SaaS vendors have a strategic challenge on their hands. Generative AI is being used to implement the kind of workflows and tasks that were the bread and butter of SaaS systems. Emerging coding tools are making it easier to build the data management capabilities that these systems are based on. The fear is that a new generation of "SaaS killers" will appear to disrupt existing vendors.

Despite this threat, AI offers more of an opportunity than a threat to established SaaS vendors. AI can breathe new life into established SaaS platforms in three important respects: it can accelerate feature development in complex code bases, allow vendors to modernise their systems, and - most importantly of all - enable new ways for users to engage by providing expert guidance and time-saving automation.

Pace of feature development

SaaS systems are notoriously difficult to maintain once they have accumulated a lot of customers. Some degree of technical debt is inevitable as the demands of the domain diverge from the design of the solution. The years spent accommodating customer demands as reusable functionality can leave a tangled mess of configuration switches and flags. Issues around performance, scalability, and resilience become more urgent, while technical entropy sets in as frameworks and tools drift out of active support.

Engineers move on and essential knowledge of the inner guts of the system is either lost or concentrated in a very small group of people. The result can be a byzantine mess that is often poorly understood by the engineers who are working on the system. Engineer productivity slows to a crawl and many problems with the system can feel too big to solve in any reasonable timeframe.

Given this context, there's little wonder that AI coding tools have generated so much excitement, even if this enthusiasm has been driven in part by the familiar "low code" myth of being able to dispense with engineers. The real benefits of coding tools are more nuanced than this.

The emerging generation of tools - Cursor, Replit, Claude Code and the like - are beginning to feel seriously impressive. We may be on our way to establishing a new set of abstractions that change the way that software engineers work, much in the same way that 3GLs saved us from having to work directly with assembly language.

Where these tools currently excel is in building applications from scratch or working with relatively small and self-contained code bases. They are less effective when it comes to the meat-and-drink of SaaS development, i.e. bug-fixing or enhancing a large and tangled code bases. The "work" here is in understanding complexity before making the necessary surgical intervention of a few lines of code. Given that LLMs can still be overwhelmed by too much context, the current crop of coding tools are less able to add value here.

There is an enormous prize waiting for anybody who manages to nail the problem of discerning context in large code bases. There are a lot of solutions out there who are using LLMs to parse code bases into graphs or beautiful visualisations, but so far these tools struggle to reliably navigate years of accumulated complexity.

AI coding tools are often justified by that slippery old concept of developer productivity, but it's important to understand that software systems are more than just code. They are based around groups of people who build a shared mental model around how the domain works. This is especially the case for SaaS as customers buy into a service model made of software and the team of engineers who write and operate the software.

The fight against entropy and obsolescence

AI coding tools may not replace engineers, but they have already made them more productive in a way that brings new projects into view. This is particularly the case for managing larger legacy code bases that suffer from creeping technical obsolescence. Frameworks go out of date, some technology choices don't age well, but the cost of change is often too great to contemplate any migration.

AI's code generation capabilities make some of the more common technical improvement challenges for SaaS systems feel possible. You can deliver major migrations to manage cost, such as moving from SQL Server to Postgres, despite having to port over thousands of stored procedures. You are not trapped on older, abandoned versions of UI frameworks such as AngularJS. You can accelerate the "lift and shift" work required to move from older versions of Python, Java, or .Net.

LLMs won't completely automate the work of upgrades or migrations. In each case, you still need engineers to iron out the creases, but at least 80% of the work can be done through automation. These projects become actionable, where before they were time sinks that never survived any prioritisation discussions.

AI can help to extend the lifetime of systems by requiring less effort to keep them up-to-date. More engineering time can be expended on building new features and adding new value. However, the real prize for SaaS vendors may have less to do with engineering productivity and fighting technical obsolescence. It offers a chance to change the relationship with the user by transforming systems into expert guides and assistants.

Agentic enhancement

Despite AI's potential for accelerating development, it's unlikely to usher in a new generation of disrupters that sweep away established vendors. Many SaaS systems embody years of accumulated knowledge based on a deep understanding of the processes and usage patterns that underpin the domain. SaaS isn't just about capabilities, as there is a trust and compliance barrier based on years of security certifications and customer relationships. This sets them apart and creates a moat that can protect them against disruptive AI-driven entrants.

AI offers a chance to breathe new life into SaaS systems by creating more intuitive ways of interacting with them. Many SaaS applications model complex processes so have an impenetrable UX to match. Vendors can build an agentic layer to leverage the years of accumulated domain expertise and battle-tested capabilities and offer a natural language interface.

A conversational UI based on tool-using agents could become the dominant means by which users engage with SaaS applications. The underlying application will always be there, but a conversational assistant will be guiding the user, providing expert guidance, and automating away the more tedious and repetitive tasks. The user should always remain in control, while the agents support them in making better decisions and being more productive.

The demand on established code bases will change as task orchestration will increasingly be delegated to an agentic layer. The focus will be more on exposing data and behaviour, while agents work with users to plan and execute workflows. The underlying architectures may not change, while the way in which users interact with them can be transformed without needing to make costly enhancements to complex code bases.

Until recently, the long-term lifecycle for SaaS applications often involved a long, slow heat death, with the system eventually being overwhelmed by technical obsolescence. Now there is an opportunity to transform them into a foundation layer for agentic domain experts that help people to navigate complex processes. SaaS vendors have an opportunity to thrive by embracing AI and using it to make all that accumulated knowledge more accessible to their users.

Filed under AI, Agents, Strategy and Rants.