How to survive your first AI project

I recently gave a talk at an internal company event of a Swiss startup. The topic was "How to build AI products (that people actually want)".

While preparing for this, I reflected on the AI projects I have done and asked myself: What were the things that worked? And why do AI projects rarely work as expected – especially first-time projects?

In today's newsletter, I'd like to share some of these insights with you so that your next AI project doesn't bite the dust.

Let’s dive in!

I’m still blown away by the response to my AI LIVE Workshop announcement last week. Thanks to everyone who’s joined and participated in our first session two days ago. If you didn’t have a chance to join, there’s still time!

I'm making a limited one-time payment option available to you, which is still half the price you'd normally pay for an average Maven cohort course. Check it out:

The Real Reasons AI Projects Fail

"80% of all AI projects fail" – this, or similar numbers are often floating around. I can confirm from my own experience that the fail rate for AI projects is much higher than for "non-AI projects" - probably twice as high.

So are all AI projects doomed for failure?

Let’s revisit that data critically.

First, most AI projects that fail are first-time AI projects. I think that makes a big difference. When you're doing something new for the first time, the chances of failure are inherently much higher than when you're doing things you've done a gazillion times. But there's more. The problem is that most organizations don't even get to the point where they can try their second, third, or fourth AI project because they burned all the "goodwill" when their first, often very ambitious, AI project failed.

That's why getting the first step right is absolutely critical. But before diving into what makes AI projects successful, let’s take a closer look at why they often fail in the first place.

1. Intrinsic Complexity

AI projects come with an inherent level of complexity, making them difficult to plan and execute effectively.

There are essentially three levels of project complexity:

First, “classical” IT projects often fail anyway — 17% fail so hard, they even threaten the company’s survival.

Next are data projects, which add "data" as a core component to IT projects, introducing additional complexity. Data can be unpredictable and hard to manage, leading data-heavy projects to fail even more frequently than traditional IT projects.

Finally, AI projects are essentially data projects combined with complex algorithms. This means they inherit all the challenges of classical IT and data projects, while introducing AI as another complexity layer.

Many decision-makers are unaware of this interdependency. They assume AI will simplify things, but in reality, it gets more difficult. Even rolling out seemingly "simple" stuff like Copilot or ChatGPT can introduce numerous unforeseen challenges — whether technical (such as inconsistent software outputs) or non-technical (like user resistance to adopting the AI solution). Underestimating complexity in AI projects is a major cause of failure. It only gets worse when combined with the second big mistake.

2. Unrealistic Expectations

Look how vendors like Microsoft announce their new AI solutions. It's utopia. AI handles all the tedious tasks, finds information, helps us make better decisions, and boosts productivity so much that we can finally relax and hit the beach. (Or complete even more work, depending on your preference.)

The problem is that these promises usually do not live up to reality. Seth Earley, whom I recently joined on the Earley AI Podcast, calls this "aspirational functionality" — a term I really like. Vendor demos typically showcase where they want their solutions to be "soon", not where they actually are right now.

So, if you’re managing your first AI project and communicate these same "aspirational functionality" internally, without clearly labeling them as such, it’s a recipe for disaster and unmet expectations.

3. The "Shiny Object" Syndrome

Once people become aware of AI (which is easy these days), it's only a matter of time before they come across an AI tool that grabs their attention. What follows is a "solution looking for problems" mode. A project is started simply to "drive innovation". This is known as the "shiny object syndrome". If there's "AI" in the name, decision makers often get so excited that they forget to ask if it's actually solving a real problem and if it's a good fit for the organization. Too often, it's not.

For example, if your company buys a predictive analytics tool simply because it sounds hot, chances are you don't have the data to make accurate predictions or the organizational analytics maturity to act on those predictions. At that point, the tool is essentially useless, no matter how well it's implemented.

4. Ignoring User Experience

User experience is often overlooked in AI projects. The reality is that even the most advanced AI solution won't be a hit if the people who are supposed to use it don't know how to interact with it.

Sure, you can offer onboarding workshops and tutorials, but the reality is that education alone can only drive adoption so far. For widespread adoption, everyone needs to intuitively "get it." AI –like all tech solutions–needs to meet users where they are. Whether it's a chatbot, a recommendation engine, or an internal tool, it needs to integrate seamlessly into existing workflows. This is a concept explored in depth in my co-authored book Augmented Analytics.

If your solution is too complex or confusing, most users will simply ignore it.

Strategies For Successful AI Projects

Now that we've learned why AI projects fail, let's focus on some key strategies to increase the chances that your AI projects will be a success – especially your first one.

1. Think Roadmaps, Not Projects

As explained in my last workshop, it's critical to have a backlog of AI use cases that you can continuously iterate on. Given the inherent complexity, AI projects will always have a higher risk of failure. That's why instead of aiming for large, big-budget moonshots, it's better to break your AI ambitions down into smaller, more manageable parts that can be prototyped as Atomic AI use cases. Because it makes a big difference whether your AI project fails after two weeks, two months, or two years.

However, this only works if your use cases are aligned with a clear roadmap so that even failures become valuable learning opportunities. Otherwise, you risk getting lost in the details and losing sight of the bigger picture.

Anticipate failure from the start, but fail smart.

2. Start with a Proven Problem

When working with new technology, there are already lots of unknowns. That's why you should focus on using AI to improve the way you solve an existing issue. Don't try to invent new problems. Leverage your internal knowledge. Look inside, not outside. Pick a well-understood problem and explore how applying AI might provide an advantage. Never innovate on both the problem and the solution at the same time (unless you don't care about quick ROI).

The Pain Points and Bottlenecks approach is a great way to get started. When in doubt, choose a use case that has already proven to work somewhere else!

3. Focus on Augmented Use Cases

Set aside any "AI doing all the work" fantasies for a moment. Instead, focus on using AI to assist and enhance existing workflows rather than replacing them entirely. Augmented AI use cases have a much higher chance of success than fully automated ones, and they often serve as a stepping stone toward more advanced solutions.

For examples, instead of diving straight into building a customer support chatbot to handle all incoming inquiries, why not start with a simple tool that helps your internal support agents respond to queries more efficiently? Once the tool is working well, you can gradually offer it as a self-service option.

Prioritize augmentation over automation to further reduce risk and get your first AI projects up and running faster.

4. Iterate Fast, Learn Faster, and Let People Know

Don't aim for a perfect AI system on your first try. Start small, iterate quickly, and learn from each iteration. Communicate this approach clearly to all project stakeholders!

The faster you get your AI into the hands of real users, the faster you'll gather valuable feedback that will help you improve the system.

Conclusion

Here's the most important thing to remember: Your first AI project shouldn't be your last - it should be the stepping stone into a larger journey. Treating AI projects like traditional software projects with a "ship and forget" mindset misses the point.

Create a roadmap with multiple use cases and aim for quick wins. This way, you reduce risk, learn from each iteration, and build confidence as you go.

Surviving your first AI project isn’t about making a huge splash with a groundbreaking solution. It’s about managing expectations, solving real problems, and building momentum with small, incremental wins.

If you need help with that, feel free to reach out or tune in to my upcoming workshops.

Good luck & see you next Friday!

Tobias