I burned $120,000 and three months of engineering time on an AI dashboard that nobody used. Not a single person.
When I was building RemoteTeam, before Gusto acquired us, I thought I had the perfect vision for our analytics. I wanted a command center. I wanted predictive models flashing red and green. I wanted our investors to look at the screen and think we were operating in the year 2030. I hired a brilliant frontend developer and a data scientist, and we locked ourselves in a room to build the ultimate business intelligence tool. We were convinced we were building the future of work.
Instead, I got a very expensive lesson in what happens when you prioritize flashy visuals over actual business utility.
I've made over 200 angel investments in companies like Anthropic, OpenAI, Scale AI, and Hugging Face. I see pitch decks every single week promising the ultimate AI analytics solution. Most of them are making the exact same mistakes I made. They build complex systems that look amazing in a demo but fail completely in the real world. I saw this pattern earlier when my previous company MovieLaLa was acquired by Gfycat. Founders get obsessed with the tech and forget the user. They fall in love with the algorithm and ignore the human being who has to look at the screen every morning.
Here is the brutal truth about AI dashboards. Most of them are useless. They are vanity projects disguised as data science. But when you actually get them right, they can transform how a company operates. After that initial $120K failure, I tore everything down and started over. The second version wasn't as pretty, but it improved our team's decision-making speed by 40%.
Here are the brutal truths I learned about building AI dashboards that actually drive business growth.
1. Pretty Charts Are the Enemy of Action
The biggest trap in AI visualization is the desire to make things look cool. We spent weeks tweaking D3.js animations and building 3D scatter plots to show employee engagement metrics. We had heat maps that pulsed based on server load. We had network graphs showing how different teams interacted on Slack. It looked like a scene from a sci-fi movie. We were so proud of it.
But when our head of operations looked at it, she asked a simple question. "What am I supposed to do with this?"
She didn't care about the 3D scatter plot. She needed to know which remote employees were at risk of churning this month. That was it. A simple list sorted by probability would have been infinitely more valuable than our interactive galaxy of data points.
When you build an AI dashboard, every single pixel needs to answer a specific business question. If a chart doesn't immediately tell the user what action to take next, delete it. The best AI dashboards I see at companies like Scale AI don't look like spaceships. They look like incredibly smart, opinionated spreadsheets. They tell you exactly what is broken and exactly how to fix it.
I tell founders all the time to stop optimizing for the investor update and start optimizing for the Monday morning operations meeting. If your dashboard looks great on a slide but confuses your sales manager, you have built a toy. Real business tools are often boring to look at. They are dense with utility, not animations. The goal is not to entertain your employees. The goal is to make them faster and more accurate.
2. Your Data Is Messier Than You Think
You cannot build a smart dashboard on top of stupid data.
I see founders make this mistake constantly. They buy an expensive AI analytics tool, plug it into their messy CRM, and expect magic. They think the AI will somehow clean up the fact that their sales team hasn't logged a call properly in six months. They assume the machine learning model will magically infer the missing values and correct the timezone mismatches.
AI is not a janitor. If you feed it garbage, it will just give you highly sophisticated, beautifully visualized garbage.
Before we built the second version of our dashboard at RemoteTeam, we spent a full month just fixing our data pipelines. We focused on three core areas:
- Standardizing event tracking across all our web and mobile clients to ensure consistency.
- Forcing strict validation rules on user inputs to prevent bad data entry at the source.
- Purging historical data that we knew was corrupted, incomplete, or fundamentally flawed.
It was boring, tedious work. Nobody writes Medium posts about data cleaning. But it was the only reason our predictive models eventually worked. We found out that 30% of our historical user activity data was duplicated because of a bug in an old API endpoint. If we had trained our AI on that data, it would have given us wildly inaccurate churn predictions. We would have been making strategic decisions based on a software glitch.
If you are not willing to spend 80% of your time cleaning and structuring your data, you have no business building an AI dashboard. The glamorous part of AI is the final 5%. The other 95% is plumbing. You have to be willing to do the dirty work before you get to play with the shiny new models.
3. Complexity Kills Adoption
In my book "Becoming Top 1%", I talk about the importance of extreme focus. This applies directly to how you design internal tools.
Our first dashboard had 45 different metrics. We tracked everything from server response times to the average number of Slack messages sent per user. We tracked the time it took to close a support ticket, the bounce rate on our landing pages, and the utilization rate of our cloud infrastructure. We thought we were being comprehensive. In reality, we were just causing cognitive overload.
When people are overwhelmed by data, they ignore it. They go back to their gut feelings. They go back to asking the data team to pull custom reports. They open up Excel and start doing things manually because they don't trust the massive wall of numbers on the screen.
The dashboard that actually worked had exactly five numbers on it. Five.
We used AI to synthesize all the underlying complexity into those five core metrics. If one of those numbers looked wrong, you could click into it to see the underlying factors. But the default view was aggressively simple.
You have to treat your internal users with the same respect you treat your customers. If your dashboard requires a training manual, you have failed. The AI should be doing the hard work of synthesis, presenting only the signal and hiding the noise. It should act as an executive assistant, reading through thousands of pages of reports and handing you a single sheet of paper with the bullet points that actually matter. If your dashboard looks like an airplane cockpit, you are doing it wrong.
4. Real-Time Is Usually a Waste of Money
Everyone wants real-time data. It sounds impressive. But very few businesses actually need it.
We spent a massive amount of engineering resources trying to get our AI models to update predictions in real-time. We wanted the dashboard to reflect every single user action the second it happened. We set up Kafka streams. We built complex event-driven architectures. We wanted to see the needle move the exact moment a customer clicked a button.
Do you know how often our executive team actually looked at the dashboard? Once a day. During our morning standup.
We were paying massive cloud computing bills to update predictive models 24/7 for an audience that only checked the results at 9:00 AM. It was pure ego. We were building a system for a use case that didn't exist in our company.
Unless you are building a high-frequency trading algorithm or monitoring critical infrastructure, you probably don't need real-time AI analytics. Batch processing your data once a night is cheaper, more reliable, and perfectly adequate for 99% of business decisions. Save your money and spend it on hiring better engineers instead.
I see startups burning through their seed rounds paying for massive cloud infrastructure just to power real-time dashboards that nobody looks at on the weekends. It is a complete waste of capital. Build for the actual cadence of your business. If you make decisions weekly, your data only needs to be updated weekly. Stop paying the AWS tax for vanity metrics.
5. The AI Should Have an Opinion
A traditional dashboard tells you what happened. A good AI dashboard tells you what will happen. A great AI dashboard tells you what to do about it.
This was the biggest shift in our thinking. We stopped using AI just to draw trend lines and started using it to generate specific recommendations. We moved from descriptive analytics to prescriptive analytics.
Instead of showing a chart that said "User engagement is dropping," the dashboard would say "User engagement in the enterprise segment is down 12%. Based on historical patterns, you should run the re-engagement email sequence today."
This is where the 40% improvement in decision-making came from. We stopped forcing our team to interpret the data. We let the AI do the interpretation and present a clear hypothesis. The human operator just had to approve or reject the recommendation.
When I look at the AI startups I've invested in, the ones that grow the fastest are the ones that reduce cognitive load for their users. They don't just provide information. They provide answers. They understand that business leaders are tired of looking at charts and trying to guess what they mean. They want a system that acts like a trusted advisor, pointing out the problem and suggesting a solution. If your AI isn't sticking its neck out and making a recommendation, it is just a very expensive calculator.
6. The Build vs. Buy Dilemma is Real
When we started building our dashboard, we assumed we had to build everything from scratch. We thought our business was so unique that no off-the-shelf tool could possibly understand our metrics. This is a classic founder delusion.
We spent months writing custom Python scripts and building proprietary visualization components. We were essentially trying to reinvent Tableau or Looker, but with a fraction of the engineering resources. We wasted hundreds of hours debugging charting libraries instead of improving our core product.
What I learned is that you should only build the parts of the dashboard that represent your core competitive advantage. If your secret sauce is a proprietary machine learning model that predicts customer churn with 95% accuracy, build that model yourself. But do not build the charting library to display it.
Today, the ecosystem of AI tools is massive. You can plug your data warehouse into existing platforms that handle the visualization, the user permissions, and the basic anomaly detection out of the box. Your engineers should be focused on the unique business logic, not on making sure the tooltips render correctly on mobile devices.
Buy the infrastructure. Build the intelligence. Don't let your engineering team get distracted by building things that have already been commoditized.
7. Keep the Human in the Loop
There is a dangerous temptation to let the AI take the wheel completely. Once you have a dashboard that makes good recommendations, the next logical step seems to be automating the execution of those recommendations.
If the AI says we should send a re-engagement email, why not just let the AI send it automatically?
We tried this. It was a disaster. The AI lacked context. It didn't know that there was a major AWS outage that day, which was why engagement was down. It didn't know that it was a national holiday in our biggest market. It just saw a dip in the numbers and fired off thousands of emails that confused our customers and annoyed our sales team.
AI is incredibly good at processing large volumes of data and identifying patterns. It is terrible at understanding the messy, unpredictable context of the real world.
The best AI dashboards do not replace human judgment. They augment it. They serve up the data, highlight the anomalies, and suggest a course of action. But a human being must always make the final call. The human provides the context that the machine lacks. You need the speed of the algorithm combined with the intuition of the operator.
The Path Forward
Building an AI dashboard is not a technical challenge. It is a product design challenge.
You have to deeply understand the psychology of the people who will use it. You have to know what keeps them up at night. You have to know what decisions they are afraid to make. You have to understand their daily workflows and how they actually consume information.
If you start with the technology, if you start by asking "What cool things can we do with this new LLM?", you will fail. You will build a $120,000 toy that looks great in a pitch deck but gathers dust in the real world.
Start with the business problem. Talk to the person who actually has to make the decisions. Find out what single piece of information would make their job easier. Then, and only then, use AI to get them that information as simply and clearly as possible.
That is how you build something that actually matters. That is how you turn messy data into real business growth.
Frequently Asked Questions
How has this view evolved over time?
My thinking on most topics has changed significantly over the years. Early in my career, I held many conventional views that experience proved wrong. I try to update my beliefs when the evidence changes.
Do all experts agree with this view?
No, and that's fine. The best ideas in business are often contrarian. I share my perspective based on my experience and data, but I encourage you to seek out opposing viewpoints and form your own conclusions.
What experience informs this perspective?
This perspective comes from over a decade of building companies in Silicon Valley, two successful exits (RemoteTeam to Gusto, MovieLaLa to Gfycat), and investing in 200+ startups including Anthropic, OpenAI, and Scale AI. I write about what I've lived.
What's the most common pushback you get on this?
People often push back by citing exceptions or edge cases. And they're usually right that exceptions exist. But building a strategy around exceptions rather than patterns is a losing game for most founders.