Between the promise that AI changes everything and the suspicion that it changes nothing, there is a narrow band of work where it genuinely pays for itself. Here is how to find it.
There are two unhelpful positions on AI in business. One is that it changes everything and you are behind. The other is that it is a bubble and you should wait. Both are easier to hold than the truth, which is that AI automates a specific and identifiable kind of work extremely well, and most other kinds badly or expensively.
Knowing which is which is worth more than any particular tool.
The tasks where AI reliably pays for itself share four properties. The more of them a task has, the better the case.
The input is messy and the output is structured. Turning a stack of unformatted invoices into rows of data. Reading support emails and sorting them by what they are actually about. Extracting the terms from a contract. Software has always struggled with unstructured input; this is the thing that genuinely changed.
Being roughly right is useful. If a suggestion that is correct 90% of the time saves a person real effort, and the remaining 10% is caught easily, the economics work. If the 10% is expensive or invisible when wrong, they do not.
A person stays in the loop. The most valuable deployments draft, sort, summarise and suggest — and a person approves. This is not timidity. It is where the return actually is, because you get the speed without owning the failure mode.
The task repeats. Automating something that happens two hundred times a week is a business case. Automating something that happens twice is a hobby.
Work where being wrong is expensive. Anything with a legal, financial or safety consequence needs review, and if review takes as long as the original task you have moved the cost rather than removed it.
Work that needs one right answer. These systems are probabilistic. Ask twice, and you may get two defensible but different answers. For drafting that is fine. For anything that must reconcile, it is a liability.
Work where the rules are actually stable. If a decision follows rules you can write down, write them down. A conditional statement is cheaper to run, impossible to misinterpret, and does not need monitoring. A surprising amount of what gets proposed as AI is a form and three rules wearing a costume.
Work nobody had defined. Automation makes a process faster; it does not make an undefined process coherent. If two people in your company do the same task differently and both are right, automating it will produce something that is confidently wrong in a new third way.
Which task in this business consumes hours of skilled attention, produces a predictable kind of output, and would be no worse if it were 90% right and checked?
That question finds the opportunities. It is not a technology question, and the person who can answer it usually does not work in technology — they work in the department drowning in the task.
The projects that fail tend to start at the other end: a decision to use AI, followed by a search for somewhere to put it.
The deployments we see working are unglamorous:
Intake and routing. Everything arriving in a shared inbox, read, categorised and sent to the right person with the relevant history attached.
Extraction. Documents in, structured records out, flagged for review where confidence is low.
Drafting. First versions of replies, summaries and reports, written by a machine and finished by a person. The blank page is the expensive part.
Search across your own material. Answering questions from years of documents that nobody can realistically read.
None of these will appear in a keynote. All of them return hours every week to people who are expensive and busy.
The build is not the expensive part. Running it is.
These systems need monitoring, because quality drifts as inputs change and nothing alerts you. They need a fallback for when the service they depend on is unavailable or changes underneath you. They need someone who owns them — an automation nobody is responsible for degrades quietly until it is producing nonsense that people have learned to work around.
And they need honesty about what they are. If a customer is reading something a machine wrote, decide deliberately whether to say so. Increasingly, this is not only a courtesy.
Pick one task. Make it a real one that someone complains about, not a demonstration. Measure how long it takes today, honestly, before you change anything.
Build the smallest version that does that one task with a person approving the output. Run it for a month. Compare it against the number you wrote down.
If it saved time, widen it. If it did not, you have learned something specific about your business for a small cost — which is a considerably better outcome than a platform decision made on a slide.
If you want a second opinion on whether a particular task is worth automating, describe it to us. We will tell you when the answer is no, which happens more often than the market suggests.
Tell us what you are building and we will scope it with you.
