E-commerce: what worked on Jumia and Instagram storefronts
I started with a jewellery vendor in Surulere who runs an Instagram shop selling beaded necklaces, waist beads and anklets. She posts three times a day, answers DMs until midnight, and ships via GIG Logistics. Her bottleneck was writing captions that actually described the product—length, clasp type, whether it stretches—without sounding like a police report.
Works today: Product descriptions from a photo and three bullet points. I used Claude Sonnet 4.5 inside Kryotta's image input. Take a picture of the item, list the price, material and size, then ask for a 40-word caption with two relevant emojis. Took her caption-writing time from twelve minutes per post to under two. She now batches a week's posts on Sunday evening and her DM questions about sizing dropped by half because the captions actually answered them.
Works today: Turning a product list into a Jumia bulk-upload CSV. She had 80 SKUs in a Notes file—just names, prices and rough descriptions. Gemini Flash read the list, asked three clarifying questions (category structure, shipping weight, return policy), then output a formatted CSV with all mandatory Jumia fields filled. Uploaded first try. Cost about ₦200 in tokens; a VA would have charged ₦8,000 and taken two days.
Still needs you: Knowing which product to shoot next. I tried feeding her last month's sales data into DeepSeek V3 and asking what to restock. It suggested "blue items" because three of her top five sellers were blue. Useless. You still need to notice that customers keep asking for waist beads in size 38 and you only stock up to 34, or that wedding season is eight weeks out and gold sets always move in April.
Still needs you: Negotiating with logistics partners when a shipment goes missing. AI can draft the email, but it won't know that your GIG rep responds faster on WhatsApp than email, or that you need to reference the October delivery that also went to Ibadan to get her to prioritise your case.
Healthcare admin: scheduling and patient comms that don't need a nurse
I worked with a dental clinic in Lekki that runs on WhatsApp. Patients book via message, the receptionist confirms, and they send payment proof through Paystack before the appointment. The receptionist was spending ninety minutes a day answering "Do you take HMO?", "What time are you open Saturday?", and "Can I come today?"
Works today: A WhatsApp auto-reply script that hands off to AI for common questions. We used Llama 3.3 70B with a twelve-question FAQ document and a prompt that says "If the question isn't in the FAQ, say 'Let me connect you to the front desk' and stop." It answers insurance questions, opening hours, services offered and pricing for cleanings and fillings. Doesn't try to diagnose or book. The receptionist now handles thirty messages a day instead of ninety, and the clinic hasn't missed an urgent case because the handoff works.
Works today: Turning voice notes into appointment records. Patients leave two-minute voice notes explaining their issue. We used Gemini Flash's audio input to transcribe and summarise: chief complaint, previous treatment, preferred date. The dentist reviews the summary in fifteen seconds instead of listening to the full note. Cut her admin time by forty minutes a week.
Still needs you: Deciding whether a case is urgent. A patient messaged saying "my tooth hurts when I drink cold water." The AI flagged it as routine and suggested booking in five days. The receptionist read the follow-up message—"it's been hurting for three weeks and now my gum is swollen"—and moved her to same-day. AI reads one message at a time; you read the thread and the tone.
Still needs you: Handling payment follow-up when someone sends the wrong amount or the transfer doesn't reflect. The AI can draft the message, but you're the one who checks Paystack, knows that this patient always pays eventually, and decides whether to hold the slot or release it.
Restaurants: WhatsApp ordering and menu updates that don't need a developer
I tested this with a bukka in Ikeja that does rice, stew, proteins and sides. Orders come through WhatsApp, the owner types them into a notebook, then calls them out to the kitchen. She wanted to move to a Google Sheet so her cook could see orders in real time, but she didn't want to type every order twice.
Works today: Turning a WhatsApp order message into a spreadsheet row. Customer messages "I want jollof, two chicken, plantain, no salad." We fed that into Claude Haiku 4.5 with a prompt that maps it to columns: dish, protein, sides, exclusions, customer name, time. It appends the row to a Google Sheet via Zapier. Her cook now checks the sheet instead of waiting for her to shout orders. Errors dropped because the sheet is clearer than her handwriting.
Works today: Generating a daily menu image for her WhatsApp status. She used to pay a graphics guy ₦1,500 per day to update a Canva template with the available proteins and soups. Now she lists what's in stock, and Stable Image Core makes a clean menu card with her logo and prices in under a minute. She posts it at 10 a.m. and orders start coming by 10:15.
Still needs you: Knowing when to 86 a dish because you're running low. She has twelve portions of goat meat left at 1 p.m. and twenty orders already. AI can count, but it doesn't know that Fridays are always busier after 2 p.m., or that she can send her assistant to the market for more meat if she decides in the next thirty minutes.
Still needs you: Tasting the food and adjusting seasoning. I mention this because one owner asked if AI could analyse a photo of her stew and tell her if it needed more pepper. It can't. You still cook.
Real estate: listings and client screening that save you site visits
I worked with an agent in Abuja who lists flats and duplexes on PropertyPro and Instagram. She was spending two hours writing each listing and another hour answering questions from people who couldn't afford the rent or wanted a property type she doesn't handle.
Works today: Turning a site visit voice note into a full listing. She walks through the property, records a three-minute note describing the rooms, finishes, location and landlord's terms, then feeds it into Gemini Pro. It outputs a 200-word listing with a headline, neighbourhood context and a bullet list of features. She edits for tone and posts. Cut her listing-writing time from two hours to twenty minutes.
Works today: Screening inquiry messages before she replies. Someone messages "I need a 3-bed in Wuse 2 under ₦2 million per year." She pastes that into Claude Sonnet with her current inventory and a prompt that says "Does this client's budget and requirement match any of my properties? If no, draft a polite referral message." If there's no match, she sends the referral and moves on. Saves her from spending thirty minutes on a call that goes nowhere.
Still needs you: Reading the client to know if they're serious. Budget and requirements match, but the client asks "Can I pay in instalments over twelve months?" or "Is the landlord flexible on the deposit?" You know from experience which landlords negotiate and which don't, and whether this client will actually close or waste three site visits.
Still needs you: Taking the photos that make the listing work. AI can write the description, but if your photos show a dark parlour with clutter in the corner, the listing won't get saves. You still need to stage the shot, open the curtains, and frame it so the space looks bigger than it is.
CA and legal: document drafting and review that don't need a senior
I tested this with a CA in Victoria Island who handles tax filings, company registrations and contract reviews for small businesses. Her bottleneck was first-draft contracts and turning a client's messy records into a structured format for FIRS.
Works today: Drafting a service agreement from a client brief. Client says "I need a contract for a six-month social media retainer, ₦150,000 per month, payment on the first, deliverables are three posts per week and one Reel, client owns the content." She feeds that into Claude Sonnet 4.5 with a template prompt and gets a four-page draft with termination clauses, IP assignment and payment terms. She reviews, adjusts for the client's industry, and sends. Cut her drafting time from ninety minutes to twenty.
Works today: Extracting transaction data from bank statements and receipts. A client sends her three months of Zenith Bank statements as PDFs and a folder of receipts. She uses Gemini Pro to read the PDFs, categorise expenses (inventory, logistics, marketing, utilities), flag duplicates, and output a summary table. She spot-checks it and hands it off to her junior to finish the books. Saved her four hours of manual entry per client.
Still needs you: Deciding which clause to push back on when a client sends a vendor contract. AI can summarise the terms and flag unusual clauses, but you're the one who knows that this client's cash flow can't handle a 50 per cent upfront deposit, or that the liability cap is too low given the contract value.
Still needs you: Advising the client on tax strategy when their business model changes. They're moving from selling physical products to offering a subscription service. AI can explain the VAT implications, but you're the one who knows whether to register them as a new entity or restructure the existing one, based on their growth plan and risk tolerance.
The through-line: AI does the first pass, you do the judgement call
Every task that worked had the same shape: structured input, clear output format, and a human who checks the result before it goes out the door. Every task that failed was one where the AI had to guess what I wanted, make a decision with incomplete information, or handle an exception that wasn't in the training data.
If you're trying AI in your business, start with the task that has the clearest before-and-after. Not "make my marketing better," but "turn this product photo and price list into a 40-word Instagram caption." Not "help me grow," but "read these five customer messages and tell me which two need a reply today." You'll know in one day whether it works, and you won't spend three weeks tuning a prompt that was never going to land.
I ran all of this inside Kryotta because I needed to switch models depending on the task—Sonnet for contracts, Flash for transcriptions, Stable Image for the menu cards—without opening five different tabs. Your mileage will vary depending on your industry, but the checklist format holds: pick the task, try it on three real examples, and if it saves you thirty minutes, keep it.
Questions people ask
Which model should I start with if I've never used AI for business?
Gemini Flash. It's fast, cheap, handles text and images, and you'll know in ten minutes whether the task is one AI can do. If Flash works, you probably don't need a bigger model. If it doesn't, try Claude Sonnet before you assume AI can't help.
Do I need to learn prompting or can I just ask normal questions?
Normal questions work for simple tasks—summarise this, list the action items. For anything you'll do more than twice, write a three-sentence prompt that includes the format you want and one example. You'll get consistent output and won't have to re-explain every time.
How much does this cost in naira per month if I'm a solo business?
For the tasks I listed—captions, transcriptions, drafts, scheduling—you'll spend ₦2,000 to ₦8,000 per month depending on volume. That's assuming you're not generating fifty images a day or running reasoning models on every task. Track your usage for two weeks and you'll see the pattern.
Can I do this on my phone or do I need a laptop?
Most of it works on mobile through a browser or WhatsApp if you're using an API setup. Writing and reviewing is easier on a laptop, but I've seen vendors batch their product captions on a phone during their commute. Start where you are.
If you want to try the tasks I've described here without juggling five subscriptions, Kryotta gives you access to Claude, Gemini, Llama, Stable Image and the rest in one workspace. It's faster than switching tabs, and you can compare outputs when you're not sure which model to trust.
