Software Is Not Dead, But Will be Very Different

Introduction: The Next Generation of Software

The software market is not dead; it is being reborn around AI. There are over 10,000 VC-funded AI companies, most of them AI-native software companies with plans to encroach upon the roughly 400-billion-dollar global SaaS market. The larger, more established SaaS vendors are at risk, and their declining stock valuations underscore the urgency to redesign their systems around an LLM (foundation model) architecture and a reasoning-driven agentics framework. These vendors disrupted the prior generation of client/server companies, so they understand the disruptive impact and opportunities of technology platform shifts and the need for more extensive rearchitecting versus a more superficial bolt-on of some new features.

We are not concluding that incumbent SaaS vendors will rearchitect and win against the upcoming generation of AI-native software companies and a fragmented agentics market. We are saying they “get it” and are rearchitecting to have a better chance of adapting. They will be disrupted, but, as Bain has characterized it, they are fighting to avoid obsolescence.

Software will expand dramatically as a result of these new automation technologies; it will just look a lot different. At the applications layer, there may be greater encroachment by AI-native vendors with new processes enabled by intelligence (LLMs) at the core, and there may be greater fragmentation in the market versus traditional suites, as third-party agents encroach and impact growth rates for established vendors.

We saw this with the move from client/server to the cloud/SaaS model, where best-of-breed software solutions were embraced in the market as an alternative to the traditional monolithic suites. We may see further disaggregation with the platform shift to AI, as agents enable a fragmentation of functions and an unbundling of existing SaaS functions. There may be cannibalization as third-party agents encroach and impair growth opportunities in what is likely to be a more fragmented, unbundled market compared with traditional suite solutions.

From Bolt-On AI to Architecture Reborn

We will discuss the upcoming architectural changes and the need to do more than bolt on incremental agents to automate narrow legacy human tasks. Bolt-on functions may be quicker and look good more quickly in the marketing literature, but they risk missing the more critical and transformative opportunity to rearchitect systems around LLM intelligence and reasoning at the core and the orchestration of evolving autonomous agents.

Traditional SaaS companies have hard-coded processes that reflect industry best practices for workflows. Their user interfaces are human centric, menu driven and human click-through oriented. With AI-based systems, the user interface becomes more conversational. More importantly, it becomes possible to embed the LLM intelligence engine at the core to orchestrate agents that can reason, make decisions and take actions to perform tasks traditionally performed by humans.

Coding is automated but one should not think software is dead because coding is easier. There is a lot more work to be done in building enterprise systems than coding. Domain expertise is required to design processes and workflows; data access and management is more complicated as are proper governance and security. Global systems must reflect local conventions, not only languages but also accounting conventions differ across geographies. Complex workflows and integrations with data sources and other applications are also difficult and time consuming and should not be trivialized.

Over time the infrastructure AI innovations will likely erode many moats, however today they remain meaningful. We should not assess the changes based on what we can see today, this is a 5 – 10-year process of entering a market with a new computing platform and trying to break down the castle walls one stone at a time. Typically, applications competitors start with the low hanging fruit, the lower end of the market, where there is a fresher market opportunity with less invested in incumbent systems. We saw this with Salesforce and Workday in the last cycle.

There are process automation and workflow automation vendors addressing the difficult constraints in building enterprise systems so it is helpful to think of the upcoming disruption in the market over a multi-year timeframe, which makes the job of the investor harder to assess the eventual extent of the disruption for incumbent vendors.

AI Enables New Processes

AI enables new processes. For example, traditional customer support consists of menu based IVR front ends, frustrating cascading menus and eventually a human representative who may not be easy to understand or entirely knowledgeable. Users frequently yell “representative” in frustration just to get out of the captive menu system and speak to a person who might better understand the nature of the support call.

With AI there is a better way. Knowledgeable automated voice support virtual agents can be available 24x7, speak whichever language you choose, and be trained on all of the customer support manuals and potentially all prior support tickets and case studies. With memory and persistence, they can maintain a record of the customer interaction and next time deliver a more personalized, concierge experience.

Because intelligence sits at the core, the agent may identify that the customer is using an older version of the product that is no longer supported. The system can pivot from a cost center support function to a revenue function by recommending an upgrade. If integrated across systems, the agent can seamlessly pivot to selling a new product while also addressing billing and accounting as interoperable parts of the workflow. Agents will increasingly orchestrate complex multistep workflows that run across different systems and, with the right infrastructure, reduce the role of human involvement.

In CRM, the LLM can orchestrate agents to replace previously human tasks such as searching LinkedIn to identify job changes that result in new leads, reading emails and listening to conversations to update customer records, and drafting personalized emails for follow-up and marketing.

We recently watched a demo of an AI Native procurement system. The new capabilities were truly transformational and greatly productivity enhancing, perhaps analogous to someone moving from a typewriter to a word processor or manual ledgers to Excel. Software is not dead, it is being reborn.

Supporting Infrastructure and Agentic Frameworks

Large Language Models at the core are like the brain of an intelligent system. Agents can observe, reason, make decisions and take actions. As they get better, they will begin to orchestrate more complex, multistep workflows and, with proper guardrails such as governance, security and permissioning, can become increasingly autonomous.

These systems are probabilistic by nature of the LLM, so they are not always accurate as in a deterministic system. Human in the loop will be required to supervise agents until there are advances in reasoning, monitoring and reliability.

Agentic frameworks are the software foundations that define how AI systems reason, plan and take actions. They may support multiple agents in a system and manage how they interoperate. There are many different agentic frameworks, from hyperscalers, from LLM providers, from enterprise software companies, RPA vendors and other third parties.

The leading SaaS vendors have their own agentic frameworks that work with their respective data models and address how agents are governed, orchestrated and allowed to interoperate across the product platform. Two important de facto industry standards include MCP, which defines how agents gain access to data and other resources, and A2A, which defines how agents communicate with other agents across systems.

For fully autonomous systems, it is critical that agents can access data and talk with each other across systems. This sounds rather straightforward, but in practice it is not. What field in the data base is sales. Is sales the same as revenue. How do we define ARR or net retention rates. This is defined in a semantic layer which can insulate agents from complexities of the data model. Having consistency across companies is a particular challenge for automation across multi-vendor supply chains. Some fields of the database must be protected if confidential, such as payroll data or health records. What is the governance policy for each agent to access data.

Data across disparate systems and silos is inherently not uniform, incomplete and messy. The problem becomes much more difficult if agents from different vendors need to interoperate and share a common data model. Exposing data in a clean and consistently formatted way is a goal that most enterprises have not yet achieved internally, let alone across vendor supply chains.

As agents support more and more processes, more of the process logic migrates into the agentic layer. This evolution is productivity enhancing, but it requires significant work at the infrastructure layer to ensure governance, monitoring and security are robust.

Context graphs are helpful in understanding workflows. knowledge graphs are helpful in understanding relationships so one has context to better understand the data, which is very important and easy to demonstrate if we think of the needs of the intelligence community to understand relationships among people and for a bank to monitor related entities for money laundering or for export controls to understand potentially complex webs of related entities.

Incumbent Moats and Market Fragmentation

Incumbent SaaS vendors have protective moats. Notably, they control the customers’ data and are very restrictive in permitting others to leverage this data. They assert this is for security reasons and this is true, but it also provides a significant competitive moat. There is considerable controversy about restrictions in allowing customers to take their data out of SaaS systems and this can fuel concerns of vendor lock-in.

Domain expertise reflected in workflows can be extensive, so delivering full featured systems takes time for others to replicate. Typically, new vendors start at the low end of the market where there is less invested in specific vendors’ systems and functional requirements are fewer. As we saw with salesforce.com and Workday, vendors start at the lower end of the market and migrate up market over time with greater functionality and distribution scale.

It used to be that complex workflows were a competitive moat, but there are a number of vendors now with low-code workflow automation tools and context graphs that will make it easier to erode this moat. One must acknowledge that disruption is inevitable, so it is a question of whether incumbents are motivated and capable of moving fast and breaking things, including their per-seat business models, and reposition to leverage their market presence, distribution and scale advantages to be net beneficiaries of an expanding but changing market.

When the market transitioned from on-premise and client/server to the cloud, there was a shift to best-of-breed solutions versus integration or suites. The market requirement of a suite was diminished as customers found they could run SaaS solutions from different vendors. We suspect that the same software unbundling or disaggregation will be true with the migration to AI solutions. There is likely to be greater fragmentation so AI-native companies with robust solutions may see less friction in market adoption with superior agentic capabilities and interoperability across systems.

Incumbent SaaS vendors may see more fragmented competition from agentics players rather than head on competition for fully featured competitors. This is particularly true since most SaaS vendors have extended their product lines into broader suite solutions over the past 20 years. Competition from a fragmented market of agentics companies would not necessarily displace a SaaS vendor, but it might encroach, cannibalize and dampen growth opportunities in what is likely to be a more fragmented market versus traditional suite solutions.

We welcome your thoughts and a follow-up conversation.