta on the projector, typed a prompt into Claude, and watched it spit out three paragraphs about a lace blouse in four seconds.
"How does it know what to write?" someone asked.
"It reads your prompt and generates a response based on patterns it learned during training," I said, already regretting the jargon.
"But how much can you type before it stops understanding you?"
Good question. I started explaining tokens and context windows. Two minutes in, I could see I'd lost them. So I tried again, this time using examples they already knew: WhatsApp character limits, how a conversation gets too long to scroll back through, why sometimes a customer asks you a question and you've forgotten what they said ten messages ago. That worked. By the end of the session, everyone understood enough to actually use the tools without breaking them.
Here's the explanation that clicked, stripped of the jargon.
Tokens: not words, but close enough for planning
A token is the smallest unit of text an AI model processes. It's not exactly a word—"running" might be one token, but "unbelievable" might be split into two or three—but for practical purposes, you can estimate that 100 words equals about 130 tokens in English. In a mix of English and pidgin or when you're writing product descriptions with prices and measurements, the ratio shifts a bit, but the rule holds.
Why does this matter? Because every AI model has a limit on how many tokens it can read and write in one go. Claude Sonnet 4.5 can handle 200,000 tokens of input. Gemini Flash Lite handles 1 million tokens. DeepSeek V3 handles 64,000. If you try to paste your entire Jumia product catalogue—800 items, 50 words each, plus prices and specs—into a model with a 4,000-token limit, it'll cut off halfway through and you'll get nonsense.
I tested this by accident. I copied 150 product descriptions into Gemini Flash and asked it to rewrite them all in a more conversational tone. It worked. Then I tried the same task with GPT-OSS 20B, which has a smaller context window. It rewrote the first 40 products beautifully, then started repeating itself, then gave me half-finished sentences. I'd hit the limit.
The fix: break the job into batches. I now paste 30 products at a time, let the model finish, then move to the next batch. Takes longer, but the output is usable.
Context window: how much the model remembers in one conversation
The context window is the total amount of text—your prompts plus the model's replies—that the AI can keep in memory during a conversation. Think of it like a WhatsApp chat that only shows the last 50 messages. If you scroll up past message 51, the model has no idea what you said.
This hit me hard when I was using Claude to draft responses to customer complaints on Instagram. A woman DM'd me saying a dress didn't fit, asked for a refund, then sent five more messages over two days: her bank details, a photo of the dress, a question about exchange rates because she'd paid through Paystack in dollars, a follow-up asking why I hadn't replied yet. I copied the entire conversation into Claude and asked it to draft a response.
Claude wrote a polite reply offering a refund, but it referenced the wrong bank and ignored the exchange-rate question entirely. I scrolled up and realised I'd pasted 60 messages—about 8,000 tokens—and Claude had forgotten everything before message 30. The context window had moved forward, and the early details fell off the back.
The fix: summarise old messages before you add new ones. Now when a conversation gets long, I paste the first 20 messages, ask Claude to summarise them in three sentences, then paste that summary plus the recent messages. It works. The model keeps the important details without choking on the full history.
Hallucinations: when the model invents facts that sound real
A hallucination is when an AI generates something false but presents it as fact. It's not lying—it doesn't know it's wrong—it's just predicting the next token based on patterns, and sometimes the pattern leads somewhere untrue.
I asked Llama 3.3 70B to write a blog post about the best payment processors for Nigerian online sellers. It gave me Paystack, Flutterwave, and then invented one called "NaijaPay" that supposedly had lower fees than both. It included a fake website, fake pricing (₦50 per transaction, 0.5% fee), and a paragraph about how NaijaPay was "preferred by over 10,000 Lagos vendors." None of it was real.
I only caught it because I know the market. If I'd been writing about a topic I didn't understand—say, logistics software for cold-chain shipping—I might have published the fake facts without checking.
The fix: verify anything that sounds specific. Model names, prices, website URLs, statistics—if the AI gives you a number or a name, Google it. And never trust an AI to cite sources unless you've fed it those sources yourself.
Grounding: giving the model real information so it stops guessing
Grounding means anchoring the AI's output to specific, verified information you provide. Instead of asking "What are the delivery times for Lekki?" and letting the model guess, you paste your actual delivery schedule and say "Based on this schedule, write a customer-facing FAQ about delivery times."
I grounded Claude when I needed to update our Instagram bio and highlights. I pasted our current prices, delivery zones, payment methods, and bank details—about 400 words of facts—then asked it to rewrite the bio and draft captions for five highlights. Every fact in the output matched what I'd given it. No hallucinations, no invented prices, no made-up courier services.
Grounding works especially well for product descriptions. I paste the specs (fabric, size chart, care instructions, price) and let the model write the marketing copy. It can't invent a colour or size that doesn't exist because the source material constrains it.
The trade-off: grounding takes more setup time. You have to gather the facts first. But for anything customer-facing—prices, delivery terms, return policies—it's worth it.
Open-weight vs closed models: who can see the code, and why it matters for cost
A closed model (Claude, Gemini, GPT-OSS) is one where the company that built it keeps the underlying code private. You access it through an API or a web interface, you pay per token or per month, and you have no idea how it works inside. An open-weight model (Llama, Mistral, DeepSeek) is one where the model's architecture and parameters are public. Anyone can download it, run it on their own server, modify it, or offer it as a service.
For most Nigerian small businesses, this distinction matters for two reasons: cost and privacy.
Closed models are often more expensive. Claude Sonnet 4.5 costs a few cents per thousand tokens; Gemini Flash is cheaper; GPT-OSS sits in the middle. If you're processing 50,000 tokens a day—writing product descriptions, replying to DMs, drafting emails—that adds up. Open-weight models like Llama 3.3 70B or DeepSeek V3, when accessed through a platform like Kryotta, can be cheaper per token because the hosting costs are split across many users.
Privacy is the other factor. With a closed model, your prompts and outputs pass through the provider's servers. For most e-commerce tasks—writing a product description, drafting a social media caption—that's fine. But if you're pasting customer bank details, transaction records, or anything sensitive, you want to know where that data goes. Open-weight models, especially if you're running them locally or through a provider with clear data policies, give you more control.
I use closed models (Claude, Gemini) for customer-facing content and open-weight models (DeepSeek, Llama) for internal tasks like summarising sales reports or drafting supplier emails. The quality is similar; the cost and privacy profile differ.
Questions people ask
If I hit the token limit halfway through a task, can I just start a new chat and continue?
Yes, but the new chat won't remember the old one. You'll need to paste a summary of what you'd already done, or the model will start fresh and might contradict itself. I keep a running document with key details (prices, product names, customer requests) and paste it into each new chat.
How do I know if a model is hallucinating or just giving me a different opinion?
If it's a fact—a price, a name, a date, a statistic—you can verify it. If it's an opinion or a suggestion ("I recommend using Instagram Reels for this campaign"), that's not a hallucination, it's advice. You can disagree with advice; you can't disagree with a fact that's wrong.
Which model has the longest context window, and should I always use that one?
Gemini Flash Lite handles 1 million tokens, which is huge. But a longer context window doesn't always mean better output. For short tasks—writing one product description, drafting one email—a smaller, faster model like Claude Haiku 4.5 or Gemini Flash works fine and costs less. Use the big context window when you're actually working with a lot of text.
Can I use AI to reply to customer DMs without checking every response?
Not safely. I draft replies with AI, read them, fix anything that sounds off or wrong, then send. It takes half the time of writing from scratch, but I never let a message go out unread. One wrong price or invented delivery promise will cost you more than the time you saved.
I still train my team on this every few months, especially when we hire someone new. The jargon changes, the models get faster, but the core ideas—tokens, context, grounding, knowing when the AI is guessing—stay the same. If you're running a Nigerian business and trying to figure out which AI tools are worth the money, start by understanding these five concepts. Everything else is just features. You can try the models I mentioned, compare their output on your own tasks, and see which ones fit your workflow at Kryotta.
