Why 77% of AI Projects in SMBs Fail

91% of mid-sized companies say AI matters. 94% have not completed a single AI project. The 6% who tried? 77% failed.

These numbers come from recent studies by the German Mittelstand-Digital Zentrum, the IHK, and Bitkom. They tell a story I know first-hand.

I saw how AI is done right at Instacart: start small, learn fast, then scale. At DHL, Delivery Hero, and New Relic, I watched companies with massive data and enormous budgets set up AI projects. Some succeeded, many did not. In the German SMB market I now see the opposite: big project, long timeline, no measurable outcome.

The good news: failure has patterns. Once you know the patterns, you can avoid them.

Five reasons AI projects fail in SMBs

Over the past two years I have spoken with dozens of mid-sized companies about their AI plans. Manufacturing firms in the Eifel, service providers in the Rhineland, skilled trades in the Kreis Euskirchen. The problems repeat.

1. Poor data quality

"Garbage in, garbage out" is not a textbook slogan. It is daily reality.

An AI model is only as good as the data it trains on. And the reality inside many SMBs looks like this: customer data in three separate spreadsheets, production data in a system from 2014, order history half digital, half in paper binders.

If you launch an AI project on that data, you get results that are useless at best. At worst the results look plausible but are wrong. Then you make decisions based on faulty predictions. That is more dangerous than having no AI at all.

At Instacart we had a team that focused only on data quality. Not on models, not on algorithms. Only on clean, consistent, current data. Without that foundation the best model in the world delivers nothing.

Most SMBs do not have that team. And that is fine. But then you need to start with the data you already have and honestly evaluate whether it is good enough.

2. No specific problem to solve

"We need AI." I hear this constantly. When I ask "For what, exactly?", the room goes quiet.

Most failed AI projects begin with the technology instead of the problem. Someone attended a trade fair, heard a talk about ChatGPT, and now the company should "do something with AI." That is like buying a hammer and then looking for a house where you can drive nails.

The right approach is reversed: you have a specific problem. You lose customers because proposals go out too slowly. Your team spends 10 hours a week on manual data entry. Quality control repeatedly misses the same defect type. Then you check whether AI can solve that problem.

3. Scope too large from the start

"We are doing a complete AI transformation." That sentence is the beginning of the end.

Companies that start with a giant project almost always fail. Why? Because a holistic project is too complex, too expensive, and takes too long. After six months you have spent a lot of money but have nothing anyone uses day to day. Motivation drops, the budget gets questioned, and eventually the project dies quietly in a Teams channel nobody opens anymore.

At New Relic I watched even a billion-dollar tech company roll out AI step by step. Not as a "transformation" but as a series of small experiments. Each experiment had a clear timeframe and a measurable goal. If it worked: scale. If not: stop and learn from the mistakes.

That discipline is almost always missing in SMBs.

4. No internal champion

AI projects need someone inside the company who stays on top of it. Not the CEO who announced the project at a strategy conference. Someone who knows the daily work, talks to the team, spots resistance and removes it.

Without an internal champion, here is what happens: the external vendor delivers a solution. The team does not understand why they should change how they work. The tool gets used for two weeks and then ignored. The project counts as "implemented" but creates no value.

Change management sounds like corporate jargon. It is. But the underlying problem is real, even in a 15-person shop. People do not change their behavior just because someone installed a new tool. They change when they experience the benefit themselves.

5. Wrong vendor

This last point is awkward because I am a vendor myself. But it has to be said.

Many AI projects fail because the vendor sold the wrong solution. A consulting firm that runs a months-long assessment before anything happens. A startup that sells a platform far too complex for an SMB. An agency that puts "AI" on the website but really just glues a few API calls together.

What to look for: Does the vendor have experience with companies your size? Can they show references where measurable ROI resulted? Do they talk in months or weeks? Do they want to analyze your entire company first, or do they start with a specific process?

The pattern that works: small, fast, measurable

Now for the good news. The AI projects that succeed follow a clear pattern. I saw it at Instacart, at Delivery Hero, and with the SMBs that got it right.

The pattern in three steps:

Step 1: Pick a single process. Not two, not three. One. The process that annoys the most, eats the most time, or causes the most errors. That is your pilot.

Step 2: Deliver a measurable result in 30 days. Not in six months. In 30 days. That forces you to keep the scope small. "Can we show within 30 days that this process runs faster, better, or cheaper?" If the answer is no, the scope is too big.

Step 3: Measure, learn, then scale. Did the pilot work? What exactly did it deliver? How many hours saved, how many errors reduced, how much additional revenue? Only once you have those numbers do you decide on the next step.

Why does this work? Because small projects deliver fast feedback. You see within weeks whether something works. You can course-correct before it gets expensive. And you build internal trust, which may be the most important point. When your team sees that a small AI project actually saves hours, the willingness for the next step is there.

AI readiness: 5 questions for your company

Before you start an AI project, answer five questions honestly. None of them require technical knowledge. All of them require honesty.

1. Do you have a specific, measurable problem? Not "we want to be more efficient" but "process X takes 8 hours per week and should take 2." If you cannot describe the problem in one sentence, it is not ready for AI.

2. Do you have data on this problem? AI needs data to learn. If your orders live in three spreadsheets, two email inboxes, and a notebook, you first need to consolidate. That sounds boring. It is. But without this step, AI delivers nothing.

3. Is the process repeatable? AI excels at tasks that repeat. Writing proposals, checking invoices, capturing customer data, running quality checks. One-off tasks or tasks that differ completely every time are harder for AI.

4. Can you measure success within 30 days? If your pilot needs 6 months before you know whether it works, the scope is too big. Shrink it until the answer is yes.

5. Is there someone on the team who will drive it? Not the CEO (they have other things to do), but someone who knows the process and is motivated. A team lead, a floor manager, an office manager. Someone close to the problem.

If you can answer three or more questions with yes, you are ready for a pilot. If not, work on the gaps first. That is not failure, that is preparation.

Success vs. failure: what makes the difference

Let me tell you two stories. Both from the region, anonymized.

Company A: EUR 80,000 burned

A mechanical engineering firm with 45 employees. Management decides: "We are doing AI." A consulting house is hired. Three months of assessment, interviews, workshops. Result: a 120-page roadmap. Cost so far: EUR 35,000.

Then a software company is hired to build "AI-driven production optimization." After four more months there is a prototype. It works in the lab but not with the real machine data, because the data format was never checked. Another EUR 45,000.

Total cost: EUR 80,000. Result: a prototype nobody uses. The production team never understood and never accepted the project. Today the word AI is on that company's internal blacklist.

Company B: EUR 3,800, ROI in three weeks

A skilled-trades business with 12 employees. The problem: the office manager spends every Friday 4 hours assembling proposals from notes, photos, and voice messages.

The solution: a simple workflow with a speech-to-text tool and a template. The technician dictates notes into a phone, the tool transcribes them, and the office manager receives a pre-filled proposal draft. She just reviews and sends.

Cost: EUR 3,800 for setup, EUR 60 per month for licenses. Time saved: 3.5 hours per week. ROI: after three weeks. And the best part: the office manager drove the project herself because she was fed up with the Friday grind.

The difference between these two cases? It is not the technology. It is the approach. Company A started with the solution. Company B started with the problem.

Which processes make good starting points

Not every process is equally suited as an AI pilot. Here are categories I see again and again as strong entry points in SMBs.

Document creation covers proposals, reports, meeting notes, and summaries. Anywhere someone turns raw data into a document. This is one of the most common starting points because the effort is clearly measurable beforehand.

Customer communication includes pre-drafting email replies, auto-answering FAQ requests, and suggesting callback times. Here AI saves not only time but also improves response speed, and customers notice that.

Then there is classic data capture. Scanning and reading invoices, digitizing delivery notes, transferring business cards into a CRM. All repetitive, all well suited for automation.

Quality control is another strong candidate. Automatically inspecting product photos, detecting anomalies in measurement data. Especially in production, the potential here is enormous.

And finally: knowledge management. Making internal knowledge bases searchable, summarizing manuals, preparing onboarding material.

What does not work well: processes where the decision is complex, rare, and highly individual. Hiring, strategic partnerships, crisis management. There, AI is a tool but not autopilot.

My personal take on AI in SMBs

I am not an AI evangelist. I do not sell AI platforms. I am a software developer and founder, and I see every day what works for my clients and what does not.

My honest assessment: most SMBs do not need AI. They first need clean data, clear processes, and software that talks to each other. Once you have that, AI becomes a lever. Without it, AI becomes the next expensive disappointment.

That might sound counterintuitive from someone who builds software for SMBs. But that is exactly why I say it: I want my clients to succeed with AI. And the path leads through foundational work, not through hype.

At Instacart the first AI application that truly made money solved a tiny problem. It predicted the optimal order of items on the shopping list so the shopper walks fewer meters in the store. Not a glamorous project. No headline. But millions in saved labor hours per year. Start small, measure, scale.

That works for a 15-person shop in the Eifel too.

Your first AI step

If you have read this far, you take the topic seriously. Good. Here is what you can do this week:

1. Identify your most annoying process. Ask your team: "Which task annoys you the most?" Not the strategically most important one. The most annoying one. Because motivation is half the battle.

2. Measure the effort. How many hours per week? How many errors per month? What does it cost in euros? If you cannot answer that, here is your first step: measure for one week.

3. Check the data situation. Do you have the data an AI system would need? Is it digital? Is it consistent? If yes, you have a candidate for a pilot. If no, you know what to work on first.

4. Start small. Budget under EUR 5,000. Timeframe 30 days. One process, one team. If there is no measurable result after 30 days, stop and analyze why.

5. Get help if needed. But the right kind of help. Not a consultant who sells months-long assessments. Someone who works on the process with you in the first week.

If you have questions or want to assess whether AI makes sense for a specific process in your company: get in touch. We do this every day, for companies in the Eifel and the Rhineland.

And if you are wondering whether your employees need AI training: yes, they do. Since February 2025. More on that in our article about the EU AI Act and the training obligation.