Automate first the task where a wrong output is cheapest to catch, not the one that takes the most hours. A good first job has messy input, a result someone already reads before it goes anywhere, and mistakes that would be obvious when they happen: drafting a quote from a call note that you check before sending, rather than reconciling accounts that nobody re-checks. The reason is that you cannot tell by looking which tasks a model handles well. In a 2023 field experiment with 758 Boston Consulting Group consultants, GPT-4 made people faster and better on tasks inside its capability, and made them 19 percentage points less likely to reach the right answer on a task that looked similar but was not. Your first automation is where you find out which kind of task you have, so pick one where finding out is cheap.
Ask what to automate first and you will mostly be told to list your repetitive tasks and start with the one that costs the most time. It sounds like arithmetic. Twelve hours a week on invoices, two on quotes, so invoices first.
The hours are real, but they answer the wrong question. They tell you the size of the prize if the system works. The decision in front of you is a different one: which job can you hand over and still notice, quickly and cheaply, when the output is wrong?
You cannot see from outside which tasks AI does well
The clearest evidence comes from a pre-registered experiment run with Boston Consulting Group and published as a Harvard Business School working paper in September 2023 (Dell’Acqua, Mollick, Lakhani and colleagues, Working Paper 24-013). On a set of 18 realistic consulting tasks that sat inside what GPT-4 could do, consultants with access completed 12.2% more tasks, finished 25.1% faster, and produced work rated more than 40% higher in quality than a control group.
The study also included a task chosen to sit outside it. It was built on a spreadsheet designed to look comprehensive, plus a file of interviews with company insiders that held the detail needed to read the numbers correctly. Without AI, about 84.5% of consultants got it right. With AI, 60% and 70% did in the two AI conditions. The authors’ phrase for this is a “jagged frontier”: tasks that seem similar in difficulty land on opposite sides of it, and nothing on the surface tells you which side you are on.
That failing task looks a great deal like a small business. The figures in your system look complete. The thing that changes what they mean (the customer who always pays late but always pays, the supplier whose price list is out of date) lives in someone’s head or an email thread. A model reading only the figures gives you a fluent, confident answer, and the confidence is not a signal of anything.
Three tests for a first job
If you cannot know in advance whether a task is inside the frontier, the first job should be one where being wrong is survivable and visible. Put each candidate through three questions.
- Does a person already look at the output before it matters? A quote you read before it goes to the customer has a check built in. An overnight reconciliation that only gets opened when something breaks does not.
- Would a wrong output look wrong? A quote with the wrong unit price is easy to spot. A summary that quietly leaves out the one clause that mattered reads exactly like a correct one.
- Is the input messy enough to need a model at all? Free-text call notes, emails written in five different styles and scanned delivery notes are model work. A form that already has the right fields is not.
The job that passes all three is rarely the biggest one on your list, and that is the point. Working out which of your slow jobs pass is most of the decision, and it is the part people skip because the expensive job is the one that hurts.
Time it before you build, because you will misjudge the after
The second reason to start small is that people are poor judges of whether AI is helping them. In a randomised trial METR published in July 2025, 16 experienced open-source developers worked through 246 real issues in code they knew well. Before starting they expected AI tools to make them 24% faster. Measured, they took 19% longer with the tools. Afterwards, they still believed the tools had made them about 20% faster.
The limits matter here. That was software, with early-2025 tools, and a small sample, and METR says plainly it does not show that AI slows most people down. Its February 2026 follow-up says developers are probably faster with current tools, but that its own new data is only very weak evidence of how much. What the study does show is the thing that matters for you: the people doing the work could not feel which direction the change went.
So before anything is built, write down for two weeks how long the job takes, how often it comes out wrong, and who catches it. Without that, the only measure of success you will have is how the new system feels to use, which is the measure that study shows to be unreliable.
When the right answer is not AI
A fair amount of what turns up on an “automate this” list does not need a language model. Reminder emails sent on a fixed schedule, a quote built from a price list, an invoice chased at fourteen and thirty days: these are rules. A template, a spreadsheet lookup or the scheduling feature already in your accounting software will do them more cheaply, the same way every time, and when they break they break loudly rather than fluently.
There is also a cost to the advice above, and it should be said plainly. The job that is cheapest to check usually saves fewer hours than the job that hurts most. For the first month you are trading some of the saving for information: whether the system holds up on your material, what its mistakes look like, and whether the people using it actually check its output. That information is what tells you whether the bigger job is safe to hand over at all.
Common questions
- What should a small business automate first with AI?
- The task where a wrong output is cheapest to catch: one with messy input such as call notes or emails, where a person already reviews the result before it reaches a customer, and where a mistake would be obvious. Drafting quotes or replies for review is a typical example. Start there even if a larger task costs more hours, because the first job is how you learn whether AI holds up on your material.
- Should I automate the task that takes the most time?
- Not first. Hours tell you how much a working system would be worth, not whether it will work or how you would notice when it fails. A 2023 Harvard Business School and Boston Consulting Group experiment found AI helped consultants substantially on some tasks and made them less accurate on one that looked similar, so it pays to find out on a task where errors are cheap.
- How do I know if AI is actually saving time?
- Measure the task before anything is built: time taken, how often it comes out wrong and who catches it, over a couple of weeks. Do not rely on how it feels afterwards. In METR's 2025 trial, experienced developers believed AI had made them about 20% faster when they had in fact taken 19% longer.
- Which business tasks do not need AI?
- Anything where the input is already structured and the rule is fixed: scheduled reminders, quotes calculated from a price list, invoice chasing on set dates. A template, a spreadsheet formula or a built-in feature of your existing software is cheaper and more predictable, and it fails visibly rather than producing a plausible wrong answer.