Schedule a call

The Five Places AI Actually Pays

Jul 13, 2026

Across every business we have worked in, the returns cluster in the same five places. All five share one property, and knowing what it is tells you where to look in your own.

The results are not evenly spread. Most of what gets attempted produces very little, and the things that work keep turning out to be variations of the same five.

They share one property, which is the useful part: each was blocked because a step needed judgement, judgement meant a person, and a person could not do it at volume.

What the adoption data suggests

Statistics Canada, in a release dated 16 June 2025, found that among Canadian businesses using AI the leading applications were text analytics at 35.7%, data analytics at 26.4%, and virtual agents or chatbots at 24.8%.

Two of the top three are reading and interpretation rather than generation. That matches what we see: the durable wins are in understanding what arrives, not in producing more output. It is unglamorous, and it is where the hours are.

One: reading what arrives

Documents in any format, emails, photographs of paperwork, handwriting, forms filled in by hand. Extracting what matters and writing it into the system that runs the business.

This is the largest category by a distance. Every business receives unstructured material and every business has somebody keying it in. The work is high volume, low judgement per item, and it was completely unautomatable until recently because the input varies infinitely.

Invoices, delivery notes, certificates, applications, site photographs, supplier statements. If it arrives as a picture or a paragraph and leaves as a field in a system, it belongs here.

Two: judgement at volume

Which of four hundred enquiries deserves attention first. Whether these two records are the same client. Which contract contains an unusual clause. Which of forty invoices does not match its order.

Each decision is easy for a competent person and takes thirty seconds. Multiply by four hundred and it becomes somebody’s week, so it either does not get done or it gets done badly under time pressure.

This is where the largest quality improvements sit, because the alternative was not a person doing it well. It was a person doing it for the first twenty cases.

Three: contact with the whole book

Not the twenty accounts somebody remembers. All four hundred, written from what is actually in each file.

Every business has a book it cannot service properly: dormant clients, old enquiries, past candidates, renewals eighteen months out. Personalised contact at that scale was never economic, so it happened for the top few percent and nothing else.

The return here tends to be revenue rather than saved hours, which makes it the easiest of the five to justify.

Four: noticing what changed

The client who has gone quieter than usual. The job drifting late. The margin that moved for a reason nobody has identified. The system whose exception rate has tripled.

These require comparing current behaviour against a baseline for every account or process, continuously. No person holds four hundred baselines, so problems get discovered when somebody complains.

Detection is worth more than most efficiency gains, because the cost of finding out late is usually a client rather than an hour.

Five: whatever waits on one person

The approval, the answer, the quote, the specification only one individual can produce.

That person is your capacity ceiling, and the work waiting on them is not usually work that needs them. It needs somebody with their context, which is a different thing, and context is what can now be supplied.

Everything upstream of the decision can be prepared: the file assembled, the options set out, the draft written. What is left for them is the judgement, which is a few minutes instead of an hour.

The test

Take anything you are considering and ask whether it fits one of the five. If it does not, be sceptical, whatever the demonstration looked like.

The projects that produce nothing are consistently the ones aimed elsewhere: a capability bought without a process attached, a chatbot answering questions nobody was asking, or a faster version of something already efficient that nobody was complaining about.

Where the returns are smallest

Worth naming, because these are the projects most often proposed and they consistently disappoint.

Producing more content or output. Volume was rarely the constraint, and more of something nobody was short of does not create value. The exception is where output was genuinely blocked by one person’s availability, which is really case five.

Chat on a website. Occasionally useful, frequently answering questions nobody was asking, and it is the version of AI most often shown to businesses that leads them to conclude none of this applies to them.

Anything replacing a process that already worked. A marginally faster version of something efficient returns almost nothing and carries all the cost and risk of a build.

None of these are harmful. They simply do not repay the effort, and they consume the first attempt, which for a business with one attempt in it is the real cost.

Pick the one that annoys everybody

Where two candidates look equally good on the arithmetic, choose the one your team dislikes most.

A first project that removes work nobody wanted to do makes the argument for the second project without anybody having to be persuaded. A first project that is technically better but touches something people were content with produces a shrug, and the next proposal gets a harder hearing.

The order matters more than most owners expect, because the second automation is always easier to fund than the first.

What they have in common operationally

They happen daily. Anything monthly rarely repays the build.
They currently wait on a person being available rather than on a decision being difficult.
The input is messy, which is why conventional software never solved them.
There is a right answer most of the time, and a person can check it.

Four conditions. Anything clearing all four is very likely to work, and most businesses have three or four candidates they could name in a minute if somebody asked.