Self improving AI for small business in IndiaAn AI that feels human and keeps improving itself.

Written by Madhur Tyagi, founder of The Human Algorithm. 8+ years, 25+ brands, since 2017.

Picture Meera. She runs a home bakery in Jaipur. Orders come in on WhatsApp all evening, so her helper Aman started using AI to draft the replies. It was fast. It was polite. Most nights it was fine.

Then one Saturday it told a customer the chocolate cake was ₹1,200. It also promised delivery “for sure by 6 PM” on a day Meera had no rider.

The customer was lovely about it. Meera still lost the money and a little sleep.

So she did something simple. She wrote seven rules on a card and stuck it on the fridge. Prices. No promises that are not on the card. Short and warm. Always end with a question. Anything about refunds, call Meera.

Now every draft gets checked against the card before Aman presses send. When a rule fails, the draft gets fixed. Every Sunday they look at what failed most and add one rule.

The AI did not get smarter on its own. The rules got smarter, because real mistakes taught them.

Meera and Aman are an example, drawn from patterns we see in real chats with Indian sellers. They are not real clients. The loop is real.

What does self improving AI really mean?

You will hear the phrase used for two very different things.

• The big one. AI that designs smarter AI by itself. It makes headlines. Researchers say it is still unproven, and it has nothing to do with your WhatsApp.
• The useful one. A loop that gets a little better every week at one job, because every draft is checked and every miss is written down. This one you can start tonight.

Developers use the useful one for code. AI writes the code, real tests run, the errors go back in, and AI fixes it. The clever part is not the AI. The tests do the grading.

There is research behind the idea too. A 2023 paper called Self Refine showed that AI drafts often improve when feedback goes back in, round after round. Our version adds one thing: the feedback comes from your rules and your customers, not from AI’s opinion of itself.

The Taste Test Loop: Draft, Check, Fix, Learn

Four steps, in a circle. Your rules sit in the middle. Skip the check, and AI is guessing. Skip the learning, and you fix the same mistake forever.

Notice the last step. That is the part most people skip, and it is where the “self improving” really lives. A loop that never learns is just a spell check.

Where can a small business use it?

Anywhere the same kind of message goes out again and again, and you can write the rule in one line.

Tool: test a reply against your rules

This is the Check step, in your hands. Type your price, paste a reply you sent or one AI drafted, and see which rules it passes. Nothing leaves this page.

A rule tester like this is simple on purpose. It catches the costly mistakes: a wrong price, a promise you cannot keep, an angry customer left with a bot. Those are the ones that lose money and trust. For a deeper look at a reply’s warmth, try the free Reply Score.

Your Rule Card, and 5 scripts to copy

Start with the card. One page, 5 to 7 rules, written in your own words. Then use the checker and fixer as two separate chats, so the cook never tastes his own dish.

Real mistakes make great rules. A coffee brand’s chat assistant once told a buyer a pack was out of stock while its own page said Add to Cart. Our breakdown of that wrong answer shows how a single rule would have caught it.

Is the loop worth it for this task?

Not every task deserves a loop. Tick what is true for the task you have in mind.

What should you never hand to the loop?

A loop makes drafts better. It should never press send on things that need a heart or a decision.

• Refunds and complaints. An upset customer wants a person, not a better sentence.
• Apologies for real mistakes. Say sorry yourself. It is worth it.
• New prices and discounts. These are your decisions, every time.
• Health, legal or money advice. Hand it to a person who knows.
• A message from you to someone new. Your name, your yes.

If you want a wider safety check before you give any AI tool access to your accounts, read our 10 point AI agent safety check.

How we use it at The Human Algorithm.

Our promise is one line: an AI that feels human and keeps improving itself. Feels human always leads. I built our Revenue Desk on this loop because I never wanted a client to wake up to a wrong price sent in their name.

“Keeps improving itself” does not mean AI marks its own homework. It means the system learns every week from your rules, your customers and our team.

We do the same with our own work: our emails, our articles and the daily plans we send. The loop is how a small team stays sharp without getting cold.

Self improving AI is not magic. It is a cook, a taster and a notebook, every week. That is how an AI can feel human and keep improving.

About this guide: the loop idea comes from how developers test AI written code (tests grade the code, errors go back in), from a developer guide published 2 Nov 2026, and from the 2023 Self Refine paper on improving drafts with feedback. The Taste Test Loop, the Rule Card and the tester are ours, labelled as ours. Meera and Aman are an example. No numbers in this guide come from studies we could not check. Madhur Tyagi is the founder of The Human Algorithm., a growth and automation practice in New Delhi that helps Indian founders, coaches and creators reply faster, follow up kindly and earn trust in their DMs and WhatsApp. 8+ years, 25+ brands, since 2017.