This is a really interesting angle. Removing the UI for data consumers feels counterintuitive but also kind of obvious once you think about it — most stakeholders just want answers, not dashboards.
How are you handling trust and validation of the outputs?
semantic-layer first approach, provide the tools for analytics/ai engineers to manage this well. It reduces errors and ambiguity to the absolute minimum
My brother Alex and I have been building in the data space for over a year. We started with AI-powered report creation, triggered analysis, deep-dive automation and some other stuff for data teams, finance, and ops. Custom products we'd build for each client. We had some traction, but the AI could never deliver accurately on the last 10%. Insane amounts of context and prompt engineering went into it, but clients didn't want it unless they had full trust in the numbers. Which is, frankly, impossible.
We figured out a few things.
1. If someone else builds your analysis, you're going to redo it anyway. That's just how people work. If I handed you a presentation someone else made, you'd probably rebuild it from scratch. And with numbers, it's even worse because you're going to pick them apart regardless.
2. Unless you're an analyst iterating on SQL, there is no point putting code in front of a business user. So they can learn and debug SQL? lol.
3. Self-serve failed because it was the UI that was self-serve and not the actual data. Business users do not want to learn another UI low-code tool. The world is moving in the opposite direction.
Business users should just use the tools they already know: ChatGPT/Codex, Claude Desktop, whatever. Those tools are already insanely powerful at delivering cheap UI. All you need to do is make the data self-serve, fast and reliable for agents. It feels pretty obvious that every business with data to operationalise should let people do whatever the hell they need to do with it to get the job done faster, right? It's not the data team's problem anymore cause the UI is cheap. The only thing that the data team needs to do with it is keep enriching and maintaining the self-serve data layer. They can then actually get time to do the impactful stuff.
So we built Bonnard. Data teams (Analytics Engineer, Data engineers or whoever has been handed the AI delivery project) go from data model to self-serve in about 5 minutes, all through the CLI. Zero UI on the consumer side unless they build it themselves.
Would love to get your feedback! Will hang in the comments if you have any questions.
This is a really interesting angle. Removing the UI for data consumers feels counterintuitive but also kind of obvious once you think about it — most stakeholders just want answers, not dashboards.
How are you handling trust and validation of the outputs?
semantic-layer first approach, provide the tools for analytics/ai engineers to manage this well. It reduces errors and ambiguity to the absolute minimum
My brother Alex and I have been building in the data space for over a year. We started with AI-powered report creation, triggered analysis, deep-dive automation and some other stuff for data teams, finance, and ops. Custom products we'd build for each client. We had some traction, but the AI could never deliver accurately on the last 10%. Insane amounts of context and prompt engineering went into it, but clients didn't want it unless they had full trust in the numbers. Which is, frankly, impossible.
We figured out a few things.
1. If someone else builds your analysis, you're going to redo it anyway. That's just how people work. If I handed you a presentation someone else made, you'd probably rebuild it from scratch. And with numbers, it's even worse because you're going to pick them apart regardless.
2. Unless you're an analyst iterating on SQL, there is no point putting code in front of a business user. So they can learn and debug SQL? lol.
3. Self-serve failed because it was the UI that was self-serve and not the actual data. Business users do not want to learn another UI low-code tool. The world is moving in the opposite direction.
Business users should just use the tools they already know: ChatGPT/Codex, Claude Desktop, whatever. Those tools are already insanely powerful at delivering cheap UI. All you need to do is make the data self-serve, fast and reliable for agents. It feels pretty obvious that every business with data to operationalise should let people do whatever the hell they need to do with it to get the job done faster, right? It's not the data team's problem anymore cause the UI is cheap. The only thing that the data team needs to do with it is keep enriching and maintaining the self-serve data layer. They can then actually get time to do the impactful stuff.
So we built Bonnard. Data teams (Analytics Engineer, Data engineers or whoever has been handed the AI delivery project) go from data model to self-serve in about 5 minutes, all through the CLI. Zero UI on the consumer side unless they build it themselves.
Would love to get your feedback! Will hang in the comments if you have any questions.
The trust problem is real. AI can generate analysis, but if users can’t trace the numbers themselves, they’ll always redo it manually.
Making data self-serve for agents instead of building another UI layer sounds like a much more sustainable direction.
Totally! this is what we had in mind, make sure data at source is built for agents, solves a lot of the downstream problems
people are using claude artifacts for this.
That is exactly why we built it so that people can use artifacts reliably!