Where AI projects go wrong in small businesses
The five failure patterns behind most abandoned AI pilots, and the cheap checks that catch each one before you have spent a budget finding out.
Small businesses do not usually fail at AI because the technology was not good enough. They fail earlier, in the choosing, and by the time anyone notices, the budget is gone and the conclusion drawn is “AI does not work for us.”
These are the five patterns we see most, with the check that catches each one before it costs anything.
1. Solving a problem nobody has
The project starts because AI is on the agenda, not because a specific task is painful. So somebody picks something visible and plausible, usually a chatbot on the website, and builds it.
Six months later it has answered forty questions, thirty of which were “what are your hours.”
The check: name the person whose week gets shorter, and by how many hours. If you cannot, the project has no owner and no measure, and it will be judged on novelty, which wears off.
2. Automating a process nobody agrees on
AI is very good at doing a task consistently. It is useless when three people do the task three different ways and each believes theirs is correct.
Attempting to automate in that situation surfaces the disagreement, which is genuinely valuable, but it surfaces it halfway through a build rather than at the start.
The check: write down the current process in ten steps and have two people who do it confirm it is right. If they cannot agree, that is the project, and it does not need AI.
3. Treating a confident answer as a correct one
Language models produce fluent text regardless of whether they know the answer. For summarising a document that fluency is fine. For quoting a price, citing a policy or answering a customer’s legal question, fluency without grounding is a liability with good grammar.
This is the failure that damages trust fastest, because the first wrong answer that reaches a customer ends the pilot.
The check: for every output, ask what happens if it is confidently wrong and nobody notices for a week. If the answer is expensive, the system needs citations back to source documents and a human approving anything outward facing. If the answer is “nothing much,” proceed cheaply.
4. Buying the platform before understanding the task
The market is full of platforms that do everything, priced per seat, configured over months. They make sense for organisations with a dedicated team to run them. For a business of twelve people they are a large recurring cost attached to a capability nobody has yet proven they need.
The reverse also happens: a business builds custom infrastructure for something an existing tool already does well.
The check: do the task manually with an off-the-shelf assistant for two weeks first. Paste the documents in by hand. It is tedious and it answers the only question that matters, which is whether the output is good enough to act on. Build or buy is a different decision once you know that.
5. Ignoring what it costs to run
Pilots are cheap because ten people use them occasionally. Costs arrive at rollout, and they are usually not the model calls. They are the embedding and re-embedding of documents, the storage, the retries, and the person who now owns keeping it working.
We wrote out the actual arithmetic in what a document AI system costs to run, because the gap between a pilot bill and a production bill surprises people.
The check: multiply your pilot usage by realistic adoption and add a line for the human who maintains it. If the total is uncomfortable, the project is not viable at scale and it is better to know now.
What the ones that work have in common
They are boring, narrow and measurable:
- One task, not a capability. “Draft the first version of the site report” rather than “AI for operations.”
- A person who wanted it, who was consulted, and who will use it daily.
- A quality bar defined in advance. Good enough to edit, not good enough to send. Those are different systems with different costs.
- A clear fallback. When the model is unavailable or unsure, the work still gets done the old way rather than stopping.
- A number, checked at ninety days. Hours saved, backlog cleared, response time. If nobody committed to a number, nobody can call it a success or a failure, so it will drift.
The honest first step
Pick the most repetitive text-shaped task in the business. Not the most strategic one, the most repetitive. Do it with an assistant manually for two weeks and keep notes on where it was wrong.
That fortnight costs almost nothing and tells you more than a proposal will. It also quite often ends with the finding that the task should be eliminated rather than automated, which is a better outcome than either building or buying.
If you want a straight read on whether a specific task is worth automating, describe the task and who does it now. Sometimes the answer is that software is not the fix, and that is worth hearing before you spend.
Related: when your business actually needs AI, what to automate first, and our AI tools work.
Further reading
Topics in this article
Get new articles by email
One email when something new goes up, roughly twice a month. Plain writing on what software costs and what is worth building. No sequences, no sales calls, and one click to leave.
Keep reading on the same subject
-
What a document AI system actually costs to run
Real monthly figures for retrieval systems over your own documents, what drives the cost, and the two mistakes that inflate it most.
See the real numbers -
Getting AI to answer from your own documents
How document AI actually works, why answer quality is a filing problem rather than a model problem, and what to fix before you build anything.
See how retrieval works -
When does your business actually need AI?
A practical test for whether an AI feature is worth building, the four cases where it genuinely pays, and the ones where it quietly wastes money.
Find out if you do