What AI is actually doing to businesses — and what it costs to wait

Every business owner I speak to now asks some version of the same question: is AI something we actually need to do anything about, or is it another cycle of noise that will settle down on its own?

It is a fair question. Most people asking it have lived through several rounds of technology that was going to change everything and mostly changed the invoice. So let me answer it the way I would answer any other question about a business: by costing it.

What has actually changed

Strip away the language and one thing has moved. The cost of a certain kind of work has collapsed.

Specifically: work that involves reading something, understanding it, and producing a first draft of a response. Summarising a long document. Extracting figures from a hundred invoices. Drafting a reply that follows a house style. Classifying a mailbox. Turning a rambling client email into a structured brief.

None of that is new work. Businesses have always done it. What has changed is that it used to cost an hour of somebody's time and now it costs a few pence and thirty seconds, with a person checking the output rather than producing it.

That is the whole event. Not intelligence, not replacement — a step change in the unit cost of a specific and very common category of work.

Where it is already showing up

In the businesses I work with, three patterns come up repeatedly.

The first response. Enquiries, quotes, support questions. The gap between a customer asking and a business replying used to be structural: someone had to be free. It no longer has to be. A competitor answering in two minutes with something accurate beats a business answering in two days with something considered, more often than most owners would like to believe.

The data-entry layer. Every business has one — the person or the afternoon spent moving information from one format into another. Invoices into the ledger, forms into the CRM, PDFs into a spreadsheet. This is the single most immediately affected category of work in most businesses, and it is usually the one nobody has ever measured.

The document pile. Contracts, policies, correspondence, records. The cost of asking a question across a large body of documents has fallen from "someone spends a day" to "someone asks and reads the answer". Any business whose knowledge sits in files rather than in a system has just been handed a way to use it.

The part that actually costs money

Here is where I would push back on how this is usually discussed. The risk to a business that does not adopt is rarely the dramatic one. Nobody is put out of business by a chatbot.

The risk is margin, and it compounds quietly.

Consider two competitors of similar size. One reduces the time spent on quoting, first-line response and data entry by a third. They have not fired anyone. They have simply released capacity, and that capacity goes into work that wins revenue. The other keeps doing it the old way, at the old cost.

In year one the difference is barely visible — a few points of margin, easily attributed to something else. By year three the first business is quoting faster, following up more, and carrying the same overhead against more revenue. The second is not losing because it made a bad decision. It is losing because it never made one.

That is the actual cost of waiting: not a catastrophe, but a slow widening you cannot see on your own P&L, because your own P&L looks much as it always did.

There is a second cost, and it is the one I would watch more closely. Your competitors are not the only ones adjusting. Your customers are. People now expect an immediate, specific answer, because they get one everywhere else. A business that takes three days to respond is not being judged against its own past performance. It is being judged against the fastest reply that person received this week.

Where AI is the wrong answer

I would be doing my job badly if I stopped there, so — the other side.

AI is genuinely poor at anything where being confidently wrong is expensive. Anything requiring a defensible audit trail. Anything where the judgement is the product rather than the drafting. It will produce a plausible answer when it does not know, which is a specific and dangerous failure mode in a regulated business, and the whole point of a professional is being able to say "I am not sure".

It is also poor value where the underlying process should not exist. Automating a report nobody reads is worse than deleting it, because now it is cheap enough that nobody will ever question it again. This is the mistake I see most often, and it is not new — it is the same mistake businesses made with every previous wave of automation, arriving faster.

And it does not fix a business whose problem is that its data is a mess. If three systems disagree about who your customers are, adding a layer that reads all three produces confident nonsense, at speed.

What I would actually do

Not a strategy. Three things, in order.

Measure one process before you touch it. Pick the one that consumes the most hours and returns the least judgement — usually the data-entry layer. Find out what it genuinely costs in hours per month. Most businesses have never done this, which is why they cannot tell whether any change helped.

Change one thing, and only one. Something small, reversible, and measurable. The purpose of the first project is not the saving. It is finding out how your business responds to this kind of change, cheaply, before anything larger is committed.

Then decide with the number in front of you. If it paid for itself, do the next one. If it did not, you have learned something specific rather than acquiring a subscription and a feeling.

The honest summary

AI is not going to end your business, and it is not going to run it. What it has done is make a particular class of work dramatically cheaper — and businesses that do not take that cost out will carry it while their competitors do not.

That is not a technology decision. It is a margin decision, and it should be costed like every other one. The businesses that will struggle are not the ones that adopt slowly. They are the ones that never look at the number at all.