You run a factory, or a trading business, or a practice with six people in it. Every article you have read about AI was written for someone building a product. So you have reasonably concluded that this is a thing technology companies do to each other.
That conclusion is wrong, and it is expensive. Not because you are missing a revolution — most of what is promised will not happen — but because there is probably one job in your week, done by a person, that a machine could have done while everyone slept.
The misunderstanding
AI is sold as a technology. It is easier to think about as a category of work.
Specifically, four kinds of work. Read them and think about who in your business does them.
Reading things nobody structured. Invoices in a dozen layouts. Enquiries on WhatsApp. Handwritten delivery challans. Emails with the important sentence in the middle. Until recently, software could only read things that arrived in the shape it expected. That is the change. Everything else follows from it.
Sorting and matching. This invoice belongs to that purchase order. This enquiry is a complaint, not an order. This payment is against those three bills. Somebody in your office does this, and they are good at it, and it is the least interesting thing they do.
Noticing patterns in your own records. Which customers have quietly stopped ordering. Which stock moves three weeks before a season. Which day of the week your no-shows cluster on. The information is already in your system. Nobody has the hours to go and look.
Drafting. The first version of a reply, a summary, a report. Not the final version. The first one, so a person is editing instead of starting.
None of those four is technical work. All four are happening in your business right now, done by someone at seven in the evening.
And the four things it does not do
It does not decide anything you have not already defined. If two people in your company would answer a question differently, a model will not settle the argument. It will pick one of them at random, forever.
It does not fix wrong data. It makes wrong data faster and more confident. If your master list has the same customer three times under three spellings, start there and come back.
It does not replace the relationship. The reason your customer calls you and not the larger competitor is not a process. Do not automate it.
It does not know your business. Everything above depends on someone sitting down and explaining how the work actually gets done. That person is you, and it is the part nobody warns you about.
What it looks like in six businesses
Six situations, described plainly. Read them against your own week rather than as a list of features.
A factory making to order. Quotes are made from experience, jobs are made, and the month closes before anyone knows whether that quote was any good. Costing a finished job is arithmetic. The useful version is quoting the next one from what similar jobs actually consumed — material, machine time, the operation that always overruns — rather than from memory. What has to be recorded on the floor for that to work is less than most people fear.
A trading business taking orders in every format there is. A photographed purchase order. A voice note. A quantity buried in the third paragraph of an email. Somebody retypes all of it, and that is where the wrong number enters the business. Reading unstructured things into structured order lines is the clearest example of this work there is. A person still confirms before dispatch, because the failure mode is a wrong quantity rather than an obvious crash.
A distributor with stale reorder points. They were set three years ago and have not moved. Consumption has. The result is capital sitting in slow stock and stockouts on the fast lines, at the same time. A reorder point that follows actual consumption is not really an AI problem — it is arithmetic — but the forecasting layer above it is.
A founder with three companies. Three sets of books, three registrations, and no honest view of the whole. There is a trap in the obvious fix: putting everything into one ledger creates a tax exposure rather than a dashboard, so the books stay separate and a read-only layer reports across them.
A dental practice. The recall list is the asset and nobody works it. Patients who should have returned at six months are at fourteen, and the records already say who they are. Separately, no-shows cluster by day, by time, and by how long ago the appointment was booked. A practice that can see the pattern can overbook the right slot instead of every slot.
A CA firm at filing season. Documents arrive from every client at once, in every format a human can produce. Someone sorts them by client, by type and by year, then chases what is missing. The sorting is machine work and the chasing can be automatic. The judgement about what the numbers mean is neither, and is the reason the client pays.
Three ways to get it, and they cost very different amounts
Buy something. Most problems have already been productised by someone with a hundred engineers and a lower price than you could ever build at. In India you can run compliant payroll for free below twenty-five employees, as of mid-2026. Start here, always. If a product does ninety per cent of what you need, buy the product.
Build the gap. The thing no product handles. A shift pattern peculiar to your factory. An incentive calculation your industry uses and nobody else does. An approval that crosses two systems that will never talk to each other. This is small, specific software, and it is the honest case for a custom build.
Add to what you already run. Your existing system works and you are not replacing it. One capability goes in — reading the invoices, sorting the enquiries — and the rest is left alone. This is usually the cheapest of the three, and the one most often mis-sold as a migration.
How to tell whether you have a case
Four questions, in order. Answer them before you speak to anybody, including us.
Can you name the decision in one sentence? Not the process. The decision. If your sentence needs an "and", you have two projects and should do the smaller one first.
Who does it now, and how long does it take them? A rough weekly figure is enough. Multiply it out for a year. That number is your budget ceiling, and often it is smaller than people expect, which is useful to learn early.
Is the data that feeds the decision reliable? Go and look at fifty records. If they disagree with each other, you have a data project first. That is real work and sometimes it is the whole job.
What does it cost on the day it is wrong? An hour of somebody's time, or a lorry. This determines how much checking the thing needs, and checking is most of the cost.
When we would tell you not to bother
If the answer to the second question is under two hours a week, leave it alone. The build will cost more than the problem for years.
If the process is not agreed inside your own company, software will set the disagreement in concrete and charge you for it.
If the decision is the thing your customers value you for, keep making it yourself.
We put all of that in writing before quoting anything, because a firm that only ever finds reasons to build is not giving you an opinion. It is giving you an invoice.