The moment I realized my team was trusting AI to quote prices
I walked into the shop one Tuesday morning and found my assistant, Chidi, on WhatsApp with a customer, copying and pasting a product description he'd asked Claude to write. The description was confident, specific, and completely wrong. It listed our USB-C cable as "compatible with all iPhone models" and quoted a bulk discount structure we'd never offered. The customer had already sent a bank transfer request for twenty units at the invented price.
I'd been using AI for weeks to draft replies, write product copy, and summarize long customer threads. I'd told the team "this will save you time." I had not explained that the model sometimes makes things up with absolute confidence, and I definitely hadn't trained them on when to double-check the output. That ₦47,000 pricing error taught me I needed a way to explain hallucinations that didn't require anyone to understand how a language model works.
Over the next month I tried five different explanations with my three-person team—two handle WhatsApp sales, one manages our Instagram shop. They're all smart, none are technical, and all of them now use AI daily to answer questions about our phone accessories and power banks. One explanation worked. The other four got polite nods and then the same mistakes kept happening.
1. The technical explanation: "It predicts the next word based on patterns"
I started where every AI article starts. I explained that the model doesn't know things, it predicts the next word based on patterns it saw in training data, and sometimes those predictions are plausible but wrong. I used the phrase "statistical likelihood" twice. I could see their eyes glaze over.
The problem: this explanation is accurate but doesn't tell someone when to worry. Chidi nodded, said "okay, I understand," and the next day asked Gemini Flash to write a reply promising same-day delivery to Abuja. We don't deliver to Abuja at all.
Why it failed: it's an explanation of how, not a decision rule. If you don't know what a prediction model is, being told the AI uses one doesn't help you spot the moment it's guessing about your stock levels.
2. The "confident liar" framing: "It makes things up but sounds sure"
I tried reframing it as a personality flaw. "The AI is like someone who hates saying 'I don't know,' so when you ask it a question it doesn't have the answer to, it invents something that sounds right. It won't warn you. It'll just state the fake answer like it's fact."
This got more traction. Blessing, who runs our Instagram account, started saying "wait, did we actually say that?" before posting AI-written captions. But she still couldn't predict which statements were likely to be invented. She'd check the obvious stuff—prices, product names—and miss the subtle errors, like when Claude Sonnet wrote that our power bank "supports fast charging for all Android devices." It supports fast charging for devices with USB Power Delivery. Not the same thing, and the customer who bought it for a three-year-old Samsung was annoyed.
Why it half-worked: it made people suspicious, but not systematically careful. They'd check some things and assume the rest was fine.
3. The Jumia listing test: "Ask it something you can verify on Jumia right now"
I got more specific. I opened Jumia on my phone, found one of our competitor's listings for a phone case, and asked Gemini Pro to "write a product description for this case." Then I compared the output to the actual listing. The model got the material right, invented a "scratch-resistant coating" that wasn't mentioned anywhere, and claimed the case was "rated 4.7 stars by over 300 buyers" when the listing had sixty-eight reviews.
I showed this to the team and said: "This is what hallucination looks like. The model fills in details that sound like they belong, even when it has no data. If you're writing about our products, it's doing the same thing with our stock, our prices, our policies."
This was better. Tunde, my other WhatsApp assistant, started opening our inventory spreadsheet before he'd paste an AI reply that mentioned quantities or lead times. But he still trusted the model on anything that sounded like general knowledge—shipping times, payment methods, how Paystack transfers work—and that's where the next mistake happened.
Why it mostly worked: it gave them a concrete example of the kind of error to expect, but only for product details. They didn't generalize the lesson to policies or processes.
4. The two-bucket rule: "Only let it write about what you've told it"
This is the explanation that finally worked, and I wish I'd started here. I drew two columns on the whiteboard: Grounded and Guessing.
Grounded (safe to use):
- You've pasted our price list into the chat
- You've given it the customer's previous messages
- You've copied the exact product spec from our supplier
- You've told it our current delivery schedule
Guessing (double-check every word):
- You asked it to describe a product without giving it the details
- You asked it general questions ("how long does Paystack take to process transfers?")
- You asked it about a policy we haven't spelled out in the chat
- You asked it to compare our product to a competitor's
I told them: "If the information came from you in this conversation, the AI can work with it. If you're asking it to know something you didn't tell it, it's guessing. Confident, plausible guessing—but still guessing."
Blessing immediately said, "Oh, so I need to paste the product details first, then ask it to write the caption." Yes. Exactly.
Why it worked: it gave them a simple decision tree. Before you trust the output, ask yourself: did I give it the facts, or am I hoping it already knows them?
5. The one-sentence version I now use for new hires: "It remixes what you tell it; it doesn't look things up"
After a month of using the two-bucket rule, I distilled it further. When I hired someone new in March, I said: "The AI doesn't have access to our inventory, our prices, or our policies. It can only work with what you paste into the chat. If you want it to write a reply about a product, paste the product details first. If you want it to quote a price, paste the price list. It remixes what you give it—it doesn't look things up."
She got it in one conversation. No follow-up mistakes.
The longer explanation is still useful for training people on why this matters, but the one-sentence version is what they remember when they're busy and a customer is waiting.
What "grounding" actually means when you're selling on WhatsApp
Grounding is just the formal term for "giving the AI the facts before you ask it to write." In Kryotta I'll paste our current stock levels into the chat, then ask Claude Sonnet 4.5 to draft a reply to a customer asking about availability. The model writes: "We have twelve units of the 10,000mAh power bank in black, and five in white. I can reserve one for you if you'd like to pay by bank transfer today." That's grounded—it's working from the numbers I gave it.
If I'd just asked "write a reply about power bank availability" without the context, it would invent plausible-sounding stock levels and I'd be correcting the customer an hour later.
The technical explanation is that grounding means providing retrieved or verified information that the model references when generating output, which reduces hallucination by anchoring the response in real data. The practical explanation is: paste the facts first, then ask for the reply.
The checklist I printed and stuck to the wall
After the two-bucket conversation, I made a one-page checklist. It's taped next to the desk where we handle WhatsApp orders:
Before you send an AI-written reply:
☐ Did I paste the product details, price, or policy into the chat?
☐ Is the AI quoting something I told it, or something it "knows"?
☐ If it mentions stock, delivery time, or payment terms, did I give it those facts?
☐ If I'm not sure, can I verify the claim in five seconds? (Check the spreadsheet, check Jumia, check our Paystack dashboard.)
If you answer no to any of those, rewrite that part by hand.
It's not a perfect system. Tunde still occasionally forgets and asks Gemini to "write a reply about our return policy" without pasting the policy, and the model invents a fourteen-day window we don't offer. But mistakes dropped from two or three per week to one every two or three weeks, and no customer has received a wrong price since February.
What changed after my team understood hallucinations
We're faster. Chidi drafts twenty WhatsApp replies in the time it used to take him to write six, because he's not agonizing over wording—he's pasting context and letting Claude handle tone. Blessing posts Instagram captions daily instead of twice a week, because she's not staring at a blank screen. Tunde closes sales calls with order confirmations that used to take him ten minutes to type.
But we're also more careful in a specific way. Everyone knows the moment they need to stop and check: when the AI is writing about something they didn't explicitly tell it. That's the shift. They're not scared of the tool, and they're not blindly trusting it. They've got a mental model of when it's reliable (remixing their input) and when it's risky (filling in gaps with guesses).
The cost in mistakes was real—₦47,000 on that first pricing error, another ₦12,000 when we promised a delivery date we couldn't meet, and several annoyed customers who had to be walked back from incorrect product claims. The cost in time to fix the training was about two hours of explanation and one hour making the checklist. I'd pay that again.
Questions people ask
Q: Which AI model hallucinates least when writing product descriptions?
Claude Sonnet 4.5 and Gemini Pro are both solid if you give them grounding context—paste the product spec, then ask for the description. Gemini Flash is faster and cheaper but more likely to add unsupported claims if you're vague in the prompt. I've had good results with Llama 3.3 70B for straightforward listings. The model matters less than whether you've given it the facts to work from.
Q: Can I just tell the AI "don't make things up" in the prompt?
I've tried. It doesn't work reliably. You'll get output that sounds more cautious—lots of "may" and "typically"—but it still invents details when it doesn't have them. The only method that consistently works is giving it the information first, in the same chat, right before you ask it to write.
Q: How do I explain this to someone who's never used AI before?
Start with "it remixes what you tell it; it doesn't look things up." Then show them one example of a hallucination—ask the model a question you know the answer to, without giving it context, and point out where it guessed. That's usually enough. The two-bucket rule (grounded vs. guessing) is for people who are already using AI daily and need a decision framework.
Q: Is grounding the same as fact-checking?
No. Grounding means giving the AI verified information before it writes. Fact-checking means reviewing the output after. You need both. Grounding reduces the chance of hallucinations; fact-checking catches the ones that slip through. My team does both: they paste context (grounding), then they check any claim that could cost us money or trust (fact-checking).
If you're training a team to use AI for WhatsApp sales, Instagram posts, or customer service, the two-bucket rule is the clearest place to start. It's not about understanding transformers or training data—it's about knowing when you've given the model enough to work with, and when you're hoping it knows something it doesn't. Try it with one person for a week, watch where they still make mistakes, then adjust the checklist. You can set this up in Kryotta and test it with your own product list this afternoon.
