The integration of AI into software development has revolutionised the field, enabling rapid code generation and prototyping. AI Full Stack Engineers - AI systems that can handle all frontend, backend and database tasks—have become invaluable for quickly iterating on ideas and building functional prototypes.

However, when it comes to developing production-class SaaS applications, these AI-driven tools reveal significant limitations. Production environments demand robustness, security, scalability, and compliance—qualities that require the nuanced judgment and oversight of experienced Software Developers.

This post argues that AI should serve as an augmentation tool for software developers rather than a replacement. We explore the concept of Vibe Coding⁸, where non-programmers, or Citizen Developers, collaborate with software developers under the guidance of an LLM.

This approach with Vibe Coding leverages the strengths of both AI and human expertise while building on the trusted Salesforce Lightning Platform, a PaaS designed to support secure and scalable SaaS applications.

AI Full Stack Engineers

AI Full Stack Engineers are AI systems capable of generating code across the entire software development stack, from user interfaces to databases. These tools, such as GitHub Copilot and Replit Ghostwriter, excel at prototyping due to their ability to quickly translate natural language prompts into functional code (Chen et al., 2021²).

For example, an AI Full Stack Engineer can generate a basic app dashboard in minutes, allowing teams to test concepts and gather feedback early in the development cycle.

However, whilst these prototypes are useful for validation, they often lack the depth required for production environments. Prototypes generated by AI may not account for edge cases, security vulnerabilities, or performance bottlenecks - all key issues that are critical in production-class applications (Dahlin, 2021³).

Thus, whilst AI Full Stack Engineers are powerful for initial development, they are not sufficient for delivering trusted, scalable SaaS apps.

Production-class SaaS applications must meet stringent requirements for security, scalability, and maintainability. These apps often handle sensitive data, integrate with complex systems, and serve large user bases—all of which demand a level of precision and reliability that AI alone cannot guarantee. Key limitations of AI in this context include:

Security Risks: AI-generated code may inadvertently introduce vulnerabilities, such as SQL injection or cross-site scripting, due to a lack of context about broader system (Perry et al., 2022⁶).

Scalability Challenges: AI tools may not optimize for performance or anticipate the need for horizontal scaling, leading to bottlenecks as user demand grows (Bass et al., 2019¹).

Business Logic Complexity: AI struggles to fully grasp intricate business rules or compliance requirements, which are often context-specific, require human judgment (Holstein et al., 2019⁴).

Maintenance and Debugging: AI-generated code can be difficult to maintain, as it may lack the modularity and documentation that human developers prioritize (Lwakatare et al., 2020⁵).

These limitations underscore the need for human oversight in the development of production-class SaaS apps.

Augmenting Software Developers

Rather than replacing Software Developers, AI should augment their capabilities, enhancing productivity and creativity. In this model, AI tools assist Software Developers by generating boilerplate code, suggesting optimisations, or identifying bugs, allowing Software Developers to focus on higher-level tasks such as architecture design and business logic (Dahlin, 2021³).

One promising approach is Vibe Coding, where a non-programmer (known as Citizen Developer) collaborates with one or more Software Developers to co-create a SaaS app. Guided by an LLM powering an AI Agent, the Citizen Developer can contribute domain expertise and high-level requirements, whilst the Software Developers ensure technical rigour.

This pairing leverages the strengths of both parties: the Citizen Developer’s understanding of the problem space and the Software Developer’s ability to implement scalable, secure solutions (Holstein et al., 2019⁴).

Adding Salesforce

As illustrated above, the complete strategy for combining Vibe Coding with Citizen Developers paired with Software Developers, requires committed stakeholders, a solid business case and a trusted underlying technology.

As the name implies: Design Thinking⁹ is thinking (and acting) like a designer. Being curious, restless and constantly challenging business-as-usual. It is all about solving problems in a human-oriented way. In order to generate receptivity and rapport, empathy is the key to success.

Inspired by the Stanford d.school, at Being Guided our Design Thinking is delivered as six iterative stages: Empathize; Define; Ideate; Prototype; Test; and, Implement. This offers a structured yet flexible framework to better understand and challenge assumptions and redefine problems.

Design Thinking maximises stakeholder engagement in digital innovation and is strengthened by Value Engineering. mapping out a solid business case for timely initiatives.

Value Engineering was originally conceived by Lawrence D. Miles¹⁰, a General Electric engineer. Miles' techniques have saved design engineers, manufacturing engineers, purchasing agents and service providers millions of dollars.

You can see more about how we apply Design Thinking and Value Engineering at Being Guided elsewhere in this Journal: Introducing XCELD.

A trusted Platform-as-a-Service (PaaS) is essential for building production-class SaaS apps. Salesforce Lightning Platform, for example, provides a secure, scalable foundation with built-in features such as:

Security and Compliance: Salesforce offers robust security controls, including encryption, access management, and compliance with standards like GDPR and HIPAA (Salesforce, 2023⁷).

Scalability: The platform automatically handles load balancing and scaling, ensuring that apps can support growing user bases without performance degradation (Bass et al., 2019¹).

Integration: Salesforce’s ecosystem enables seamless integration with other enterprise systems, a critical requirement for SaaS apps (Lwakatare et al., 2020⁵).

By building on a trusted PaaS, paired Citizen Developers and Software Developers can focus on application logic rather than infrastructure, further amplifying the benefits of AI augmentation versus reinventing wheels that characterises Full Stack Development.

Vibe Coding in Action

Consider a scenario where a knowledvge-intensive services company needs a custom SaaS app to manage client portfolios. Using Vibe Coding, a Citizen Developer (an analyst) describes the app’s requirements in natural language, and an LLM generates initial code snippets. A paired Software Developer then refines this code, ensuring it meets best practices, security standards and integrates with existing business logic on the Salesforce Lightning Platform.

In contrast, relying solely on an AI Full Stack Engineer might produce a functional prototype, but it would likely lack the necessary compliance features (e.g., audit trails) and scalability to handle thousands of users. The collaborative Vibe Coding approach, augmented by AI and built on a trusted PaaS, ensures the final product is both innovative and production-ready.

Conclusion

Whilst AI Full Stack Engineers are transformative for prototyping, they are not yet capable of delivering trusted, scalable, production-class SaaS applications without human oversight. The optimal approach is to use AI to augment Software Developers, leveraging collaborative methods like Vibe Coding, where Citizen Developers and Software Developers work together, guided by a Large Language Model (LLM) powering an AI Agent.

By building on a trusted PaaS such as the Salesforce Lightning Platform, at Being Guided we can ensure that your SaaS apps meet the rigorous demands of production-class environments. This combination of human expertise, AI augmentation, and a reliable platform represents the future of SaaS development.

This is also when at Beding Guided we apply Fierce Reduction: the practice of aggressively simplifying all business processes, tasks and information systems by removing redundant or non-essential elements before considering SaaS app development.

Next?

Let's Meet to explore this further.

References

1. Bass, L., Clements, P., & Kazman, R. (2019). Software architecture in practice (4th ed.). Addison-Wesley.
   https://www.oreilly.com/library/view/software-architecture-in/9780136885979/
2. Chen, M., Tworek, J., Jun, H., Yuan, Q., Pinto, H. P. D. O., Kaplan, J., ... & Zaremba, W. (2021). Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374.
   https://arxiv.org/abs/2107.03374
3. Dahlin, M. (2021). The role of AI in software engineering: A review. Journal of Systems and Software, 174, 110887.
   https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3633525
4. Holstein, K., Wortman Vaughan, J., Daumé III, H., Dudík, M., & Wallach, H. (2019). Improving fairness in machine learning systems: What do industry practitioners need? Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, 1-16.
   https://arxiv.org/abs/1812.05239
5. Lwakatare, L. E., Raj, A., Bosch, J., Olsson, H. H., & Crnkovic, I. (2020). A taxonomy of software engineering challenges for machine learning systems: An empirical investigation. Journal of Systems and Software, 162, 110496.
   https://colab.ws/articles/10.1007%2F978-3-030-19034-7_14
6. Perry, D. E., Porter, A. A., & Votta, L. G. (2022). Empirical studies of software engineering: A roadmap. ACM Computing Surveys, 54(3), 1-36.
   https://dl.acm.org/doi/10.1007/s10664-009-9121-0
7. Salesforce. (2023). Salesforce Security Guide. Retrieved from: https://developer.salesforce.com/docs/atlas.en-us.securityImplGuide.meta/securityImplGuide/security_guide.htm
8. Karpathy, A. [@karpathy]. (2025, February 2). There’s a new kind of coding I call “vibe coding”, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. [Post]. X. https://x.com/karpathy/status/1753472166197080428
9. The Hasso Plattner Institute of Design. (2004) Stanford d.school. https://dschool.stanford.edu/about
10. Miles, L.D. (1947). The Lawrence D. Miles Value Engineering Reference Center Collection.
    https://minds.wiscon.edu/handle/1793/301