In short: AI is a tool, and tools earn their place by returning more than they cost. Most businesses never work out whether theirs does, because the obvious cost is a small monthly subscription and the real costs are invisible: the time spent fighting it to get a usable answer, the time spent checking the answer, and the occasional wrong output that nobody caught. Get those three wrong and you can spend a year feeling productive while going backwards.
A tool is judged on what it returns
I have watched a lot of technology come through small and medium businesses over thirty-two years, and the ones that stuck were never the impressive ones. They were the ones where someone could tell you, in a sentence, what it saved.
AI is being bought differently. It is being bought because everyone is buying it, which is the same reason businesses bought websites nobody read in 1999 and apps nobody downloaded in 2012. Impressive is not a business case. The question has not changed and it is not a complicated one: what does this return, against what it costs.
The trouble is that the costs are not where people look.
The three costs nobody counts
The subscription is the smallest one. Twenty or thirty dollars a seat a month is real money but it is rarely what decides this. If the only thing on your side of the ledger is the licence fee, you have not finished the sum.
The time spent fighting it. This is the big one and almost nobody measures it. Someone asks a vague question, gets a vague answer, rephrases, gets something closer, rephrases again, gives up and writes it themselves. Forty minutes gone on a task that would have taken twenty-five. It happens constantly and it never shows up anywhere, because nobody logs a timesheet entry called “arguing with a chatbot”. Multiply it across a team and it is a meaningful number that no one has ever added up.
The time spent checking it. This is the one that actually decides whether AI pays, and I will come back to it because it deserves its own section.
And then the fourth, which is not really a cost so much as a risk: the wrong answer that got through. AI does not fail loudly. It produces something fluent, confident and plausible, and if nobody who knows better reads it carefully, it goes out. We have measured this on our own systems, where models given a hostile instruction produced professional-looking recommendations that would have destroyed a machine’s ability to recover. In a business the equivalent is a quote with the wrong figure in it, or an email to a client that is subtly not true. The cost of that is not on any invoice, but it is not zero.
The checking cost is what decides it
Here is the test I use, and it cuts through most of the argument.
If verifying the output takes as long as producing it would have, the return is zero. If it takes longer, which happens more often than people admit, the return is negative and you have bought yourself a slower process with extra steps.
So the useful question for any task is not “can AI do this”. It nearly always can, to some standard. The question is how expensive is it to tell whether it did it right.
That single question sorts almost everything:
Where the return is large. High volume, repetitive, and easy to check at a glance. First drafts you were going to rewrite anyway. Summarising documents you will skim regardless. Restructuring or reformatting something that already exists. Extracting data from files where a wrong figure is obvious. Answering the same customer question for the four hundredth time. In all of these, a bad output is cheap to spot and cheap to discard.
Where the return is thin or negative. Low volume, high stakes, hard to verify, or requiring context that only lives in your head. Anything where being subtly wrong is expensive and being subtly wrong is hard to detect. Anything you would have to read three times to trust. If checking it properly means reconstructing the reasoning yourself, you have not delegated the work, you have added a step to it.
That is not an argument against AI. It is an argument for pointing it at the right half of your business.
The same tool, wildly different returns
Two businesses buy the same subscription. One gets a return that pays for it many times over. The other concludes AI is overrated. The tool was identical.
The difference is nearly always the operator. Someone who supplies the actual document, states what they want, gives an example of the output shape and knows enough to judge the result gets usable work back the first time. Someone who types a vague question and hopes gets a generic answer, spends twenty minutes trying to fix it, and forms a permanent opinion about the technology.
I have written separately about what that skill actually looks like, because it is learnable and it is the highest-leverage thing most people could change. But for the purposes of the sum: the operator is a variable in the ROI calculation, and usually the largest one. Any return figure quoted to you without reference to who is doing the work is close to meaningless.
How to actually work it out
Pick one task. Not the whole business, one task that happens often enough to matter.
- Time it as it is now. Honestly, including the interruptions. Most people underestimate their own routine work by a third.
- Put a rate on that time. Yours or your staff member’s, whatever is genuine.
- Do it with AI for a fortnight. Not once. A single good result proves nothing and a single bad one proves nothing either.
- Time the whole thing, including the checking and the redos. This is the step people skip and it is the step that contains the answer.
- Compare, and be willing to hear that it was worse.
Then repeat for the next task. You will very quickly find that the answer is not “AI is worth it” or “AI is not worth it”. It is that AI is strongly worth it for four things in your business, marginal for three, and actively counterproductive for two. That is a far more useful conclusion than either of the ones people arrive at without measuring, and it is the only one you can act on.
The honest version
Some businesses should spend a lot on this. Some should spend very little. A café mostly needs coffee made and staff rostered, and the truthful recommendation there is a little help with ordering and paperwork. A trades business living in quotes and follow-ups, or a practice buried in bookings and reports, can hand over a genuinely large share of that work and feel it within weeks.
Both are correct answers. The wrong answer is a number that does not match your actual problem list, in either direction, arrived at without anyone doing the sum.
If you want that sum done properly rather than guessed at, that is what an AI opportunity audit is: we work out which tasks in your business would genuinely return more than they cost, put a rough dollar figure against each one, and give you a written roadmap you own. Sometimes the finding is that you should not spend much at all yet, and we will tell you that, because a report that says so has saved you more than one that does not.
Drafting and structure: AI, working from my notes and client work, and disclosed openly as with everything on this site at /experiment.