Every breakthrough in technology tends to nail one or two things really well, but there’s usually a catch somewhere. Think about it, most innovations excel at making things faster, more efficient, or more reliable, but rarely all three at once.
AI has absolutely crushed it on speed and efficiency when it comes to building applications. What used to take weeks can now happen in minutes. But here’s the thing that keeps everyone up at night: AI is inherently unpredictable. It’s probabilistic by nature, not deterministic, which makes it feel unreliable for anything mission-critical.
And let’s be honest, when you’re building internal tools that your team actually needs to use every day, “mostly works” isn’t good enough.
At ToolJet, we’ve been tackling this exact problem. We’re an internal application-building platform that helps both technical and non-technical users create apps that actually solve business problems. And while we were already pretty good at the reliable part, we knew AI could make us incredibly fast and efficient.
The question was: how do you get AI’s superpowers without sacrificing the dependability that production applications demand?
Here’s how we’re tackling it:
- Understanding AI’s strengths & weaknesses — and designing workflows that play to both
- Transparency in the process — so users always know what’s happening and why
Let’s dive in.
AI’s strengths & weaknesses
“The first 90 percent of the code accounts for the first 90 percent of the development time. The remaining 10 percent of the code accounts for the other 90 percent of the development time.” — Tom Cargill on the 90/90 rule.
It is painfully true that the final stretch from ‘mostly working’ to ‘production ready’ is often the hardest part of any project.
And here’s the thing: AI is absolutely incredible at that first 90%. It can take you from zero to ‘holy crap, this actually works’ faster than any human ever could. But when it comes to that brutal final 10%, the precise tweaks, the edge cases, the ‘can you just move this button 2 pixels to the left’ moments. That’s where AI starts to struggle, and the back-and-forth conversation becomes inefficient compared to just clicking and dragging.
This is why we believe the future isn’t about AI doing everything, but about AI doing what it’s brilliant at (that explosive 0–90% creation phase) and then seamlessly handing off to humans for what we’re brilliant at (that meticulous 90–100% refinement phase).
We’ve used this principle directly in our product and split up the application-building process into two interfaces to mirror how people can actually get the best out of AI:
The creation phase: AI first interface
We start users with a clean chat interface paired with a live canvas. During this phase, everything runs through AI: no components, no menus, no distractions. Just you, the AI, and your ideas coming to life. This keeps the cognitive load low and lets AI do what it does best: turning your ideas into something real.
Creation phase with AI
The refinement phase: Visual interface
Once AI generates that first working version, the interface transforms to reveal all our low-code editing tools. Now you can click, drag, adjust colours, and fine-tune layouts with the precision that visual editing provides. Of course, AI doesn’t disappear; it’s still there when you need to make bulk changes or tackle something complex. But for those ‘make this button blue’ moments? Point, click, done.
Refinement phase with GUI
The result is that users get the best of both worlds: AI’s creative horsepower for the heavy lifting, and human-friendly tools for the detail work.
Balancing AI conversation with visual structure
We apply this same creation-meets-control principle even within the AI-first phase. While the AI is building your app, it needs to gather requirements and fill in gaps through conversation. But here’s the thing: even during this creative phase, it’s often easier to review and edit a structured document than to parse through long paragraphs of AI text.
That’s why our AI doesn’t just chat. It generates structured artefacts like the spec document you see in the image below. Instead of buried details in conversation, you get a clean, scannable overview of exactly what’s being built. Need to change something? Much easier to spot and edit when it’s laid out clearly.
App specification artefact
Guiding users with intent-aware actions
Another challenge with open-ended AI chat is that users can type literally anything, making it hard for us to predict what they’ll say next. But here’s what we realised: once someone starts building an app, we actually do know their intent; they want to move forward in the process.
Hence, we’ve added action buttons like “Next” that give users a clear, reliable path forward. The AI conversation becomes more focused, and users get that satisfying sense of progress without the guesswork.
Action button in chat
By understanding where AI shines and where it struggles, we’ve built an interface that plays to its strengths while covering its blind spots. This results in an AI that feels less like a magic black box and more like a reliable partner, one that knows when to take the lead, when to step back, and how to keep users confidently moving forward throughout the entire process. That’s how we’re making AI not just powerful but reliable.
Transparency
For those who have seen Silicon Valley, remember when Gilfoyle told AI to debug their code, and it just deleted everything?
It’s possible ‘Son of Anton’ decided that the most efficient way to get rid of all the bugs was to get rid of all the software, which is technically and statistically correct. But artificial neural nets are sort of a black box, so we’ll never know for sure.
Yeah, this is why transparency is a core part of our approach. We want users to see the steps AI is taking so they’re never left wondering, ‘wait, what just happened?’
So now, when a user tells ToolJet to build them an application, instead of immediately creating it, which may or may not align with the user’s vision, the process is broken down into two key checkpoints:
- Specifications of the app (think objective, navigation, core features)
- Setting up a database
This makes everything easier to review since it’s in digestible chunks. Users can examine and modify each piece to ensure the AI truly understands their requirements before moving forward.
That’s not all, we also show users exactly where they are in the process through a progress bar, giving them context about what’s coming next before the application is actually generated. And throughout each step, we display detailed information about how the AI is executing each task, so you’re always in the loop about what’s happening to your application.
Detailed loading state of each step
By making every AI decision visible and reviewable, we eliminate the biggest source of user anxiety: not knowing what the AI is doing or why. When users can see the reasoning behind each step and course-correct in real-time, they develop trust in the system. Transparency doesn’t just make AI more understandable, it makes it genuinely more reliable because users become active participants in ensuring the output matches their intent.
What’s next?
We’ve made AI reliable.
Next, we’re working on making it even more powerful, enabling complex modifications, not just simple suggestions.
Follow our GitHub repo to stay updated on what we’re refining next.
Stay tuned.
The post How ToolJet is making AI reliable appeared first on ToolJet.