Every app is adding AI right now. On the surface, that seems like an obvious win: AI makes complex interfaces easier to use, expands who can use them, and helps existing users work faster.
But most companies aren’t rethinking their products from an AI-first perspective. They’re adding intelligence to individual tools without reconsidering how the work actually gets done.
I think of this as the difference between local AI and global AI.
Local AI
Each tool gets smarter, but the boundaries between tools remain. The user still has to move context, coordinate people, and stitch the answer together.
Global AI
The system is designed around the full outcome. The AI can move across the relevant context, take action, validate the result, and keep working.
Making every tool smarter isn’t enough
Consider the modern data stack.
Your BI tool—whether it’s Looker, Power BI, or ThoughtSpot—might use AI to help you build dashboards incredibly fast.
Your transformation layer, such as dbt, might use AI to help you write transformations.
Your data dictionary might use AI to help people find the right metric and understand how it’s defined.
Your project management system might use AI to organize everything the team is working on.
Your meeting-notes app might use AI to capture the context behind important decisions.
Each of these is a legitimate improvement. But taken together, they reveal a deeper problem: a human being is still required to connect all the systems.
Imagine asking the AI inside your BI tool:
Why is this number on the dashboard wrong?
That AI can inspect the dashboard, but it may not be able to inspect the transformation that produced the number. Someone must take the question—and all its context—to the person working in dbt. That person then investigates the transformation, perhaps discovering that they need information from the data dictionary, an external pipeline, a project ticket, or a recent meeting.
You may have made every individual step more efficient without making the overall problem meaningfully faster to solve.
The AI takes five minutes. Getting someone on the data engineering team to prioritize the question might take two days.
The real cost isn’t always the work itself. It’s the context switching, coordination, waiting, and repeated transfer of information between people and systems.
Global AIAI-first products should be designed around the entire flow
Traditional software is organized around tools. AI-first software should be organized around outcomes.
Instead of asking, “How do we add AI to this dashboard builder?” we should ask:
What would an AI need to investigate, solve, and validate a data problem from beginning to end?
To do that, the AI needs substantially more context.
It needs to understand:
- ✓How the dashboard is built and who uses it
- ✓What every element on the dashboard represents
- ✓How each metric is defined
- ✓Which transformations produce those metrics
- ✓What queries and underlying building blocks are involved
- ✓Whether the data comes from internal or external sources
- ✓When the source was last updated
- ✓When uploaded data was last refreshed
- ✓How the data is maintained and monitored
- ✓Whether an API or enrichment pipeline sits upstream
- ✓Whether a failure in that pipeline could explain the incorrect number
Crucially, the AI must be able to act across that entire system—not merely describe what someone else should do next.
It should be able to move from the lowest level of the data to the dashboard, and from the dashboard back down to the source. It should be able to build, test, deploy, inspect the results, validate the numbers, and revise its work.
That’s the difference between making a tool smarter and creating a system that can actually do the work.
What we built at Polyform
This is why we built Polyform as an end-to-end system.
When we give our AI work to do, it has the context and capabilities necessary to execute across the full data lifecycle. It can investigate the underlying data, understand transformations and business definitions, build the output, test it, deploy it, inspect the resulting dashboard, and validate the numbers.
The value doesn’t come from adding a chatbot to every part of the data stack. It comes from eliminating the boundaries that force humans to carry context between those parts.
I believe more companies need to think this way. AI-first product design isn’t about making each existing interface slightly easier to use. It’s about designing the system so an AI has the global context and agency required to deliver the outcome.
And yes, I know what you’re thinking:
Shouldn’t ChatGPT or Claude be the brain, with every other system connected through MCP?
I’m writing about that next.