Ask your software "which items are moving slowly?" and get an answer in plain English. Speak the question instead of typing it. Drop a supplier invoice in and let it pick out the supplier, the items, the quantities, the tax and the total. Ask why sales fell last week and get an attempt at an explanation.

All of that is genuinely useful, and we are building towards it.

There is one part of the conversation that usually goes missing:

Every time AI has to read, listen, interpret, reason or generate an answer, somebody is paying for that work.

Why this is different from ordinary software

Traditional business software has a particular economic shape. Once it is built, opening yesterday's sales report for the hundredth time costs the software company essentially nothing. The work was done once, at build time.

A language model is not like that. The question you type, the business data handed to the model, the documents it reads and the answer it produces all contribute to usage. More capable models cost more than simpler ones. Voice, images, long documents and long conversations all add to it.

None of which is an argument against AI in business software. It is an argument for putting AI where it earns its keep, and for being honest about where it does not.

"Show me today's sales" does not need intelligence

An owner asks: what are today's sales?

That number already exists. The software can calculate it exactly with an ordinary database query and show ₹1,84,720 — correct to the rupee, instantly, at no processing cost. Sending invoice rows to a language model so it can hand back the same figure is adding intelligence where none was required, and doing it less precisely.

The same is true of most of the questions owners ask every day:

• What is my closing stock?
• Which customers owe me money?
• What did we collect today?
• What are my top ten selling items?
• Which invoices are overdue?
• What did Branch 2 sell yesterday?

Those are database questions and software should answer them directly.

AI starts to earn its place when the question shifts from what happened to something more like what should I be paying attention to. For example:

Sales look similar to last month, but there is less money in the bank. What changed?

Now several things may have to be connected: receivables, purchases, stock that has built up, supplier payments, margins, returns. Nothing in the database answers that on its own. That is a good place to spend AI.

Voice is convenient, and the meter is running

Voice can make software much easier for an owner who does not want to navigate ten reports. Instead of finding the right screen, he asks:

Which customers promised payment this week and still have not paid?

Useful. But that one sentence may involve converting speech to text, working out what was meant, fetching the right business data, handing the model enough context to answer, generating the answer and possibly reading it back aloud.

One question will not trouble anybody. Hundreds of questions a day, across hundreds of customers, is a different proposition.

The answer is not to avoid voice. It is to make the software do the cheap, exact work first and ask the model to do only the part that genuinely needs language. The database finds the overdue customers. AI turns that into a sentence.

Reading supplier invoices, compared honestly with typing

Invoice reading looks close to magic in a demonstration. Upload a supplier invoice and the system picks out the supplier, the invoice number, the items, quantities, rates, tax and total. Where purchase invoices are long, that saves real time.

It is worth comparing the whole process honestly, though.

A person entering an invoice has a labour cost and makes occasional mistakes. AI reading an invoice has a processing cost and also makes mistakes. An unclear photograph, an unusual layout, an abbreviated item name, a free quantity, an odd tax treatment or a badly printed figure will still send somebody back to the original document.

So AI does not always remove data entry. Often it changes the work from enter everything to let AI enter it, then check what matters.

That is an excellent trade on an invoice with 150 lines. It is a poor one on an invoice with four. The question is not "can AI read this invoice?" — it can. The question is whether using it here saves enough human effort to be worth the processing cost and the checking that still has to happen.

The expensive version of a report you already had

There is a category worth being careful about, which is analysis for the look of it.

An owner asks for an AI analysis of his dead stock. The software already knows the purchase date, the last sale date, the quantity, the value and the movement. An ordinary report can identify everything with no sale in 90, 180 or 365 days instantly and exactly. Running all of that through a model every morning to produce a nicely worded paragraph may feel advanced, but the business answer never needed AI. That is where it quietly becomes decoration.

The better arrangement is that the software finds the exception and AI explains it.

The report determines that 23 items have not moved in 180 days and that ₹3.8 lakh is sitting in them. Then AI can help with the question that actually needs judgement:

Given recent sales, the season and what we paid for these, which should I discount, which should I move to another branch, and which should I simply stop buying?

That is worth paying for. Reproducing a report is not.

The opposite mistake

It would be just as wrong to become so wary of the cost that useful intelligence gets refused.

If one piece of analysis prevents an unnecessary ₹50,000 purchase, surfaces a margin leak, flags a collection that needed chasing, or catches a stock problem a month earlier than anyone would have noticed, the processing cost of that analysis is trivial against what it saved.

So the aim is not the least possible AI. It is the most useful intelligence per rupee spent on it. Those are not the same target, and the difference matters.

How we think AI belongs in Stock2Track

Our position is that AI should sit on top of reliable data rather than replace the engine underneath it.

Stock2Track should go on calculating sales, tax, stock, margins, receivables and reorder levels with ordinary business logic, because those must be exactly right and a language model is the wrong tool for arithmetic the system already knows how to do.

AI then helps with the layer above: understanding a question asked in ordinary words, connecting information sitting in different parts of the business, summarising what deserves attention, explaining something that has changed, reading documents, suggesting what to do next.

Where a normal report answers the question accurately, the normal report should answer it. That makes the system faster, more predictable and cheaper — and it avoids a risk that matters more than cost, which is asking a model to work out something the software already knows precisely.

There should probably be an AI budget, like an SMS budget

Businesses already understand services that are charged by use. SMS costs money. WhatsApp conversations can. Payment gateways take a cut. Cloud storage is billed by what you store.

AI is likely to become another of those. For many shops a small monthly allowance will be more than enough. A larger business running invoice extraction, voice conversations and frequent analysis will use considerably more.

Costs will keep moving. Competition, smaller specialised models, caching and models that run locally may well make a lot of this cheaper over time. But software being designed now has to treat AI usage as a real operating input rather than free computing.

The customer should be able to see it, too. Where a job can be done without AI, that option should stay available. Where AI usage could become significant, the software should show that plainly rather than burying it in a bill nobody can explain.

Four questions before adding AI to anything

1. Can ordinary software already answer this accurately?
2. Is AI saving meaningful human effort?
3. Is AI finding something that would be hard to discover otherwise?
4. Is the value greater than the processing and the checking it still needs?

If the honest answer is that we have made an existing report talk in a more fashionable way, we have not improved anything. If AI removes an hour of work, connects something nobody would have connected by hand, or brings a real problem forward before money is lost, it is doing the job.

That is the balance we are aiming for. Not AI everywhere — AI where intelligence is actually required.