The load-shedding problem that made me test AI on real orders
I own a small restaurant in Observatory, Cape Town—twenty-four seats, open for lunch and dinner, takeaway and delivery through our own WhatsApp line. We're not on Uber Eats. We don't have a call centre. When Stage 4 load-shedding hit us during the dinner rush in March, the phone was useless, the router was dead, and I had twelve WhatsApp orders stacked up on my mobile data. I could cook or I could type. Not both.
I'd been experimenting with Kryotta's AI models to draft Instagram captions and menu descriptions, but I'd never let one touch a live customer conversation. That night I had no choice. I set up a simple workflow: every WhatsApp order query got copied into Kryotta, I fed it to Claude Sonnet 4.5 with a short prompt ("You're a friendly restaurant assistant. Answer this clearly and fast"), and I sent the reply with minimal editing. Over the next two weeks I handled forty order queries this way—some during load-shedding, some during normal hours so I could compare. I tracked which tasks the AI closed cleanly, which ones I had to rescue, and what each conversation cost me in tokens and time.
One category of query saved me six minutes per response and never needed a follow-up. Another nearly lost me a regular customer. Here's the checklist I built, the models I tested, and the cost in rands per query.
Tasks AI handles alone: menu questions, bookings, delivery zones
These are the queries I now trust AI to finish without me reading every word before I send.
Menu and ingredient questions. "Do you have vegan options?" "Is the lamb halal?" "What's in the butternut soup?" Claude Sonnet 4.5 answered twenty-two of these across the forty queries. I gave it our full menu as context—two pages, prices in rands, allergen notes included—and it pulled the right details every time. One customer asked if the chicken pie contained mushrooms. Claude said no, listed the actual ingredients, and suggested the vegetable curry as a mushroom-free alternative. The customer ordered both. Average response time: forty seconds. Cost per query: R0.08 to R0.12 depending on how much menu context I included.
Table bookings for standard times. "Table for four on Friday at 7 p.m.?" "Can we book Saturday lunch?" I tested this with Gemini Flash because it's faster and cheaper for simple structured tasks. I fed it our booking rules (no reservations for fewer than three people, kitchen closes at 9.30 p.m., we're shut Mondays) and our live availability from a shared Google Sheet. Flash confirmed eight bookings out of ten correctly. The two it missed were edge cases—one wanted a 10 p.m. seating, one asked for a table "around sunset" without naming a day—and I caught both before sending because I was still nervous. Once I trusted the pattern, I let it run. Cost per booking confirmation: R0.03 to R0.05.
Delivery zones and ETAs. "Do you deliver to Rondebosch?" "How long for a delivery to Woodstock?" I used Claude Haiku 4.5 for these because they're repetitive and I wanted the cost down. I gave it a list of suburbs we cover, a flat R40 delivery fee, and typical travel times. Haiku handled nine delivery questions cleanly. One customer asked if we deliver to Hout Bay. Haiku said no, explained our fifteen-kilometre radius, and suggested collection instead. The customer collected. Cost per query: R0.02.
These three categories accounted for thirty-one of the forty queries. I spent R2.80 in total and saved roughly three hours I would have spent typing the same answers on my phone during prep.
Tasks that need a human safety net: complaints, dietary restrictions, payment confusion
These are the queries where AI gave me a starting draft but I had to rewrite or escalate.
Complaints about food or service. Two complaints arrived in the test period. One said the lamb curry was too salty. One said delivery took ninety minutes and the food was cold. I tried Claude Sonnet 4.5 on both, asking it to apologize and offer a solution. The apology tone was fine—warm, not robotic—but the solution it suggested (a 20 per cent discount on the next order) wasn't our policy. We offer a full refund or a replacement meal, no halfway discount. I caught it before sending, rewrote the offer, and both customers accepted the refund via EFT. The AI saved me two minutes on the apology paragraph, but I wouldn't send a complaint reply unedited. Cost per draft: R0.15. Time saved: minimal.
Custom dietary requests. Three queries asked for substitutions or allergen swaps. "Can I get the chicken burger without the bun?" "Is there a way to make the pasta dairy-free?" "I'm allergic to sesame—what's safe?" Claude gave plausible answers, but two were wrong. It said we could swap the pasta cream sauce for olive oil (we can, but it also has parmesan in the pasta itself, which the AI missed). It said the burger bun was the only gluten in that dish (the chicken is dusted in flour). I had to check the prep sheets and rewrite all three answers. The AI got me halfway there, but these need a human who knows the kitchen. Cost per attempt: R0.10. Risk: high if you don't double-check.
Payment and order status questions. Four customers asked "Did my payment go through?" or "Where's my order?" after paying via SnapScan or EFT. AI can't check your bank account or your delivery driver's location. I tried giving Claude access to a text file I update with order statuses, but it still couldn't answer "Did you receive my R340 EFT?"—it just said "Let me check" and waited. I had to look up the payment myself and reply manually. These queries are simple, but they need live data AI doesn't have unless you connect it to Yoco or your bank API. I haven't done that yet. Cost per failed attempt: R0.06. Time wasted: three minutes per query while I figured out the AI couldn't help.
Which models I'd use for which task, and why
I tested Claude Sonnet 4.5, Claude Haiku 4.5, Gemini Flash, and Gemini Pro across the forty queries. Here's what I'd assign to each.
Claude Sonnet 4.5 for anything that needs tone and judgement—menu questions where the customer's asking in a roundabout way ("What's your lightest meal?"), booking requests that aren't straightforward ("We're a group of six but two are kids—do we need a reservation?"), or any query where you want the reply to sound like a person who cares. It's the most expensive of the four (R0.08 to R0.15 per query depending on length), but it's the only one I'd trust to write something I don't edit. I used it for twenty-two queries and only rewrote two.
Claude Haiku 4.5 for high-volume, low-stakes repetition—delivery zones, opening hours, "Do you take card?" It's fast, it's cheap (R0.02 to R0.05), and it doesn't try to be clever. I used it for nine queries and all nine were fine. If you're handling fifty WhatsApp messages a day and half are the same three questions, Haiku will save you an hour and cost you less than a rand.
Gemini Flash for structured tasks like booking confirmations or order summaries where you're filling in blanks rather than writing prose. It's faster than Claude and slightly cheaper (R0.03 to R0.06), but the tone is flatter. I used it for eight queries. Seven were perfect. One sounded like a form letter, and I added two sentences by hand. If you're okay with "functional and polite" rather than "warm and conversational," Flash works.
Gemini Pro I barely used. It's better than Flash at nuance, but it costs nearly as much as Claude Sonnet and I didn't see the benefit for restaurant queries. I tried it on one complaint and one custom dietary question. Both needed heavy editing. I'd skip it unless you're doing something more complex than order management.
The cost in rands and the time in minutes: what I actually spent
Forty queries over two weeks. Total AI cost: R3.20. Total time saved: four hours and ten minutes compared to typing every reply from scratch on my phone. Average cost per query: R0.08. Average time per query: six minutes with AI, twelve minutes without.
The biggest saving wasn't the rands—it was being able to answer twelve orders during load-shedding without stopping mid-service to type paragraphs on a phone screen. The second-biggest saving was never again typing "We deliver within fifteen kilometres of Observatory, R40 flat fee, usually forty-five minutes" for the ninth time in a week.
The biggest risk was the two dietary questions I almost sent without checking. If you're using AI for anything that touches allergens, refunds, or complaints, read the draft twice and know your own policies. The AI will give you something plausible. Plausible isn't the same as correct.
The checklist I'd give another restaurant owner testing this
Run a small test—twenty or thirty real queries—before you trust AI with your main WhatsApp line during dinner service. Here's the list I wish I'd had.
1. Give the AI your full menu, prices, and allergen notes in plain text. Don't assume it knows your lamb curry is halal or your chips are fried in the same oil as the calamari. Write it down once, paste it into every prompt as context. Claude Sonnet handled twenty-two menu questions without a mistake because I gave it a two-page menu doc every time.
2. Write down your booking rules, delivery radius, and payment methods as a checklist. "We take SnapScan, Yoco, EFT. No cash on delivery. We deliver within fifteen kilometres. Bookings for three or more only." The AI will follow the list. If the list is incomplete, the AI will guess. Guessing costs you a follow-up message or a lost booking.
3. Start with the boring, repetitive queries—delivery zones, opening hours, "Do you have WiFi?" These are low-risk and high-volume. If the AI gets one wrong, you'll catch it fast and the customer won't care much. I used Haiku for nine of these and saved an hour over two weeks.
4. Keep complaints, dietary restrictions, and payment questions manual until you're confident. I tried automating two complaints and three allergen queries. I had to rewrite all five. The apology tone was fine, but the solutions were wrong or incomplete. Let the AI draft, but read it like you're the customer's mother and you're allergic to sesame.
5. Track cost per query and time saved per category. I used a spreadsheet: date, query type, model used, cost in rands, time saved in minutes, did I edit the reply. After twenty queries I could see that menu questions were a perfect fit for Claude Sonnet, booking confirmations worked with Gemini Flash, and payment questions were a waste of time. Without the tracking, I'd still be guessing.
6. Test during normal hours first, not during load-shedding or a Friday night rush. I got lucky—my first AI-assisted dinner service was a quiet Tuesday, so when one reply needed fixing I had time to catch it. If I'd started on a Saturday with Stage 6 and thirty orders, I'd have sent something wrong and spent the next day apologizing.
Questions people ask
Can I connect AI to my Yoco terminal or bank account so it checks payments automatically?
Not directly through Kryotta yet—it's a text-and-image workspace, not a payments integration. You'd need a developer to build a bridge between your bank API and the AI, or you'd manually update a text file with order statuses and feed that to the AI as context. I tried the text file approach and it was clunky. For now, I check payments myself and let AI handle everything else.
Which model is cheapest for high-volume WhatsApp queries?
Claude Haiku 4.5 at R0.02 to R0.05 per query. I used it for delivery zones and opening hours—nine queries, total cost under fifty cents. If you're answering the same question fifty times a week, Haiku will save you hours and cost you less than a rand a day.
What happens if the AI gives a wrong answer and a customer gets sick?
You're liable, not the AI. If you're using AI for allergen or dietary questions, double-check every reply against your actual prep sheets before you send it. I almost told a customer the pasta was dairy-free when it wasn't. I caught it because I read the draft twice. Don't trust plausible—check correct.
Does this work if I'm on Mr D or Uber Eats instead of WhatsApp?
The workflow's the same—copy the customer question, paste it into Kryotta, send the reply—but those platforms rate you on response time, and adding an AI step might slow you down unless you're very fast. I only tested this on direct WhatsApp orders where I control the conversation. If you're on a delivery platform, try it on ten queries during a quiet shift and time yourself.
I'll keep using AI for menu questions, bookings, and delivery zones. I'll keep complaints and allergen queries manual until I've done another hundred and I'm certain I won't miss an edge case. If you're in Joburg or Cape Town and you're losing orders during load-shedding because you can't type fast enough on mobile data, try this with twenty queries and track the cost. It won't replace your front-of-house, but it'll keep you answering customers when the lights are out.
If you want to test Compare or Auto routing with your own restaurant queries, you can start at Kryotta with Claude, Gemini, and the other models I used here—no integration required, just copy, paste, and send.
