Did AI actually save time, or did I just spend thirty hours teaching it to write "your appointment is confirmed"?
I run a small business-admin consultancy in Pune. Most of my clients are solo operators or two-person teams: a CA firm that handles GST returns for twenty e-commerce sellers, a dermatology clinic with one doctor and a receptionist, a tuition centre that runs doubt-solving sessions on WhatsApp, a real-estate broker with thirty rental listings in Kharadi and Viman Nagar, a café near Fergusson College, and a woman who sells handmade jewellery on Meesho and Flipkart. Every one of them had asked me the same question in October: "Can AI handle any of this, or will it just make more work?"
I didn't have a proper answer. I'd used Claude to draft emails and summarise contracts, but I hadn't tested it against the specific, repetitive tasks that fill a working day in India—Meesho order replies during Diwali peak, WhatsApp doubts from Class 10 students at 9 PM, GST invoice checks, appointment confirmations in English and Marathi, Zomato review replies, property listing summaries for brokers who need to send fifty a week. So I made a deal with all six clients: I'd spend sixty days using AI for one or two recurring tasks in each business, track what worked, and report back. No cherry-picking, no skipping the awkward stuff. If a model wrote a GST reply that would confuse a seller, I'd say so.
I used Kryotta because it let me switch between Claude, Gemini, Llama, and DeepSeek without opening six browser tabs, and because Auto routing picked the best model for each task without me having to guess. I kept a log in Google Sheets: task, model used, time saved, what broke, what I still had to fix by hand.
Here's what actually happened in six Indian industries over two months.
E-commerce: Meesho and Flipkart order replies during Diwali peak—Auto routing picked Gemini Flash, and it handled 80 per cent
My jewellery-seller client gets fifty to ninety order messages a day during Diwali season. Most are variations on five questions: "When will it arrive?" "Can I change the delivery address?" "I want to return this, the colour doesn't match." "Do you have this in gold-plated?" "Payment done, when dispatch?"
I wrote a prompt template for each scenario and fed it real order details—order number, courier tracking ID, product SKU, return window. Kryotta's Auto routing picked Gemini Flash for 70 per cent of the replies and Claude Haiku 4.5 for the rest. Flash was faster and cheaper; Haiku came in when the query had a complaint or needed a softer tone.
What worked: Flash wrote clear, polite replies in under three seconds. "Your order #MEE12345 will reach Kharadi by 28th October. Tracking: [link]. Let me know if the address needs updating." It handled address changes, return instructions, and dispatch confirmations without me touching the text. My client copied the reply, pasted it into Meesho chat, sent. Done.
What didn't: Flash struggled with edge cases. One customer wrote, "I ordered silver jhumkas but got gold-plated, my mother-in-law will be upset, can you send silver ones by tomorrow?" Flash replied with the standard return policy but missed the urgency and the family context. I rewrote it by hand: "I'm sorry about the mix-up—let me send silver jhumkas today by Speed Post, you'll have them by tomorrow evening. Keep the gold-plated pair as an apology." That took two minutes, but it saved a one-star review.
Over sixty days, AI drafted 1,200 replies. My client edited or rewrote 240. Time saved: about ninety minutes a day during peak season. Cost: ₹180 in Kryotta credits for the whole period.
Education: WhatsApp doubt-solving for Class 10 maths—Claude Sonnet 4.5 explained well, but parents wanted voice notes
My tuition-centre client runs doubt-solving sessions on WhatsApp Business. Students send photos of maths problems at 9 PM; she replies by 10 PM with a step-by-step explanation. She was spending two hours a night on this and asked if AI could take over.
I tested Claude Sonnet 4.5 because it's good at reasoning through multi-step problems. I fed it photos of quadratic equations, trigonometry, and coordinate geometry questions from Maharashtra Board textbooks. The prompt was: "Explain this step-by-step for a Class 10 student. Use simple language. Show every step."
What worked: Sonnet wrote clear, patient explanations. "To solve x² + 5x + 6 = 0, we'll factorise. Look for two numbers that multiply to 6 and add to 5. That's 2 and 3. So (x + 2)(x + 3) = 0. Now set each bracket to zero: x = –2 or x = –3." Students understood. Parents were happy.
What didn't: Some parents wanted voice notes, not text. They'd reply, "Can you explain this on a call? My son doesn't read long messages." AI can't do that yet. My client ended up recording five-minute voice notes for 20 per cent of doubts, which took longer than typing. Also, Sonnet occasionally skipped a step if the problem had three or four parts. I had to check every explanation before my client sent it, which added ten minutes a night.
Over sixty days, AI handled 180 doubts. My client recorded voice notes for 35 and edited Sonnet's text for another 20. Time saved: about forty minutes a night, five nights a week. She kept using it.
Real estate: property listing summaries for brokers—Llama 3.3 70B was fast and boring, which is what brokers want
My real-estate client sends fifty listing summaries a week to other brokers on WhatsApp. Each summary has to include square footage, rent, deposit, furnishing, proximity to IT parks, and any deal-breakers (no pets, vegetarian tenants only, etc.). He was copying and pasting from old messages and rewriting the same phrases every time.
I wrote a prompt: "Summarise this rental listing in 60 words. Include rent, deposit, square feet, furnishing, and location. Keep it factual." I fed it listing details from his spreadsheet and tested Llama 3.3 70B, Gemini Flash, and Claude Haiku. Auto routing picked Llama for 90 per cent of the summaries.
What worked: Llama wrote fast, plain summaries. "2 BHK, 950 sq ft, semi-furnished, ₹22,000/month + ₹1.5L deposit. Kharadi, 10 min to EON IT Park. Lift, covered parking, vegetarian tenants only. Available 1st Nov." Brokers replied faster because the summary had everything they needed in one message. My client's WhatsApp reply rate went from 40 per cent to 65 per cent.
What didn't: Llama occasionally dropped a detail if the listing had too many clauses. One summary left out "no pets," and a broker sent a family with a Labrador to view the flat. My client now checks every summary before sending, which takes thirty seconds per listing. Still faster than writing from scratch.
Over sixty days, AI wrote 480 summaries. My client edited 50. Time saved: about three hours a week.
Healthcare admin: appointment confirmations in English and Marathi—Gemini Pro handled both languages, but patients still called to confirm
My dermatology-clinic client sends fifty appointment confirmations a week by WhatsApp. Half her patients prefer Marathi; half prefer English. Each confirmation has to include date, time, doctor's name, clinic address, and a reminder to bring previous prescriptions.
I wrote a bilingual prompt: "Write an appointment confirmation in [language]. Include date, time, doctor's name, address, and a line asking the patient to bring previous prescriptions. Keep it polite and short." I tested Gemini Pro because it handles Marathi better than Claude or Llama.
What worked: Pro wrote clear confirmations in both languages. English: "Your appointment with Dr. Deshmukh is confirmed for 3rd Nov, 11 AM, at Shivaji Nagar clinic. Please bring previous prescriptions. Call 98765-43210 if you need to reschedule." Marathi: "तुमची भेट डॉ. देशमुख यांच्यासोबत ३ नोव्हेंबर, सकाळी ११ वाजता, शिवाजीनगर क्लिनिकमध्ये निश्चित झाली आहे. कृपया आधीचे प्रिस्क्रिप्शन घेऊन या." Patients replied "OK" or "Thank you," which meant they'd read it.
What didn't: 30 per cent of patients still called the clinic to confirm, even after receiving the WhatsApp message. My client thinks it's a habit, not a clarity issue. Also, Pro occasionally used formal Marathi that older patients found stiff. My client now edits the Marathi confirmations to sound warmer, which takes twenty seconds each.
Over sixty days, AI wrote 480 confirmations. My client edited 60. Time saved: about ninety minutes a week. Patients stopped showing up late because the confirmation included the full address and a Maps link.
CA and legal: GST invoice checks for e-commerce sellers—DeepSeek V3 caught errors, but I still wouldn't trust it with a tax notice
My CA client checks GST invoices for twenty e-commerce sellers who ship across India. Most errors are small—wrong HSN code, missing IGST on an interstate sale, incorrect tax rate—but they add up during audits. He was spending four hours a week on manual checks and asked if AI could flag errors before he reviewed the invoices.
I wrote a prompt: "Check this GST invoice for errors. Look for: wrong HSN code, missing IGST on interstate sales, incorrect tax rate, missing buyer GSTIN for B2B sales. List any errors clearly." I tested DeepSeek V3 because it's good at structured reasoning and cheaper than Claude. Auto routing picked DeepSeek for 85 per cent of the checks.
What worked: DeepSeek caught most common errors. "Line 3: HSN code 6203 is for men's trousers, but the item description says 'women's kurti'—should be 6204. Line 7: Sale to Maharashtra from Karnataka, but you've charged CGST+SGST instead of IGST. Line 9: B2B sale, but buyer GSTIN is missing." My client fixed the errors before filing, which saved follow-up emails and penalty interest.
What didn't: DeepSeek missed two errors over sixty days—one was a decimal-point mistake in the taxable value (₹1,200 written as ₹12,000), and the other was a judgment call about whether a product counted as a composite supply. My client caught both during his final review. He says he still wouldn't trust AI to handle a tax notice or a GST audit query, but it's useful for first-pass checks.
Over sixty days, AI checked 320 invoices. My client corrected DeepSeek's output for 15 and found 2 errors DeepSeek missed. Time saved: about two hours a week.
Restaurants: Zomato review replies—Claude Haiku 4.5 was polite and fast, but one-star reviews still need a human
My café client gets fifteen to twenty Zomato reviews a week. Most are four or five stars ("Great filter coffee, cosy place"), but three or four are one or two stars ("Waited 25 minutes for a sandwich, staff was rude"). She was replying to every review by hand and asked if AI could draft replies.
I wrote two prompts: one for positive reviews ("Thank the customer warmly, mention something specific they liked, invite them back") and one for negative reviews ("Apologise sincerely, acknowledge the specific issue, offer to make it right"). I tested Claude Haiku 4.5 because it's good at tone. Auto routing picked Haiku for 95 per cent of the replies.
What worked: Haiku wrote warm, specific replies to positive reviews. "Thank you, Priya! So glad you loved the filter coffee—we roast the beans fresh every Monday. Hope to see you again soon." Customers replied with heart emojis. For mild complaints (three stars, "Sandwich was cold"), Haiku wrote polite apologies: "Sorry the sandwich wasn't hot, Rohan—we'll make sure it's served fresh next time. Please give us another chance."
What didn't: One-star reviews need a human. One customer wrote, "Worst experience, manager shouted at me in front of other customers, will never come back." Haiku replied, "We're sorry to hear that—please DM us so we can make it right." My client rewrote it: "I'm really sorry that happened, and it's not acceptable. I'd like to speak with you directly to understand what went wrong and make sure it doesn't happen again. Please call me at [number] or DM, whichever is easier." That took five minutes, but it turned the customer around—she came back two weeks later and left a four-star review.
Over sixty days, AI drafted 180 replies. My client rewrote 12. Time saved: about forty minutes a week.
Questions people ask
Which model did Auto routing pick most often?
Gemini Flash for speed tasks (e-commerce replies, listing summaries), Claude Sonnet 4.5 for reasoning (maths doubts), DeepSeek V3 for structured checks (GST invoices), and Claude Haiku 4.5 for tone-sensitive replies (Zomato reviews). Auto routing saved me from guessing which model to use for each task.
Did AI actually save money, or did I just spend sixty days editing its output?
It saved time on high-volume, low-stakes tasks—order replies, listing summaries, appointment confirmations. My clients spent less time typing and more time on work that needed judgment. Editing took 10–20 per cent of the time the original task would have taken. Total Kryotta cost across six clients for sixty days: ₹850.
What tasks should I not hand to AI yet?
One-star reviews, tax notices, urgent complaints with family or emotional context, and anything where getting it wrong costs you a client or a penalty. AI drafts; you decide.
Do I need to know prompting, or can I just type "write a reply"?
You need to be specific once, then reuse the prompt. "Write a reply" gets you generic text. "Write a Meesho order reply for [order number], include tracking [link], confirm delivery by [date], keep it under 40 words" gets you something you can send. Write the prompt properly the first time, save it, and use it fifty times.
I still check every output before my clients send it, but I'm checking six businesses' worth of work in two hours instead of six. That's the difference. If you're doing the same task twenty times a week—order replies, appointment confirmations, review responses, listing summaries—AI will draft it faster than you can type. You just have to decide if the draft is good enough to send.
You can try the same models I used at Kryotta—Auto routing picks the best one for each task, and you'll see the cost per reply before you send anything.
