AI for a Small Business

Every June, Anand Stationers answers the same question about four hundred times: what does my child need for class six at St Karen’s, and what will it cost.

The shop’s WhatsApp number carries most of it. A parent sends a school name and a class. Somebody behind the counter finds the right booklist on the board, types out the titles, adds the uniform items, works out a total and sends it back. It takes between six and eleven minutes depending on how legible the booklist is and how many follow-up questions arrive.

Sumit Anand, who runs the shop with his father, spent five days counting properly rather than remembering. The tally came to 212 booklist enquiries in a single ordinary week of June, at an average of just under eight minutes each. That is twenty-eight hours, spread across four people who also have a shop floor to run. In the same week the shop wrote eleven credit-account reminder letters, which took forty minutes in total and which everybody hated.

The hated task was the one Sumit had planned to hand over. The tally moved it to the bottom of the list in an afternoon, which is what a tally is for.

So: the booklists. Frequent, obviously. Made of words, obviously. The third filter is where it stopped being obvious.

If a booklist reply comes back wrong, does anybody notice cheaply? A wrong greeting is visible in a second. A wrong title is not. A parent who is told their daughter needs the 2024 edition of a grammar book when the school has moved to the 2026 one buys the wrong book, discovers it in the second week of term, and comes back angry at a shop that had one job. At about two thousand four hundred rupees a kit, a run of thirty wrong replies is a real number, and it is the kind of number that also costs the shop a family for five years.

Sumit’s father made the objection plainly. The whole shop, he said, is thirty-one lists that change every year and nobody else in Patna keeps properly. If the machine starts guessing at them, we are selling the one thing we have.

The other side of the argument was not hypothetical either. Twenty-eight hours a week is one person for most of June. In the previous season the shop had lost custom simply by being slow: parents who sent a message at nine in the evening and got a reply at eleven the next morning, by which time they had gone to the shop opposite the bus stand. Sumit could name four.

So the choice was between leaving twenty-eight hours on the table for a second year, and putting a plausibility machine in front of the one record his customers trusted. Both sides had a cost and neither was small.

What broke the deadlock was drawing the task out as boxes rather than treating it as one job. Reading the message, deciding which school and class it refers to, finding the right list, typing it out, pricing it, and answering the follow-up. Six boxes, not one. Only two of them were drafting.

The box that actually consumed the minutes was the fourth, typing out between eleven and twenty-four lines from a photocopy pinned to a board. The box that caused the arguments was the second, where a parent writes “Karen’s” and means one of two schools in Patna with similar names, or writes “class 6” when the school’s list distinguishes between the two sections. Getting that box wrong sends a correct list to the wrong family, which is a failure the shop had been making about twice a week already, by hand, at four in the afternoon in the fourth week of June.

That changed the shape of the question. Sumit had been asking whether a model could be trusted to write a booklist reply. The map asked something narrower and answerable: whether it could be trusted to choose between thirty-one named documents, and to copy one out without changing it.

Meghana Rao had done everything the right way round, which is why the argument that followed was worth having.

She had written the facts sheet first: thirty-five minutes, one page, prices by fabric and border, the delivery areas within and beyond Mysuru city limits, the fourteen-day exchange rule in her mother’s own words, and four questions customers ask that her mother answers badly when tired. She had pasted in five real replies her mother had sent, labelled EXAMPLE, with an instruction not to reuse their content. She had added the bracket: if the facts sheet does not answer part of this, write CHECK and what is missing, instead of guessing.

Then she built the ten-case test out of real messages from the last month rather than inventing them. Four ordinary price enquiries. Two where the customer had not said which fabric. Two awkward ones: a woman writing in a mix of Kannada and English about a wedding date, and a man who wanted three saris couriered to three different addresses. One outside the shop’s range. One with a trap.

Nine passed. On the tenth, a customer asked what the shop charges for blouse stitching and how long it takes. Kadambari Silks stopped offering stitching in 2024. The reply that came back was warm, correctly formatted, in her mother’s rhythm, and said seven hundred and fifty rupees with a five-day turnaround.

Neither number had ever existed. The facts sheet did not mention stitching at all, so there was nothing to contradict, and the bracket had not fired because a price for stitching is exactly the sort of thing that feels natural to supply.

Meghana wanted to ship. Nine out of ten, she argued, and the failure was an edge case about a service nobody asks for any more. Her mother, Sharada, had a different view, which she put as a question: if it will do that for stitching, what will it do for fall-and-pico, for dyeing, for the alteration work they also stopped?

The fix available in the lesson was to close the book: answer only from the facts sheet, and where the answer is not there, reply NOT IN FACTS and nothing else. Meghana tried it and it worked immediately on the stitching case. It also broke two of the nine that had been passing. A customer asking about an Ilkal sari got NOT IN FACTS, because the sheet said Ilkal cotton and the strict instruction had made the assistant literal-minded. A customer asking about delivery to Hebbal got the same, because the sheet listed areas and not that one.

So the decision had a cost on both sides, and it was a real one for a shop whose whole advantage over the showroom on Sayyaji Rao Road is that somebody answers warmly and quickly. A system that refuses when it is unsure is one you can hand to the boy who works Saturdays. A system that improvises is not. But a system that refuses a paying customer asking about a sari sitting on the shelf behind you is a lost sale, and there were two of those in ten.

Meghana made the case for scoring the tenth as half a pass. The tone was right, the format was right, and only one sentence was wrong. Her mother refused, and the reason is worth setting down because it is the whole discipline: a reply with perfect tone and an invented price is a failure, and calling it half a pass is how a business talks itself into sending something unsafe two hundred times a month.

There was one more thing neither of them had expected. Of the ten cases, the pair that had been hardest to write were the two where the customer had not said enough: no fabric named, no budget, no date. Those are the cases the bracket exists for, and they were the two Meghana had nearly left out on the grounds that they were not really enquiries.

Thirty-eight pages, four schedules, and a buying agent who had already mentioned twice that another unit was waiting.

R. Velan’s son Arun did what a great many small businesses would now do, and did it well. He uploaded the agreement and worked through it methodically rather than asking one sweeping question. He asked for a one-page summary: parties, term, price and how it changes, notice, termination, liability, and what happens to their data. He asked for clause 12 in plain English and what it would mean if they wanted to stop after eighteen months. He asked what was unusual compared with a standard supply agreement of this type.

All of that came back usefully, with quotes and clause numbers he could check, and he checked them. The plain-English explanation of the indemnity was the most valuable forty seconds of the weekend: contract English is a dialect, and Velan had been guessing at it for eleven years.

Then Arun asked the question that mattered most to his father, because the unit’s survival has always depended on running three buyers at once: does this contract contain an exclusivity clause.

The answer was no exclusivity clause is present in this agreement.

It came back in the same calm register as everything else, and it was the only answer in the whole session that could not be verified, because there was nothing to verify. A clause that is not retrieved and a clause that does not exist produce identical output.

Arun knew this in the abstract. The decision in front of him was not abstract. A trade lawyer in Coimbatore would see them on Monday afternoon for about eighteen thousand rupees for two hours, which is real money against an order worth around fourteen lakh in its first year. Taking Monday afternoon also meant signing on Tuesday at the earliest, and the agent had made it clear the buyer wanted a signature before their own board meeting. Velan had lost an order in 2023 by being three days slow with paperwork and had not stopped mentioning it.

On the other side: if the contract did tie up the machines, the unit would spend a year unable to take the two smaller buyers who had carried them through the last monsoon, and would find out in August when the floor was full.

There was a third consideration that Arun raised and his father dismissed, and which was the right consideration. The agreement itself contains a confidentiality clause forbidding disclosure of its terms to any third party without consent. Uploading it to a consumer AI service is arguably a disclosure. Whether it is depends on the wording and on the jurisdiction and is genuinely unsettled, and nobody in Tiruppur was going to resolve it on a Saturday. What they could do was stop adding to the exposure, so the rest of the weekend’s work was done on a small open model running on Arun’s laptop, which is slower and less fluent and sends nothing anywhere.

Arun then went back to the document and stopped asking it whether something was absent. He went through it section by section instead, asking for every obligation on Velan Knits in each one, with the quote and the clause number, and refusing to let one question cover more than one section. It took about fifty minutes and produced four pages of extracted obligations, most of them entirely ordinary: packing specifications, an inspection right, a late-delivery deduction of half a per cent a week.

He also asked for every date, notice period and deadline in the agreement, section by section, because his father had once lost a year to a renewal nobody had diarised.

Ayesha Qureshi had been disciplined about it, which is the only reason the conversation six weeks in was worth having at all.

Every draft had been read before sending. She had kept a tally of the ones where she disagreed with something material: not tone, not a clumsy sentence, but a fact, a price, a promise. Across two hundred and forty consecutive drafts there were four.

Four in two hundred and forty is a better record than most, and it is what made her want the next step. The sorting step had told her what her nights looked like: of roughly forty messages arriving between six in the evening and nine in the morning, twenty-two were answerable by anybody holding the price list. Drafting and reading those was costing about six hours a week, most of it in evenings she was not being paid for, and the salon two streets away was answering at eleven at night while she was not.

So: switch on auto-send for the price-list pile. Twenty-two messages a night, six hours a week back, and an evening response time her competitor could not match.

Her business partner, Nikhil, asked to see the four disagreements before agreeing, and they turned out to be the argument.

One was a bot cheerfully offering a Sunday slot at the Banjara Hills branch, which does not open on Sundays. One had quoted the price of a keratin treatment the salon does not do, assembled out of nothing. Two had quoted a cut and colour at the junior rate when the customer had asked for a named senior stylist, where the difference is about nine hundred rupees.

Three of the four were in the pile Ayesha wanted to automate. That is not a coincidence: the price-list pile is where the numbers are, and numbers are where the silent errors live.

There was also a point Nikhil had read about and repeated badly, which Ayesha checked herself. A tribunal in Canada had held an airline responsible for what its website assistant told a customer about a fare policy; the airline’s argument that the bot was a separate entity responsible for its own words had failed. The detail is jurisdictional and a Canadian tribunal decides nothing in Telangana. The shape of it survives the journey: a bot on your site speaks as your business, and if it quotes a price you may be held to it.

Against that, the cost of doing nothing was not nothing. Ayesha could name six enquiries in a month that had gone quiet between eleven at night and the following afternoon. At an average booking of about two thousand two hundred rupees, six a month is not a rounding error in a salon of that size, and the evenings were genuinely costing her.

Nikhil’s counter-proposal was the one that turned out to be right, and it sounded like a retreat when he made it. Take a different pile. There were nine messages a night that asked where the Kondapur branch is, whether there is parking, which branch does bridal work, and what time they close on Wednesdays. None of those carries a price or a commitment. The worst outcome from getting one wrong is a customer sent to the wrong car park.

Ayesha’s objection was arithmetic: nine messages is a hundred minutes a week, not six hours, and it was the six hours she wanted back. Nikhil’s answer was that exposure is not proportional to saving. The nine-message pile contains nothing the salon could be held to. The twenty-two-message pile contains a binding statement roughly one time in eighty, which at twenty-two a night is about one a week, sent in her name while she was asleep, to a customer who would reasonably expect her to honour it.

The cooperative had two hundred and forty pieces, a paper register with one line per weaver per piece, and a secretary, Kanchan Devi, who had never written a product description in her life.

The method was not in doubt. Get the facts into a spreadsheet first: one row per sari, one column per attribute: warp, weft, zari, length, blouse piece included or not, weaver’s village, loom days, price. Write five descriptions by hand, properly, in the cooperative’s own voice. Feed the rest in batches of twenty with a rule that a missing fact is written as MISSING and never filled.

Assembling the spreadsheet took two and a half days, which nobody had budgeted for and which turned out to be the project. The generation took about twenty minutes in total.

Then the question that had no obvious answer: how many of the two hundred and forty do you read before you publish.

Checking a row properly (every attribute against the weaver’s slip) took about ninety seconds. Two hundred and forty rows is six hours. With eleven days, a Diwali deadline, and two people who also had to photograph everything, six hours was not free.

The arithmetic that decides it is short. Check twenty at random, find nothing wrong, and the true error rate could still plausibly be as high as fifteen per cent: about thirty-six bad listings sitting in the catalogue. Check fifty and the ceiling falls to around six per cent. Check a hundred and it is about three.

What made the decision uncomfortable was that not every field carries the same consequence. A slightly flat sentence about the drape of a tussar costs nothing. A wrong fibre content costs a customer who washes a silk sari as though it were cotton, and it is a listing violation on every marketplace that exists: platforms do not prohibit generated descriptions and every one of them prohibits inaccurate attributes. The cooperative pays the return.

Kanchan’s co-worker argued for twenty, on the grounds that the first batch had been visibly good and the window was closing. Kanchan had been a weaver for twenty-two years before she took the register, and her objection was not statistical. She said that a wrong wash instruction on a Bhagalpur silk is not a bad listing, it is a ruined sari and a family telling everybody in their street.

The argument for twenty was better than it sounds, and it was not only about the deadline. The cooperative had thirty-one pieces whose zari column was blank in the register, because the weaver’s slip for those had been filled in by somebody who did not record it. Checking those rows against the slips would find nothing, because the slips do not say either. The only way to settle them was to go to the weavers, in four villages, which was a day nobody had.

What neither of them had examined was whether the two hundred and forty rows needed the same treatment at all. A description of drape and a declaration of fibre content are not the same kind of claim, and they do not fail in the same way. One is prose, written from a column, and a mistake reads as a slightly odd sentence. The other is an attribute the marketplace matches, filters and enforces, and a mistake in it is a return the cooperative pays for and a listing the platform can pull.

Read that way, the question stopped being how many of the two hundred and forty to check and became which fields on all of them, and which fields on a sample.

The quotes arrived in the same week and none of them could be compared with any of the others.

Kapoor Papers, the incumbent, quotes per ream. Supplier B quotes per kilogram. Supplier C quotes per pack of five hundred sheets, which is not a ream. Kapoor includes delivery within Ludhiana. B adds a delivery charge above three hundred kilograms and does not say what it is. C is silent on delivery altogether. Payment terms are immediate, thirty days and forty-five days respectively. Kapoor’s price is valid for twelve months, B’s for thirty days, and C’s document does not mention validity.

Deepak Bhatia had been meaning to do this comparison for three years and had not, which is the ordinary condition of supplier quotes in a small business and the reason suppliers write them this way.

The extraction took about fifteen minutes. Each quote went in with an instruction to return a fixed set of fields (price per single sheet in rupees, delivery cost per sheet at an order of five hundred kilograms, payment terms in days, minimum order, price validity, and anything conditional) using only figures stated, writing NOT STATED where a quote does not say, and estimating nothing.

Every extracted number was then checked against the original document, which took another twenty minutes and found one error. It was Deepak’s, not the model’s: he had been reading Supplier C’s pack of five hundred as a ream of four hundred and eighty for the whole of the previous week, which had made C look four per cent cheaper than it was.

Normalised, B came out roughly seven per cent below Kapoor per sheet. On nine tonnes a year that is about one lakh ninety thousand rupees, which in a four-machine press is a machine service and a half.

The NOT STATED entries were the reason Deepak did not act on it that afternoon. B’s delivery charge above three hundred kilograms was blank, and B’s price was valid for thirty days on a contract he wanted to hold for a year.

There was also the thing a spreadsheet cannot hold. In 2023, when a wedding-card customer defaulted on four lakh, Kapoor had carried Deepak for sixty days without being asked. In 2024 they had delivered on a Sunday morning during a rush. His son’s view was that eleven years of goodwill is not a discount and that one lakh ninety is one lakh ninety. His own view was that the Sunday had saved a customer worth more than that.

Neither of them could settle it with the table in front of them, because the table does not have a column for whether a supplier answers the phone in August. What the table did do was end the part of the argument that had been going round for three years, which was whether B was actually cheaper. Until that afternoon nobody in the press could have said, because nobody had put the three quotes into the same units, and a quote written in kilograms against a quote written in reams is not a comparison that can be done in your head.

Deepak’s son made the other point that mattered. Two of the three quotes had a NOT STATED field, and a blank field always resolves in the supplier’s favour if you let it. A price you have not asked about is not zero; it is a number the supplier has chosen not to show you at the stage where you are deciding.

So before the relationship question could be answered honestly, the blanks had to be filled in writing.

The confirmations had been running at rung three for eleven weeks: the model reads the order that arrives by WhatsApp or phone, pulls out six fields, drafts the confirmation, and leaves it prepared but unsent. Hiren Vora approves. Approving takes about four seconds, and the queue takes him twelve minutes a day.

His nephew Jay, who had built the thing, wanted to climb to rung four: let it send, notify Hiren, leave an undo window. Twelve minutes a day is an hour a week, and Hiren had not rejected a single item in five weeks.

Hiren was inclined to agree, and the argument for it was the one in the lesson: fifty consecutive clean items is a reasonable bar for moving up a rung, and this was far past fifty.

Then came the Friday. The broadband dropped for about ninety seconds in the middle of the afternoon batch. The automation’s call to the messaging service timed out, so it retried, as a sensible piece of software does. Twenty-three workshops received two identical confirmations. Four of them, reading the second as a separate order, sent stock back or rang to cancel one, and one workshop in Gondal put in a duplicate order for a clutch assembly that Shreeji then had to take back.

Nothing about the model had gone wrong. The drafts were correct. What had gone wrong is that a call which times out leaves two possible worlds (the action never happened, or it happened and the confirmation was lost on the way back) and from where the caller stands those look identical.

That reframed the decision. Jay’s case for rung four had rested on the error rate of the model, and the incident had nothing to do with the model. Hiren’s question was the one worth asking: if the system can send a thing twice while I am approving each one, what does it do when nobody is approving at all?

There was a second, quieter problem in the same conversation. Five weeks at a zero per cent rejection rate has two possible explanations. Either the automation is genuinely reliable, or Hiren had stopped looking. Twelve minutes for a hundred and eighty items a week is about four seconds each, and four seconds is long enough to read a confirmation properly only if the thing you have to check is in front of you.

It was not. The queue showed each confirmation as a paragraph: the workshop’s name, a courteous sentence, the parts, the total, a promised date, and a line about transport. To verify it, Hiren had to hold the original order in his head and read the paragraph against that memory, which is a different and much harder operation than comparing two things side by side. In week one he had done it. By week nine he was recognising the shape of a correct confirmation and approving on the strength of the shape.

Jay also raised the cost question, and it cut the other way from how he intended. At Shreeji’s volume the model costs a few rupees a week; nothing in the decision was about the bill. What rung four would have bought was twelve minutes a day of Hiren’s attention, and what it would have sold was the only moment at which a wrong part number is cheap to catch. A wrong part number on a confirmation is caught in four seconds by the man who took the order, or in four days by a workshop in Jamnagar that has fitted the wrong clutch.

That asymmetry is what the rungs are actually about. The model’s error rate is the same at rung three and rung four. What changes is how long an error lives before anybody sees it, and for a distributor the answer at rung four is however long it takes a workshop to open the box.

Ravi Meena had done the hard part, which is why the result was so nearly damaging.

He had defined the thing he was predicting rather than leaving it vague. A member counted as lapsed if they had attended at least eight times and had not attended in forty-five days, where forty-five came from looking at the distribution of gaps between visits and picking a point well beyond where regular members sit. About eleven per cent of members lapsed in a quarter.

He had built the customer table rather than analysing the raw entry log: one row per member, with first visit, last visit, visits in the last ninety days, average gap between visits, and membership type. Nine thousand four hundred rows across four years.

Then he pointed a free AutoML tool at it and asked it to predict the lapse column. It compared a dozen methods, tuned them, and reported ninety-four per cent.

The vouchers were already at the printer.

His daughter, who is in the second year of a statistics degree in Pune, asked one question on the phone: what would it have to know, on the day you want the prediction, to be that right about what a person is going to do.

Human decisions do not predict at ninety-four per cent. A number like that in a small business is not a triumph, it is a symptom, and the thing to look for is a column that only exists once the answer is known.

There were two. The export carried win_back_sms_sent, which the front desk sends only to members they have already decided are drifting: the model had learned to read Iron Yard’s own judgement back to it. It also carried locker_refund_date, which is populated only when somebody clears out, and is therefore a record of the outcome wearing the costume of an input.

Removing them meant telling his partner that the afternoon’s impressive number had measured nothing, having already talked about it twice. It also meant the vouchers were a decision without an analysis behind it, four days before a print deadline.

The other side was worse and slower. A model that scores ninety-four by reading the front desk’s own opinion will, in use, flag the members the front desk already flags, and Ravi would have spent a quarter congratulating himself on a list he could have written by hand.

There was a subtler version of the same fault sitting underneath, and it would have survived removing both columns. The tool had split the data at random by default, which lets a model learn from March while predicting February. Iron Yard’s membership behaves seasonally (a January intake, a collapse in the week of Holi, a thin May) and a method allowed to see the later months while predicting the earlier ones will look excellent and will never be able to do that in real life. Ravi had to find the option to split by date, and in that tool it was not the default.

He also had to face a question he had been avoiding, which is what he would do with the list. The plan had been a phone call and a voucher. Ravi’s partner wanted the flagged members moved down the priority list for personal-training slots, on the grounds that they were leaving anyway. That is the version to refuse: a model trained on the gym’s own past behaviour, used to decide who gets less attention, makes its own prediction come true and takes the member with it. Using it to decide who gets extra attention makes a wrong prediction cost a phone call and give somebody an unexpectedly good week, which is the cheapest kind of error available.

The message arrived inside the thread. Not a new email, not a lookalike address: the eleventh message in a conversation that had been running since the tender, quoting the previous message underneath it in the normal way.

It was in the project manager’s voice, which the practice knew well after five months. It referred to the site meeting on the twelfth and to the revised mullion drawing, both real. It said the contractor had changed banks after a merger, gave a new account number and IFSC code, attached a scanned letter on the contractor’s letterhead, and apologised for the timing. The English was faultless and the tone was exactly right.

Everything the practice had ever used to spot a fraudulent email was absent. There was no odd phrasing, no generic greeting, no wrong name, nothing that read as translated. Those cues had been doing the work for fifteen years and they had stopped working, for reasons that have nothing to do with whether Anantha & Rao use AI themselves.

Priya, the junior architect who handles payments on Thursdays, had the transfer screen open. Two months earlier the practice had written down a rule after an evening reading about this: any change to bank details, and any payment above two lakh, is verified by ringing a number already in their own records, not the number in the email, not the caller display, not the signature block.

She did not want to ring. Three reasons, all of them ordinary. It felt like telling a contractor you did not believe them, five months into a relationship that had been good. The contractor’s site team was waiting on the payment to release steel from the fabricator, and a day’s slip would push a crane hire that had been booked for a fortnight. And Srinivasan, the partner who would normally decide, was on a site visit in Sriperumbudur with his phone in his pocket.

The cost on the other side did not announce itself in the same way, which is the whole difficulty. If the message was genuine, ringing cost ten minutes and a small awkwardness. If it was not, four lakh sixty leaves the practice’s account into an account that will be emptied within the hour, and there is no realistic route back.

The only thing on the screen that pointed either way was the timing. The message had arrived on the afternoon the invoice was due, which is the moment at which a payment instruction is least likely to be questioned and most likely to be acted on quickly. Urgency was the one cue that had survived.

Priya did what people do in that position: she looked for more evidence in the message itself. The letterhead was right. The GST number on the scan matched the one on the invoices. The signature looked like the signature. She compared the sender’s address character by character and it was identical, because it was identical: the message had been sent from the project manager’s own mailbox.

That is the point at which the ordinary approach runs out. Every additional minute spent examining the message was a minute spent playing a game the attacker had already won, because the message had been written to survive exactly that examination. There was nothing left in it to find.

The rule the practice had written two months earlier does not ask anybody to judge a message. It asks a question about a category (is this a change of bank details, is this over two lakh) and it has the same answer whether the email is genuine or not. Priya’s discomfort was that applying it felt like an accusation. It was not: it was a process that fires on the honest instructions too, which is the only reason it can be applied by a junior architect on a Thursday without anybody’s permission.