You’ve heard the question: “Why not just vibe code it?”
So let’s go at it head-on. Ridge AI or Claude Code?
Spoiler alert: It's not either-or. Claude is great for ad-hoc, directional analytics. It’s a fantastic tool and the foundation for a lot of what is achievable in AI today.
But when you need something high-quality, lasting, and shareable, choose an AI-native product that combines that power with a purpose-built harness.
Common issues with vibe-coded analytics
1. Deployment is a headache: maintenance, security, updates.
This is the easiest factor to underestimate. Not the chart itself, but everything around it: auth, permissions, data modeling, performance, partitioning, embedding, data refresh, monitoring, and maintenance.
In a Reddit thread asking, “Anyone here tried building embedded analytics in-house? Regrets or worth it?” one commenter put it well:
“The charts aren’t the hard part, it’s auth, permissions, data models, and performance that eat your time. Teams I’ve seen build in house got flexibility but also signed up for permanent maintenance.”
Ouch.
2. AI can make hidden decisions.
An LLM may make choices you didn’t want or expect:
- It may choose the wrong aggregation.
- It may use the wrong denominator.
- It may sort time incorrectly.
- It may hard-code a number that looks right but is not.
- It may create a beautiful view that tells the wrong story.
When Claude makes mistakes, they can be non-obvious. The dashboard looks great. It gets shared. Unless someone checks the logic carefully, the mistake can become part of the story. In analytics, that is dangerous.
Andy Cotgreave wrote a great post on this exact issue: Hallucinations in AI Analytics: Still Real and Dangerous. In his example, Claude generated a polished dashboard with numbers that looked compelling. But a key number was wrong. Worse, it was hard-coded in the HTML. Once the mistake was in the dashboard, it became permanent.
I’ve seen this happen with interactivity as well. Some charts change, others don’t, and it’s not obvious why.
At Ridge, we make assumptions explicit and build analytics to be natively interactive. Numbers are never hard-coded, and metrics and charts respond to filters and selections. You can always inspect the SQL behind queries in the Data Agent to verify what’s there.
3. Quality needs guardrails.
Many product leaders I’ve spoken to are excited about adding a data agent to their product so customers can ask questions of their data. AI enables powerful, open-ended experiences. But that power needs a harness that:
- Validates the schema
- Validates the SQL
- Detects failure and retries
- Presents understandable visual results
- Applies good old-fashioned UX design
- Helps the end user understand the story in the data and trust the results
- Monitors quality over time
This is where a demo and a product are very different.
At Ridge, as we built our Data Agent and Build Agent, we found that they needed all of the above to reach a consistent level of quality. Our product ships with a purpose-built harness that keeps improving over time.
A demo can be impressive with a handful of good answers. A product needs to keep working across many users, many questions, many datasets, and many edge cases.
Good analytics need a governed data environment, both for security and also for the "wait, what number is that?" problem. That's achievable with CLaude, but it may take another few teams of people to rebuild that governance into a form that supports everyone Clauding their way to an answer.
And here’s one thing you don’t want: human as harness. Some analytics teams are becoming the “Check this for me, will you?” layer for analytics built with Claude. That’s not saving time, that’s shifting time to another team.
4. Shared understanding suffers.
When everyone is doing bespoke analytics, there’s no shared artifact. That usually means there’s no shared understanding.
Great analytics also reflect visual best practices, so they are as understandable as possible to as many people as possible. This is especially important for embedded analytics, where users are often not analysts by trade. Customer-facing reporting has to help your customers understand what’s happening when your team isn’t in the room.
At Ridge, we apply decades of best practices in how people understand data to time, categories, aggregation, and more, so the result is easy for your audience to understand.
You can see this in how the Data Agent responds to different questions with different kinds of charts:
5. It takes massive amounts of time.
Building something that does all these things in Claude Code is certainly possible. But it is a full-time job for more than one person.
Anthropic’s post on its own internal analytics does not describe a low-toil project. In fact, the number of processes they manage and people involved is overwhelming. Granted, this is for their entire analytics practice, not just embedded analytics.
Others who have built apps and blogged about them report the same pattern. A very experienced practitioner can build a first draft in a matter of hours, but it takes taste and experience to guide the AI. Then you still have maintenance, updates, and hardening to do.
If you’re trying to build this in-house, you’re incurring an opportunity cost in lost time and focus.
Who wins?
This is less about “which is always better?” and more about the situation.
Claude is fantastic for ad-hoc, one-time, directional analysis. When you’re exploring an area and wondering if it’s worth analyzing further, AI-based analytics can give you exactly what you want, if you guide it carefully.
But when the stakes are higher, use a purpose-built solution.
Choose Ridge when:
- You want embedded, durable, and reliable analytics
- Your analytics need to be ready for a broad internal or external audience
- The experience will represent your brand or product
- You need high-quality results, custom theming, and strong performance
- You don’t want your team spending time building and maintaining all the things around the chart
Don’t use Claude directly for that. Use a purpose-built solution like Ridge AI.
After decades of working in analytics, we know analytics need structure. They need guardrails. They need visual best practices. They need deployment, permissions, performance, and maintenance handled as part of the product experience.
We believe AI can make working with data delightful and easy. That’s the bar you should hold.
