The low-code/no-code industry stands at a crossroads. As artificial intelligence reshapes software development, traditional low-code platforms face both existential challenges and unprecedented opportunities. In a world where generating code from an idea is cheaper and easier than even dragging and dropping, it's time for a deep rethink of what low-code development means in an AI-first world.

The Existential Challenge

Historically, low code platforms have had two core value propositions that have now eroded:

• Convenience: Build once, use over and over. Low-code platforms enabled repeat usage of pre-built core components of the dev stack, drastically reducing time to delivery. The rise of AI coding assistants has begun to challenge this core value proposition of low-code platforms. Tools like Windsurf, Cursor and Devin can now translate natural language directly into functional code, promising an even more optimal path of development than low-code with the retained flexibility of traditional programming.
• Reliability in Abstraction: Built well for you. By providing well-developed high quality building blocks, low-code platforms provided greater determinism and reliability in the applications created through them. This value has been greatly undermined by their initial approach of simply wrapping ChatGPT or similar models into their platforms When users drag and drop components in traditional low-code environments, they expect consistent, reliable behavior, but the unfiltered probabilistic nature of AI models has seeped through to these apps, resulting in near-complete erosion of reliable outcomes.

This raises a crucial question: In a world where AI can generate production-ready code from plain English in minutes and a low-code app suffers the unreliability equivalent to their worst prompt, what role do low-code development platforms play?

The Path Forward

The story might look grim, but it isn't necessarily one of inevitable disruption, but of a potential reimagination. Forward-thinking low-code platforms should rethink their role in an AI-powered development ecosystem along three dimensions:

AI for Low-Code Development

Low-code platforms, at their cores, are orchestrations over pre-defined code blocks. Low code platforms hold the unique benefit of having access and distribution into companies, as well as integrations with customers’ acceptable software utilities. They should explore 'guardrailed codegen' as an offering - merging the best of both worlds. This can manifest as the following:

• Words2Workflows: Generating a low-code application constrained in the customer's utility universe from prompts, to enable AI-driven development securely and prevent misuse of packages, DBs, etc with significantly lesser debugging needed. They could leverage their vast distribution and data on desired applications created previously to finetune specialized models for these outputs, auto-regressing over their existing components.
• NLI for Existing App Refinement: Low-code platforms can embrace the greater freedom offered through code generation by making it easier for developers to transform individual components of their application through text.
• Automated Quality assurance: Developer productivity apps - the core use-case for low-code, have long suffered from a lack of QA focus. Low-code platforms could autogenerate and run test-cases for apps, measuring regressions and suggesting optimizations.

This approach enables low-code platforms to lean further into, instead of away from, their existing core proposition of reliable convenience in app development.

AI in Low-Code Applications

Almost every low-code developer is looking to integrate AI into their applications. Whether it be structured data extraction, summarization or some other task - the applications of AI in the low-code world are endless. But the answer to enabling endless possibilities is not as simple as just an input box for a ChatGPT prompts. Here's how low-code companies should be thinking about offering AI capabilities as app integrations:

• Specialized, Reliable AI components: Instead of relying on general-purpose AI models, leading platforms should be developing specialized, fine-tuned AI capabilities that ensure more deterministic output for specific tasks. This maintains the predictability users expect while leveraging AI's power for common tasks, just like these platforms did for UI components. Be opinionated, be reliable.
• Streamlined finetuning: For tasks where fine-tuned components don't already exist, they could stick to enabling a starting prompt window, but provide a one-click switch to greater reliability fine-tuned models once sufficient input-output data is collected. (auto-finetuning)
• Pre-build Prompt templates: In the absence of the willingness to invest in fine-tuned components, low-code platforms could at the least provide well-fleshed out prompt templates for basic tasks and ensure that they are well-maintained even with constantly fluctuating base models.

This would enable low-code platforms to continue to keep their promise of greater app-level consistency while enabling access to cutting-edge AI integrations.

Low-Code Development for AI

Perhaps equally importantly, low-code platforms are in pole position themselves to emerge as essential infrastructure for AI development. Here’s how:

• Data labelling interface generation: One emerging need for AI teams is labelling smaller amounts of data internally along custom dimensions. In the past, most data was labeled across standardized dimensions in high volume - incentivizing a services industry (e.g Scale AI). But with AI, especially PEFT, teams often only need to label a few hundred samples internally along completely new dimensions each time - making this a perfect use-case for low-code apps to integrate themselves into the AI lifecycle.
• AI Application Monitoring: Most AI monitoring and observability solutions are very akin to the products of low-code platforms and there is great willingness to build vs buy from application layer companies - given the hyper-specific nature of their monitoring needs (much like labelling). By providing a base application template for low-code testing and monitoring of prompts, these platforms could lower the barrier to the build case for AI evaluation.

This leans into the ideal product profile for low-code applications: variable orchestration pathways with consistent building blocks and desired outcomes that are not necessarily in the critical product path - to open up a new, highly lucrative market.

The Future of Low-Code

To survive in the coming decade, low-code platforms will need to ensure they do two things:

• Embrace, not compete with, AI-augmented development
• Prohibit AI from eroding their core value proposition of reliable determinism.

This will require low-code platforms to think deeply about integrating AI into their product and leaning into their deep data reservoirs and customer integrations - ensuring that they knowledgeably trade off between exposing third party AI systems and building their own. For businesses and developers, this transformation means access to more powerful, intelligent development tools that combine the best of both worlds: the accessibility of low-code and the sophistication of AI-powered development. The low-code industry isn't dying – it's evolving. The platforms that embrace this evolution, treating AI not as a threat but as a catalyst for innovation, will define the future of software development.

If you're currently developing AI for a low-code platform and exploring how to build meaningful AI capabilities, reach out! Emissary is an AI infrastructure platform that greatly simplifies developing reliable and optimized AI models through model specialization (finetuning + much more) and we'd love to chat about how we can help realize this transition :)