Good Transformer
Is AI worth it for a small business? The honest maths
Most AI pilots deliver no measurable return, but a few deliver a lot. We explain the difference, and give a simple test for whether AI will actually pay for your business.
On the current evidence, most AI projects return nothing you can measure: MIT put it at 95%. For a small business, AI is worth it when you point it at a task that is frequent, costly and repeatable, and then redesign the work around it.
The honest reason most firms see no return is not that the tools are weak. It is that they buy a tool and change nothing about how the work is actually done. This post gives you a simple test for spotting the tasks that are worth it, before you spend.
If you are weighing whether AI is worth the money and the disruption, this is written for you. We will look at what the 95% figure really means, why most pilots return nothing, and what the small minority who succeed do differently. Then we give you a scored test you can run this week to tell a strong case from a weak one. For the pounds-and-pence side, pair this with our guide to how to calculate AI ROI.
The 95% figure, and what it actually means
The number comes from MIT NANDA's The GenAI Divide: State of AI in Business 2025, a July 2025 study drawing on more than 300 AI initiatives, 52 interviews and 153 senior leaders. Its finding is blunt: "95% of organizations are getting zero return", while "just 5% of integrated AI pilots are extracting millions in value" and the rest see no measurable impact on the bottom line.
It is easy to read that as "AI does not work". That is the wrong lesson, because the same report shows the 5% pulling real value out of the same tools everyone else has. The gap is not the technology. It is what firms do with it.
A small business does not need to be in the top 5% of anything. It needs to avoid the mistake that puts firms in the 95%, which is buying software and hoping the return arrives on its own.
Why most pilots return nothing
The losing pattern is familiar. A firm buys licences, a few keen people try the tool, everyone agrees it is impressive, and then nothing in the business actually changes. The same emails get written the same way, just with a chatbot open in another tab.
Spend goes out, no process is redesigned, and when someone asks what the return was, there is nothing to point at. Three things separate that from a return that shows up in the accounts.
The first is aim. AI helps most on a specific, repeated task, not on "the business" in general. A tool pointed at everything tends to help with nothing in particular.
The second is redesign. The value comes from changing the work, not bolting AI onto the side of it. If the AI drafts a reply but a person still writes it from scratch out of habit, the time is not saved. The saving is real only when the new way of working replaces the old one.
The third is honesty about the baseline. If you never recorded how long a task took before, you cannot show that it is faster now. Most firms in the 95% simply never measured, so their return is unprovable even where it exists. This is the failure we cover in why your AI pilot didn't scale.
What the 5% do differently
The firms getting value are not the ones with the biggest budgets or the cleverest models. They do three unglamorous things well.
They pick a narrow, high-volume task and go deep on it, rather than spreading a general tool thinly across the whole team. They change the process so the AI output is the new default, not an optional extra a few people remember to use. And they measure one clear thing before and after, so the return is a fact rather than a feeling.
There is a useful buy-or-build signal in the MIT work too. Tools bought from specialist vendors and configured to the firm succeed far more often than internal builds. For almost every small business the message is to buy and adapt, not to build your own. The energy is better spent on the workflow than on the technology.
One honest caution is worth adding about the size of the saving. Gains are real but they vary, and they tend to be largest on repeatable tasks and for less experienced staff.
In the NBER study Generative AI at Work, a 2023 analysis of 5,179 customer-support staff, an AI assistant raised issues resolved per hour by 14% on average. The gain reached around 34% for the newest workers, while the most experienced saw little change. Do not borrow someone else's percentage. Measure your own task, on your own work.
The worth-it test
Here is a test you can run on any candidate task before you spend a penny. We call it the worth-it test. It scores three things that, together, decide whether AI will pay: how often the task happens, how much each instance costs you, and how repeatable it is. Score each factor from 1 to 5, add them up, and read the band.
| Factor | Score 1 | Score 3 | Score 5 |
|---|---|---|---|
| Frequency (how often does this task happen?) | Rarely, a few times a month | Weekly, or a few times a week | Many times a day, across the team |
| Cost (what does each instance cost in time or money?) | Trivial, a couple of minutes | Moderate, 10 to 30 minutes of ordinary time | Expensive, an hour or more, or senior time |
| Repeatability (how similar is it each time?) | Bespoke, judgement-heavy, different every time | Broadly similar with variation | Same shape every time, house rules apply |
Add the three scores for a total out of 15, then read the band:
- Under 8: not yet. The task is too rare, too cheap or too bespoke for AI to pay back the effort of changing the workflow. Leave it and look elsewhere.
- 8 to 11: trial it. There is a plausible case. Run a short, measured trial on this one task with a baseline and a go/no-go number before you commit.
- 12 or more: strong case. This is exactly the shape of task the 5% target. Redesign the workflow around it and measure the result.
The test does two useful things at once. It stops you spending on tasks that will never pay, which is where most of the wasted 95% goes. And it points you at the frequent, costly, repeatable work where the return actually lives. To turn a passing task into a live project, use it alongside how to choose an AI use case.
A worked example
Take a fictional eight-person professional-services firm, and score two tasks. The numbers are illustrative; the point is the method.
Task one: first-draft replies to routine client enquiry emails. These come in perhaps 30 times a day across the team, so frequency scores 5. Each one takes a mid-level person around ten minutes, so cost scores 3. The enquiries fall into a handful of familiar types answered in a house style, so repeatability scores 4. Total: 12.
That is a strong case. The firm should redesign the workflow so an AI draft is the starting point for every routine enquiry. It should measure the time per reply before and after, and expect a real saving that compounds across the day.
Task two: a bespoke strategy recommendation for the firm's largest client. It happens once, so frequency scores 1. It is expensive and senior, so cost scores 5. Every instance is different and judgement-led, so repeatability scores 1. Total: 7.
That is a "not yet". AI can help around the edges, such as background research or a first tidy of notes, but this is not a task to build a project on. Pointing scarce effort here is exactly how firms end up in the 95%. Same firm, same tools, opposite verdicts.
When the honest answer is "not yet"
Sometimes the test comes back low across the board, and the honest answer is that AI is not worth it for your business yet. That is a perfectly good result. It is far cheaper to learn it from a scorecard than from a year of licence fees and a return no one can find.
"Not yet" is rarely "not ever". As your volumes grow, or as a task settles into a repeatable shape, a factor that scored 2 today may score 4 next year. The worth-it test is worth re-running as the business changes.
What it protects you from is the expensive middle ground: spending real money on a vague hope, with no task in focus and no way to tell whether it worked. That, and not the technology, is what the 95% figure is really measuring. Where AI pays, it pays because the work was redesigned around a task that deserved it, a distinction we go deeper on in AI productivity versus business value.
Frequently asked questions
Is AI actually worth it for a small business? It can be, on the right task. The evidence is that most firms see no measurable return because they buy a tool and change nothing about how the work is done. AI pays when you point it at a frequent, costly, repeatable task and redesign the workflow around it. The worth-it test above tells you whether a given task qualifies.
Why do so many AI projects fail? MIT found 95% of organisations get no measurable return, mostly because they never aimed the tool at a specific task, never changed the process, and never measured a baseline. The failure is one of adoption and design, not of the technology itself.
How can you tell if a task is a good fit for AI? Score it on frequency, cost and repeatability, each from 1 to 5. A total of 12 or more is a strong case, 8 to 11 is worth a measured trial, and under 8 means leave it for now. High-frequency, repeatable, moderately costly tasks are where the return lives.
Should a small business build its own AI? Almost never. Tools bought from specialist vendors and configured to the firm succeed far more often than internal builds. For most small businesses the effort belongs in the workflow, not in building technology.
Want a straight answer on whether AI is worth it for your firm, and which task to start with? Book a call and we will run the worth-it test with you on your real work, and tell you honestly where the return is, and where it is not.