AI is not only for technology companies

You quote a job from experience. You make it. Somewhere near the month end, the numbers arrive and you learn whether that quote was any good.

The problem is not the arithmetic. Costing a finished job is straightforward. The problem is that the next quote is still made from memory, because nothing connects what a job actually consumed to what similar jobs will consume.

What to try first, before any software: for one month, record actual material and actual hours against ten jobs you have already quoted. Compare. Most factories find the gap sits in one operation, not spread across all of them. If it does, you have a specific problem worth solving rather than a costing system worth buying.

A photographed purchase order. A voice note. A challan written by hand. An email with the quantities buried in the third paragraph. Someone in your office retypes all of it into the system, and that is where the wrong quantity enters your business.

This is the clearest example of what AI is actually good at. Reading things nobody structured is the job it does well, and order intake is that job in its purest form.

Two honest limits. A person must still confirm before dispatch, because the failure is a wrong quantity rather than an obvious crash. And it will not fix a customer who orders ambiguously — it will just record the ambiguity faster.

Start by counting how many orders arrive in a week, and in how many formats.

Traders, distributors, manufacturers. Two systems, two numbers, and an argument every month about which one to believe.

Software is blamed for this more often than it deserves. Most of the gap comes from timing and habit rather than code. Goods arrive before the paperwork does. A return is recorded on one side and not the other. Samples and breakages leave the building without leaving a record. A transfer between locations is counted at one end only. None of that is a system fault, and replacing the system does not fix any of it.

Before buying anything, take one fast-moving item and trace every movement of it for a fortnight, by hand. The cause usually reveals itself within days, and it is usually a process nobody wrote down rather than a system that needs replacing.

You have an ageing report. It tells you who is late. It does not tell you the thing you actually need, which is who is drifting.

A customer who paid in thirty days last year, forty-five this year and sixty last quarter is a different problem from one who has always paid at ninety. The first is a relationship changing. The second is a payment term you never agreed to but have been honouring anyway.

That drift is a pattern already sitting in your ledger. Nobody has the hours to go and look for it, which is precisely the kind of work worth automating.

Start by sorting last year's payments by customer and by days taken. The names that surprise you are the list.

Nobody is idle. Everyone is busy. And if someone asked you what consumed last week, you could not answer with anything better than an impression.

Time hides in four places, and they are the same four in most companies. Re-keying, where the same figure is typed into a second system. Chasing, where somebody waits for an answer and follows up three times. Checking, where work is verified because it has been wrong before. And waiting, where a decision sits with one person for two days.

Pick the process that annoys you most. For a fortnight, ask the people doing it to note how long each of those four takes. That measurement costs nothing and tends to end the argument about where to start.

You can see what you sold. You cannot see which products still make money once discounts, freight and returns are counted, which customers are quietly ordering less, or what you should stop doing altogether. And last month's numbers reach you halfway through this one.

The usual response is to commission another dashboard, which is why most companies have several nobody opens.

A number is only worth having if a decision changes when it moves. So start from the decision. Write down the three decisions you make every month, then ask what number would settle each. Usually there are three or four, they already exist somewhere in your systems, and the reason you do not have them is that nobody has been asked that question.

Enquiries arrive. Money goes out on getting them. Nobody can connect the two, particularly when half the enquiries come by phone, by WhatsApp or from someone who was told about you.

Attribution software is built for businesses where every customer arrives through a browser. Yours does not.

The cheap version works and almost nobody does it: ask every person who enquires how they heard of you, record the answer in the same place every time, and review it after three months. That is a fortnight of discipline, not a project.

Build something only once that record exists and has told you which sources are worth tracking properly. And accept early that some spending can never be honestly attributed. A system that pretends otherwise will send you somewhere confidently wrong.

Three companies, three sets of books, three GST registrations. Somebody assembles a group view in a spreadsheet, late, with caveats, and by the time it reaches you it describes a month that has already ended.

There is a trap in the obvious fix. Merging the ledgers into one system so everything is visible at once creates a tax exposure rather than a dashboard: an internal record showing group turnover well above what a single registration reports is exactly the inference you do not want an officer to draw.

The books stay separate. What you want is a read-only layer above them that reads each set and reports without writing anything back. The interesting numbers there are usually operational rather than financial.

Everything routes through them. None of it is written down. They are going on leave in March, and you have been aware of this for two years.

This is a software problem before it is a staffing problem, and the first step is not automation. It is finding out what is actually in their head, which is always less mysterious and more specific than anyone expects.

Sit with them for two hours and write down every decision they make in a week. Then sort the list into three: decisions that follow a rule they can state, decisions that come from recognising a pattern across many cases, and decisions that need judgement about a person or a relationship.

The first group is a rule. The second is worth a model. The third should stay with a human, and often should simply be taught to a second one.

The product does ninety per cent of what you need. The remaining ten per cent is the part your business is actually built on, and no vendor will build it.

The wrong response is to commission a replacement for the whole thing. Compliant payroll in India starts free below twenty-five employees and runs at around ₹40 a head after that, as of mid-2026, against custom systems quoted in lakhs. That arithmetic is not close, and it holds across accounting, ticketing and most other categories.

So buy the platform and build the gap. One narrow piece of software that does the thing no product does, connected to the system you just bought.

The test: if a product covers ninety per cent, buy it and build the ten. If it covers forty, you have a process nobody has productised, and that is a genuine build.

Patients find you through a listings app. It works, and it is not yours. The platform holds the enquiry, the reviews and your position in the results, and it can change any of them without asking you.

There is also a regulatory dimension. The National Medical Commission (Registered Medical Practitioner) Regulations, 2023 restrict practitioners from participating in platforms that carry patient ratings, reviews and promotion. Your own website is the channel your council is most comfortable with.

A practice site can do real work: book, remind, and collect history, photographs and concerns before the patient arrives, which saves chair time on the day.

It must be built the right way. A tool that grades a patient's smile or tells them what treatment they need is an outcome claim and a solicitation. The same tool, producing a summary for you and a list of things to discuss for them, is neither.

Physiotherapists, tutors, salons, veterinary clinics. Two problems, and they look similar from a distance.

The first is no-shows, which cluster. Particular days, particular times, particular gaps between booking and appointment. Once you can see the pattern you can overbook the right slot rather than every slot.

The second is dormancy. People who should have returned at six months are at sixteen, and the records already know who they are. Nobody has the hours to go through them each week.

Work out two numbers before spending anything: what an empty slot costs you, and how many patients or clients have not returned when they should have. Multiply. That figure is your budget, and it is usually larger than people expect.

The ICAI Code of Ethics 13th Edition took effect on 1 April 2026. Factual, informative promotion is now permitted for firms. Solicitation, comparative claims and claims of superiority are not.

That has left a lot of practices unsure what their own website may now do, and the advice circulating is mostly written by people with no obligation to be right.

The practical question is narrower than the debate suggests. What may your site say, what may it collect, and which of the tools being sold to you sit on the permitted side of the line between telling people what you do and soliciting them.

We build the software rather than advise on ethics, so confirm anything specific with the Institute or your own adviser. But the software choices follow the rule closely, and most of them are decidable.

Effective date verified 1 September 2026.

Consultants, designers, architects, lawyers, coaches. The work gets done. Some of it never reaches an invoice.

It leaks in four places. Hours nobody recorded on the day. Scope that grew across three emails without a conversation about money. Invoices raised late, or in a batch at month end. And follow-up on unpaid ones that nobody enjoys and everyone postpones.

Only the first is really a software problem, and even then, calendars, sent mail and file histories already hold most of the evidence needed to reconstruct a defensible record.

Before buying anything, reconstruct one past month honestly and compare it to what you invoiced. If the gap is small, this is a habit. If it is large, it is worth building something.