AI strategy
AI adoption challenges: why AI projects fail and how to keep yours on track
Most AI adoption challenges come from how a project is chosen, owned and measured. Here are seven causes that come up again and again in small and mid-sized companies, the signs that give them away in the first weeks, and a simple exercise to head them off before you launch.
What failure looks like in an AI project
The AI licences a Cardiff solicitors' firm bought in January are still active five months later, when one of the partners goes through them. Two solicitors use them every day. The rest tried them a couple of times and went back to their old way of working, and the project faded out without anyone deciding to drop it. It is one of the commonest AI adoption challenges: the tools arrive and the way people work stays the same.
There are a handful of reasons why AI projects fail, and most of them have little to do with the technology. Failure here means the project produces no measurable, lasting change in the process it was meant to improve. The money is gone and leadership concludes that AI doesn't suit the business.
McKinsey's The State of AI 2025 reports that 62% of organisations are experimenting with AI agents. What takes real effort is turning a trial into a process that runs every day, that people trust and that has figures to show its value.
An accounting firm, an online shop and a law practice tend to run into the same causes, and very few of them are technical. They come down to how the project is chosen, who runs it and how it is measured, which means they can be prevented.
Why AI projects fail before they start
Three of the seven causes appear at the decision stage, before anyone has written a prompt or connected a tool.
No clear problem
The project starts with a vague aim such as "use AI in customer service". Without a specific task and a figure to improve, nobody can tell whether it is working. Describe the problem in one sentence with a process, a volume and an expected result.
Starting with the tool
Someone sees a demo, buys licences and then looks for something to use them on. The tool ends up forced into place. Start from the process and where it gets stuck, and choose the tool last, once you know what it has to do.
Nobody owns it
The project belongs to everyone, so in practice to no one. When a decision is needed, nobody makes the call. Name one owner on the business side, with time set aside in their week and the authority to change the process.
AI adoption challenges that show up along the way
The other four appear once the project is running. They are what turns a good idea into a trial that goes nowhere.
Messy or unreachable data
AI works with what it is given. If documents sit in personal folders or CRM records are half filled in, results will be patchy. Check the quality of the process data and who can reach it before building anything, and fill the gaps that matter most.
No training for the team
Licences go out with a welcome email and little else. Each person experiments alone, gets mediocre answers and gives up. Train the team on examples from their own tasks and show them how to check what AI produces.
The pilot never scales
A pilot is a limited trial with a few people and an end date. If nobody agreed in advance what result would justify rolling it out, it stays with two enthusiasts for good. Set its length, its success criterion and what happens next before it begins.
Nothing is measured
Without a baseline there is no way to prove an improvement, and the project is exposed at the first cost review. Measure time per task, volume and errors before you start, and repeat the measurement on agreed dates.
Warning signs in the third week of a pilot
A project heading for failure usually signals it early. By the third week, these signs show which cause is at work and where to step in.
Use drops after the first week
Training is usually missing, or the tool does not suit the task chosen. Ask two users what they did before and what they do now.
People redo the AI's output by hand
This points to incomplete data or unclear instructions. Count how many outputs get corrected, and why.
Nobody can say which figure is being watched
The baseline or the success criterion is missing. Without them, the final review becomes an argument about impressions.
The review meeting keeps getting postponed
The project has no owner with real time for it. A glance at the calendar shows it, yet it rarely gets taken seriously.
There is talk of rolling out with no data
The case rests on two or three keen users, with no figures behind it. A rollout on that basis tends to collapse at the first budget review.
A one-hour pre-mortem before you launch
A pre-mortem takes failure as given and asks the team to explain it. That brings out risks nobody mentions in an upbeat kick-off meeting.
- 01
Get the right people in the room
The owner, two or three of the people who will do the task every day, and whoever knows the data and tools. In a small business, five people are enough.
- 02
Jump six months ahead
Set the scene. The project has faded out and nobody uses the tool. Each person spends five minutes writing down, in silence, the reasons they imagine.
- 03
Share and group
Read the reasons aloud and sort them under the seven causes in this guide. The ones that come up repeatedly are your main risks.
- 04
Turn each risk into a check
For each risk, one action with a name and a date. Write the problem in one sentence, check the data, name the owner, plan the training or set the baseline.
- 05
Decide how the pilot ends
Before the meeting closes, write down the pilot's length, its success criterion and what happens if it is met and if it isn't.
Example
A Cardiff solicitors' firm revives its AI project
Back to the Cardiff firm, a hypothetical practice of twelve people. Rather than renew blindly, the partner goes through the seven causes with two solicitors. They find three: there was no specific problem, nobody owned the project and training amounted to one email.
They restate the project in one sentence: produce the first draft of debt recovery letters faster, in a shared format. An associate takes ownership, with two hours a week set aside. Client data is only used on accounts the firm has approved and is anonymised for training.
Before starting, they time each first draft for two weeks and note how many corrections it needs. The pre-mortem adds a risk nobody had spotted: the letter templates were scattered across personal folders. They gather them in one shared folder before the pilot.
The pilot runs for six weeks with four solicitors and a training session built on the firm's own cases. At the end, the decision rests on the time and correction figures: roll out across the firm, adjust the instructions or close the project.
What to do with an AI project that has stalled
Pause and go through the seven causes and the warning signs with the project owner. Usually what is missing is a measurable goal, training or a clear decision about the pilot. Bring back the baseline, narrow the scope to one process and set a new review date. If there is no way to show value, closing the project is a sound decision too.
Obinova's AI audit and consulting tackles these causes from the start. We look with your team at how work is done today, score each use case on impact and feasibility, data included, and leave you with a roadmap of owners and a way to measure each project. The mission continues with follow-ups several times a year.
If the team needs training, we deliver it on ChatGPT, Claude, Gemini or Microsoft Copilot using cases from their own work. If a process can be automated, we build the n8n workflow with your people and document it.
Frequently asked questions
What is the most common reason AI projects fail?
In SMEs, the most frequent cause is starting without a specific problem. Licences are bought or a tool is tested before anyone has decided which task should improve or how it will be measured. The other causes follow from there, because without a goal it is hard to name an owner, prepare the data or decide whether the pilot deserves a wider rollout.
What percentage of AI projects fail?
There is no reliable figure that applies to a small business. The studies that circulate define failure in very different ways and mostly look at large companies. What is known is that experimenting is widespread, with McKinsey reporting that 62% of organisations are trying AI agents, and that moving from trial to daily use is the hard step. The useful question is whether your project meets its own success criterion.
How long should an AI pilot last?
For a single process, four to eight weeks is usually enough to gather reliable figures. Less time leaves the team little room to get used to the new way of working. Longer often means the success criterion was never clearly set. What matters is fixing the end date, and the decision you will take, before the pilot starts.
Do you need perfect data to start with AI?
No. The data for the chosen process needs to be accessible and reliable enough for that task. An accounting firm can automate invoice logging without tidying its entire archive first. Review only the information the use case will rely on, fix what is essential and leave the rest for later.
Are AI adoption challenges different in a small business?
The causes are largely the same, with less slack to absorb them. A small business rarely has someone free to run the project, so ownership and time are the usual sticking points, and one person leaving can stall it. Decisions are quicker, though, and processes are easier to map. Keep the first project to one process, with a named owner and time set aside in their week.
How does an AI audit help prevent failure?
An audit puts the early decisions in order. It maps processes, scores each use case on impact and feasibility, reviews the available data and records who owns each project and how it will be measured. At Obinova we follow it up with reviews through the year. You receive a quote after a free 30-minute first meeting.