Scroll through any business feed for a few minutes and the message is hard to miss. Another founder announces an "AI transformation". Another agency now calls itself AI-powered. Another software product has added a chatbot to its homepage. The pressure to do something is real, and it is easy to mistake noise for necessity.
So it is worth asking the uncomfortable question directly: does your business need AI, or are you worried about being left behind?
Both are honest answers. Only one of them should decide a large technology budget.
Adoption is everywhere. Results are not.
How many companies use AI depends on who you ask. In McKinsey's 2025 global survey, 88% of organisations said they use AI in at least one business function, up from 78% the year before. The US Census Bureau, which asks a far broader sample of businesses of every size, found AI use hovering between 17% and 20% from December 2025 to May 2026, with firms of 250 or more employees at 37% and the smallest firms below 20%. Large companies are much further along than everyone else, and the gap matters if you are deciding whether you have "fallen behind".
What both sources agree on is that using AI and benefiting from it are different things. In the same McKinsey survey, nearly two-thirds of organisations were still experimenting or piloting rather than using AI across the business, and only 39% said it had any effect on enterprise-wide earnings, most of them by less than five percent.
The bluntest figure comes from MIT's NANDA initiative. Its 2025 report, The GenAI Divide, drew on interviews with 150 leaders, a survey of 350 employees and a review of more than 300 public AI deployments, and found that only about 5% of generative AI pilots produced significant revenue growth. The sample is small, so treat the exact number with care. Its explanation is the useful part: the failures came less from weak models than from tools that never adapted to how the company actually worked. The report also found that more than half of generative AI budgets go to sales and marketing, while the biggest returns were in back-office automation.
Put simply: adoption is spreading quickly, above all among large companies. Measurable results are spreading much more slowly.
Signs your business genuinely needs it
For some businesses AI is not a trend to keep up with; it is a fix for a problem they already have. You are probably in that group if several of these are true.
- A repetitive, rule-shaped task is eating hours. Processing invoices, qualifying leads, scheduling, drafting first replies to customers, tagging documents. The inputs and outputs are clear, and people spend a large part of every week on it.
- You hold data nobody has time to use. Years of sales history, support tickets or customer records that could tell you about churn, demand or pricing, if anyone could get through them.
- The problem is speed, not judgement. Customers leave because a quote, a reply or a status update takes too long, not because the answer needs careful human thought.
- You can name the number that would move. Not "we would be more efficient", but cost per lead, hours per invoice, time to first response or tickets per agent, with a figure for today.
- A competitor's speed or price is already costing you work. You can see it in deals you lost, not in a trend report.
If two or more of these describe you, AI is less a fashion than overdue infrastructure.
Signs you are riding the wave
A lot of AI spending is driven by anxiety rather than a plan. These are the warning signs.
- The goal is "an AI strategy" and no particular process is named as the target.
- The audience is outside the company. The project is chosen because it will look good in an announcement, not because it changes the accounts. This is the same pull that sends budgets to visible sales and marketing pilots rather than to the back office, where the returns were found.
- Nobody has said how success will be measured, and the tool is about to be bought anyway.
- There is no plan for training or changing the workflow. The tool is dropped into the existing process and expected to improve it on its own. This is the gap the MIT researchers described: the model is rarely the problem; the organisation not adapting around it is.
- The trigger was a competitor's announcement, not a bottleneck you found yourself.
None of this means AI does not work. It means it was bought for the wrong reason, pointed at the wrong problem, or rolled out with nothing to make it stick.
Five questions before you spend anything
Before signing an AI contract or approving a build, answer these. They take an afternoon, and they are cheaper than a failed pilot.
- Which task, metric or bottleneck are we targeting? If the answer is vague, stop here. Everything else depends on it.
- What is the number today? You cannot show an improvement you never measured. Record the baseline before anything changes, and decide what result at 90 days would count as success.
- Is the data this needs available and usable? A system that reads invoices needs the invoices, in a form it can reach, with the right permissions around them. Finding out halfway through a project that the data lives in three spreadsheets and someone's inbox is how pilots stall.
- Who owns it, and will they be trained to use it? A tool without an owner inside the business gets quietly abandoned. The owner is also the person who decides what the system may do alone and what needs a human to approve.
- Is there a simpler fix? Sometimes a clearer process, a form instead of free-text emails, or an ordinary piece of automation solves the problem without a model at all. That is a good outcome, not a failed one.
If you can answer all five clearly, you are in a far better position than most companies running pilots today. If you are stuck on the first one, you are not ready yet, and that is a perfectly reasonable place to be.
If the answer is yes, start narrow
When a task passes all five questions, the way you build the first version matters as much as the decision to build it. The projects that work tend to share a shape.
One task, end to end. Pick a single process, such as reading incoming invoices and entering them into the accounting system, and make it work completely, including the awkward cases, before touching anything else. A narrow system that runs every day teaches you more than a broad one that never leaves the demo.
Connected to where the work happens. The output has to land in the system people already use: the ERP, the CRM, the help desk. A tool that produces answers in a separate window, which someone then copies across by hand, has moved the work rather than removed it.
People on the uncertain cases. Let the system handle what it is confident about and send the rest to a person, with the reason it was unsure. That keeps quality where the business needs it and shows you, week by week, where the system is improving.
A record of what it did. Log every input, decision and action. When someone asks why a customer received a particular reply, the answer should take a minute to find, not a meeting to reconstruct.
Measured against the baseline. Compare the cost and time of a finished task, retries and human review included, with the number you recorded before. That is what tells you whether to extend the system to the next process or stop.
The short version
Plenty of businesses now use AI in some form. Far fewer can say what it returned. The difference is rarely the model; it is whether the project started from a real, measurable problem and was built into the way the work is actually done.
If you can name the task, the number and the person who owns it, AI may well be the right tool, and starting now is sensible. If you cannot yet, the most useful step is not buying anything. It is finding the bottleneck.