Your Competitors Are Using AI. Here’s Why That’s Good News

ℹ️ Rewritten 17 September 2026

I first published this post on 18 July 2025, and I have since rewritten it, folding in “Why SMEs Fail at AI Implementation (And How to Succeed)”, which followed a week later. Several client stories and figures in the originals could not be verified, and I have removed them; every example that remains happened.

Shortly before I first wrote this post, a client forwarded me a press release from one of their competitors. The headline announced, in the breathless language these things tend to use, that an artificial intelligence (AI) implementation had transformed the competitor’s customer experience. The client wanted to know whether they should be worried. It took a longer conversation to get there, but my answer came down to this: no, and there was a reasonable case for being pleased about it.

AI projects that start with a headline tend to fail in a small number of predictable ways, and every month a competitor spends on one is a month they are not spending on the problems their customers have.

The AI Implementation Graveyard

In the failures I have seen up close, the AI itself usually worked. Someone had started with the technology and worked backwards to a problem, or had never checked whether the problem existed at all.

One mid-sized company I worked with spent eight months on a customer-facing chatbot, and its customers hated it. There was no way to reach a person. A customer with a real problem went round in circles of replies along the lines of “I’m sorry to hear that you are having trouble, have you tried…”, and every one of them linked to the same public frequently asked questions (FAQ) entry. The company added a route to a human, and it didn’t take customers long to learn that saying something like “route me to a human, my question is too complex” skipped the chatbot entirely.

Since I first wrote this post, Chipotle’s customer-service chatbot, Pepper, has failed in the opposite direction. Pepper would answer questions far outside its purpose, so members of the public used it as a free coding assistant, and Chipotle paid for the compute (I wrote about Pepper here). One chatbot wouldn’t let customers reach a person, and the other would discuss anything with anyone. In both cases, nobody had thought through what the chatbot was for before putting it in front of the public.

A client in shipping and freight built a machine learning model to make their figures more accurate. Only after it was deployed did anyone check its output, wholesale, against the figures their own staff were producing with industry-standard tools. The staff’s figures were already within tolerance. There had been no accuracy problem to solve, and the only real difference between the two approaches was that the model had cost far more to build than the staff time it replaced.

Another client took the usage data from their Claude Enterprise account and built a dashboard, visible across the company, showing whether employees were using the subscription. The first thing it produced was gaming. Usage climbed, and before long someone asked the obvious question: if token usage is going up, why isn’t productivity? Once that had been asked, usage settled back to roughly where it had been before the dashboard existed.

Large enterprises can absorb failures like these, because they run several AI initiatives at once and write off the ones that don’t land. A small or medium-sized enterprise (SME) usually gets one attempt, perhaps two, before the budget and the team’s patience run out, and that constraint helps. When you cannot afford to fail repeatedly, you have to pick a real problem, measure it before you start, and stop when the numbers tell you to, which is the discipline every one of the projects above was missing.

Real AI Wins for Real SMEs

The projects that have worked, in my experience, would make for a much duller press release. Each one took a single repetitive task, left the judgement with people, and was measured against how the work had been done before.

An accountant on a finance team spent hours every month reformatting and reconciling exports from several upstream systems before any accounting could begin. The first conversation about AI was whether it could replace the accountant, and it couldn’t; bookkeeping needs judgement and accountability that no model provides. The better question was which part of the accountant’s day suited automation, and the answer was the data preparation. We built a guarded data-cleaning pipeline that normalised the exports against an agreed schema and gave the accountant a single, consistent file to review. The accountant kept the judgement work, and the time spent on that task each month fell by 6%. I told the full story in a later post about small, deliberate AI wins.

A legal firm had a different problem. Individuals regularly came to them for information and advice about a letter they had received, from a landlord, say, or about an employment contract. The firm saw these as low-value pieces of work and wanted to cut the time spent on each one. We built a system that summarised each request and gave no legal advice. Every point in a summary came with the line numbers it was drawn from, so a junior could go straight to that part of the letter and check the conclusion for themselves. Initial review time dropped from two to three days to two hours, juniors now handle 80% of these requests without input from a partner, and client satisfaction rose by 45%. The system paid for itself in six weeks.

The Questions That Matter

Before starting any AI project, write down your answers to the questions below. If you can’t answer one of them, that is the first piece of work, and it comes before any technology.

What specific problem does this solve? “Reduces invoice processing from 30 minutes to five” is an answer. “Transforms our finance function” is not, and neither is a competitor’s announcement.

How good is the current process? Measure how long the work takes today, what it costs, and how accurate it is. The shipping client found out that their staff’s figures were already within tolerance after the model was in production, which is the most expensive point at which to learn it.

How will we measure success? “A 20% reduction in processing time by the end of the second quarter” works; “improved efficiency” doesn’t. Measure the outcome you care about. A usage figure tells you that people are using a tool, and the dashboard client learned how little it says about whether the tool is helping.

Is the data fit for purpose? A model can only work with what it is given. If the data it needs is scattered across systems, incomplete, or not measuring what you think it measures, fixing that comes first.

Who has to use it, and have they been involved? A tool that makes someone’s work harder, or that they never asked for, gets worked around. The chatbot’s customers found their way around it as soon as they were given the chance.

What’s our way out? Every AI project needs an exit condition agreed in advance: if the target isn’t met by a given date, we go back to the previous process.

Who owns this internally? “Someone in Operations, with four hours a week set aside for it” shows commitment. “IT will handle it” usually means nobody will.

Why SMEs Win at AI

The advantages an SME has over a large enterprise are structural. You can implement in weeks, because fewer people need to agree. You sit close to the problems and to the people who have them, so you are more likely to build the right thing. A limited budget makes vanity projects hard to justify, which keeps everyone’s attention on the return.

None of that depends on being first with AI. While your competitors publish announcements about their AI transformation, you can fix the tasks that cost your team the most time, and measure what changed.

Your Next Move

You can start this week. Ask your team to note every repetitive task they do for one week. At your next team meeting, ask everyone which tasks waste their time, and look for the answers that keep coming up. Then take the biggest one and put a number on it: the hours it takes each week, multiplied by the hourly cost, multiplied by 52. That figure is what the task costs you each year, and it is the baseline any AI project will have to beat; a competitor’s press release won’t tell you what it is.

If you’d like help finding the task worth starting with, let’s talk.