- Most AI projects never reach production or deliver nothing measurable. The technology is rarely to blame: the problem almost always lies in the method.
- Seven mistakes keep coming back: starting with the tool, overlooking the data, aiming too big, not measuring, removing people, forgetting the teams, leaving AI without rules.
- Each one can be avoided with a simple rule, to check before you spend a single dong.
Why so many AI projects stall
The figures look alike from one study to the next. In 2024, the RAND Corporation interviewed 65 experienced data scientists and engineers: according to the estimates it cites, more than 80% of AI projects fail, twice the rate of conventional IT projects (RAND).
The trend is speeding up. According to S&P Global Market Intelligence, which surveyed more than 1,000 companies in North America and Europe, 42% of companies abandoned most of their AI initiatives in 2025, up from 17% a year earlier. On average, they scrapped 46% of their prototypes before production (CIO Dive).
Vietnam shows the same pattern, slightly later. The 2026 study by Strand Partners for AWS, based on 1,000 business leaders, shows that 61% of companies using AI are still experimenting. Only 23% have a company-wide AI strategy, 13% consistently measure what it brings in, and 46% have no reliable way of knowing (VietNamNet).
BCG looked at where the difficulties lie. Among 1,000 executives in 59 countries, around 70% of the challenges come from people and processes, 20% from technology and data, and only 10% from the algorithms themselves (BCG).
An AI project rarely fails because the AI doesn’t work. It fails because it was put in the wrong place, without clean data, without measurement and without the teams.
An SME doesn’t have a large group’s budget to absorb a failure. But it has an advantage: short decision chains, a manager who decides quickly, and processes that can be observed in a single day. Here are the seven mistakes to avoid to make the most of it.
Mistake 1: starting with the tool instead of the problem
This is the number one cause according to RAND: the problem to solve is misunderstood, or poorly explained to the people building the solution. The researchers add that chasing the latest AI advances for their own sake is one of the most frequent paths to failure.
In an SME, it often starts with a sentence: “we need a chatbot”, “we want AI agents”, “my competitor put ChatGPT on their website”. The tool is chosen before anyone has looked at the work.
Typical example. A travel agency installs an assistant on its website. It answers well, but customers mostly write on Zalo, and the real loss is elsewhere: weekend quote requests go unanswered until Monday, and some of them go to the competition. The chatbot changes nothing.
How to avoid it. State the need without saying the word “AI”: “we lose quote requests at the weekend”, “entering invoices takes three days a month”. Put a number on it in hours, errors or money. And beware of fads: getting cited by ChatGPT or Gemini, for example, is first a matter of visibility in AI answers, not an automation project. Sometimes the right answer isn’t even AI: a well-designed form, a single customer database or a clearer website is enough. Our article on where to start automating an SME offers four questions to spot the right tasks.
Mistake 2: assuming the data is ready
An AI is only as good as the data it is given. According to a 2025 Gartner survey, 63% of organizations don’t have, or aren’t sure they have, the data management practices AI needs. Gartner predicts that through 2026, 60% of AI projects without AI-ready data will be abandoned (Gartner). An engineer interviewed by RAND sums it up: 80% of AI work is the unglamorous job of preparing data.
In a Vietnamese SME, the data almost always exists. It is simply scattered: orders in Zalo, prices in Excel, payments in the banking app, sales in the POS, and part of the history in the heads of two or three people.
A real case. At Baieta, a restaurant in Ho Chi Minh City (see the restaurant example), the manager wanted to know his margin per dish. Before any calculation, the sources had to be gathered and linked: supplier invoices, bank statements and POS sales.
How to avoid it. Before choosing a model, take stock: where each piece of information lives, in what format, who updates it, and how reliable it is. Start with the sources that are already digital. Vietnam has a real asset here: every company issues e-invoices in XML format, which machines can read (see our article on automating e-invoicing).
Mistake 3: aiming too big, or getting stuck at the prototype
Two opposite traps lead to the same result. The first is the big project that tries to transform everything at once: twelve months, five departments, a budget committed in one go. The second is the prototype that impresses in a demo, then sits in a drawer because nobody planned how to connect it to the real tools.
MIT’s study of generative AI in business, published in 2025, made a lot of noise: according to its authors, 95% of the organizations studied saw no measurable return on their results six months after deployment (Virtualization Review). The figure is debated, but two findings from the report are useful for an SME. First, around half of budgets go to sales and marketing, while automating back-office tasks pays off more. Second, projects run with a specialist partner succeed about twice as often as those built alone in-house (The Stack).
How to avoid it. Break it down. A first batch of 1 to 3 measures, over 4 to 8 weeks, connected from the start to your real tools and tested alongside the old way of working. You move on to the next batch only if the gain is proven: that is the principle of our 5-step method.
Mistake 4: not measuring the “before” situation
Without a starting point, there is no way to prove a gain. The project then remains an impression (“it seems to save us time”), and it is the first to go when budgets tighten. That is exactly what the AWS study in Vietnam shows: only 13% of companies consistently measure the return on their AI projects.
Measuring also lets you stop in time, which is a form of success. McDonald’s tested AI voice ordering with IBM for about two and a half years, in around a hundred US restaurants. The target was set in advance: more than 95% accurate orders before rolling it out. A study published in 2022 put accuracy closer to 80%, and the chain ended the test in 2024 (Nation’s Restaurant News). A numerical threshold turned a potential failure into a clear decision.
How to avoid it. Before you start, write down three to five simple indicators: time spent per week, number of errors, response time, unpaid invoices, lost requests. Set the threshold that justifies continuing. Then compare at the end of each batch, in a gains report shared with management and the teams.
Mistake 5: removing people too quickly
It is tempting to see AI as a way to cut jobs. The most cited case is Klarna. In 2024, the Swedish fintech announced that its assistant was doing the work of 700 customer service agents. In May 2025, its CEO acknowledged that the cost-cutting had gone too far and started hiring again, so that every customer can always talk to a person (Bloomberg).
The second risk is legal. In 2024, a Canadian tribunal ordered Air Canada to compensate a passenger: the airline’s chatbot had wrongly told him he could claim a reduced fare after his trip. The airline argued that the chatbot was responsible for its own answers. The tribunal replied that the company is responsible for all the information on its website, whether it comes from a page or a chatbot (CBS News).
In Vietnam, Law No. 134/2025/QH15 on artificial intelligence has applied since 1 March 2026. It requires human oversight and control (The Star), and provides for labelling AI-generated content (Vietcetera).
How to avoid it. Give AI the repetitive tasks and keep decisions human. Every automation needs a validation step, a clear handover to a person for complex cases, and answers drawn from your own documents, reviewed and kept up to date. An assistant that doesn’t know the answer should say so, not make one up.
Mistake 6: forgetting the teams
An automation nobody uses brings nothing. Yet teams are often told at the end, when the tool has already been chosen. They know better than anyone the exceptions, the workarounds and where time is really lost.
Employees haven’t waited for their employer. According to the 2024 Microsoft and LinkedIn study, 78% of AI users bring their own tools to work, and only 39% have received training from their company (Microsoft). In Vietnam, 54% of businesses cite the skills shortage as their main obstacle to AI (Viet Nam News).
How to avoid it. Involve the people who do the work from the observation stage. Appoint a lead in each department. Plan a short training per role, on your own documents rather than generic examples, then share the first gains with everyone. We offer role-based training workshops designed for this purpose.
Mistake 7: letting everyone use AI without rules
When the company provides neither tools nor rules, everyone makes do with a personal account. The risk is not theoretical. In March 2023, a few weeks after allowing ChatGPT in its semiconductor division, Samsung found at least three cases where employees had pasted confidential source code or the content of a meeting into it. The group restricted the use of generative AI from May (Fortune).
The legal framework has tightened. Law No. 91/2025/QH15 on personal data protection has applied since 1 January 2026. It provides for fines of up to 5% of the previous year’s revenue for non-compliant transfers of data abroad, and up to 3 billion VND for other violations (PwC). Pasting a customer file into a consumer AI hosted abroad is therefore far from harmless.
How to avoid it. Providing good tools works better than banning. Choose the business versions of AI models, which don’t use your data for training, and decide where sensitive data lives. Write a one-page rule: what can be entrusted to AI, what never is, and who to ask when in doubt.
Before launching a project: 7 questions to ask
| Mistake to avoid | The question to ask | Good sign |
|---|---|---|
| Starting with the tool | Can we describe the problem without saying “AI”? | The need fits in one sentence, with a number. |
| Data not ready | Do we know where each piece of information lives, and in what state? | The sources are listed and already digital. |
| Aiming too big | Does the first batch fit in 4 to 8 weeks? | 1 to 3 measures, connected to the real tools. |
| Not measuring | Which “before” figures, and what threshold to continue? | 3 to 5 indicators written down before starting. |
| Removing people | Who validates, and who takes over when in doubt? | A checkpoint planned at every step. |
| Forgetting the teams | Did the users take part in the design? | A lead per department, training per role. |
| AI without rules | Which data can go into which tool? | A one-page rule and business accounts. |
What a project run in the right order looks like
We apply these rules to ourselves first, and we measure the result every month. Preparing a project now takes one day instead of a week. Accounting and HR save a week a month. Each developer saves 5 days a month thanks to AI, which helps write, test and review code.
None of these gains came from a miracle tool. They come from a series of small batches, each measured before and after, on tasks the teams themselves had flagged as tedious. To see what this could look like in your business, see typical gains by industry and some of our projects.
If you don’t know where to start, the Express AI Audit is designed precisely to avoid these seven mistakes: two days, including one at your premises, to observe the real work, measure the “before” situation and put numbers on 1 to 3 measures. The price is deducted from implementation if you decide to go further.
Frequently asked questions
Why do most AI projects fail?
Rarely because of the technology. According to BCG, around 70% of the challenges come from people and processes, 20% from technology and data, and only 10% from the algorithms. The most common causes are a poorly defined problem, scattered data, no measurement and teams left out.
What percentage of AI projects really fail?
It depends on what you call a failure. According to the estimates cited by RAND, more than 80% of AI projects don’t meet their objectives. According to S&P Global, 42% of companies abandoned most of their AI projects in 2025, up from 17% in 2024.
Does an SME have a better chance of success than a large group?
It has advantages: fast decisions, few intermediaries and processes that can be observed in a single day. Provided it starts small, with 1 to 3 costed measures, connected to its current tools and measured before and after.
How do you know whether an AI project is profitable?
By measuring the “before” situation: time spent, errors, delays, unpaid invoices, lost requests. You set a threshold that justifies continuing, then compare at the end of each 4- to 8-week batch. Without that starting point, no gain can be proven.
Should you ban ChatGPT for your employees?
Bans rarely work: employees then use their personal accounts in secret. It is better to provide business versions, which don’t use your data for training, and a one-page rule on what can be entrusted to them.
Which laws govern AI use in a business in Vietnam?
Two main texts. Law No. 91/2025/QH15 on personal data protection, in force since 1 January 2026, provides for fines of up to 5% of revenue for certain transfers of data abroad. Law No. 134/2025/QH15 on artificial intelligence, in force since 1 March 2026, requires human oversight of AI systems.
Founder of Silkwires, based in Ho Chi Minh City since 2010. A digital specialist since 2000, he leads the AI diagnostics and visibility projects.
