Please refer to our disclaimers, which can be found in the footnote of this page and here.

Contents

• Contents
• Will AI Replace Enterprise Software? Findings from CIO and CTO Interviews
• Strategic Systems Are Not Being Ripped Out
• AI Point Solutions
• AI-Native Competitors
• The Necessity of a Credible AI Roadmap
• Agentic Revenues Have a Long Ramp
• Updated AI Defence Framework
• Final Thoughts
• Software Disruption Update 4 Nov 2026

Will AI Replace Enterprise Software? Findings from CIO and CTO Interviews

The AI hype has never been stronger. We hear it in discussions and see it in articles comparing how much Anthropic and SpaceX exposure a portfolio has. We’ve heard commentators call for both companies to reach valuations in the tens of trillions. While there might be scenarios where this proves true, it ignores the many scenarios where it isn’t (e.g. if models struggle to differentiate themselves) and extrapolates the stratospheric growth across the entire global knowledge economy. The market sentiment feels distinctly euphoric, especially given how rapidly the title of "leading model" changes hands. Investors have never been so certain that Anthropic is the undisputed winner in AI, ignoring that only several months ago the same conviction was assigned to OpenAI.

It is reminiscent of the March 1998 Fortune Magazine cover declaring "How Yahoo! Won the Search Wars." Indeed, reading financial media from the 1998-2000 timeframe reveals many of the same bold, absolute predictions being made today (another good example here). Ultimately, that period proved to be a terrible time to be invested in equity markets, taking over 12 and 13 years for the S&P 500 and MSCI ACWI to recover respectively on an inflation-adjusted basis.

It has been painful on a relative basis not to own certain AI names. Equally, on both an absolute and relative basis, it has been painful to own software. We had around 33% software exposure going into SAASmageddon and currently hold around 43%. We have been adding capital selectively to existing software positions and introduced one new position in Guidewire. We believe this is one of the most attractive times to invest in select, high-quality public software companies. We would not say the same for private software companies as valuations are much higher and they are generally lower in quality.

Clearly the market disagrees with us. In our last letter, we wrote briefly on the weakness in the software sector and outlined the specific characteristics we demand in software companies to withstand the threat of AI. We concluded that strategic software vendors with strong switching costs (extending beyond mere technical integration), strict data control, and a credible AI roadmap should survive, if not thrive, by capturing greater revenue opportunities and realizing lower internal costs.

Are we being naïve? Throughout the quarter, we continued to retest our hypothesis by conducting further interviews with CIOs, CTOs, and IT professionals overseeing agentic AI deployments at large enterprises, including large tech companies, global banks, payments companies, tier 1 insurers, and telecoms to name a few. We share our insights below.

Strategic Systems Are Not Being Ripped Out

Most enterprises we spoke with are experimenting with internal AI or AI-native tools, but very few are planning to replace their strategic software vendors, such as core ERP or CRM systems. This reluctance stems from several structural realities:

• Business Disruption: These are strategic vendors, and there are real costs and risks to switching. It requires redesigning both human and technical processes and retraining staff. The new solution needs to be a step-change better to justify switching.
• Complexity and Security: Running true enterprise-grade software with the required security, data isolation, and compliance controls is exceptionally complex.
• Lack of Internal Capabilities: Many organizations cited a fundamental lack of technical expertise to build and govern these systems internally.
• Higher Total Cost of Ownership (“TCO”): When factoring in compute, maintenance, and technical debt, internal replacement often results in higher TCO.
• Opportunity Cost: Internal builds carry hidden opportunity costs, including slower time-to-market and the severe distraction of focusing engineering talent on functions that do not differentiate the core business.

To highlight this, a large insurer recently described to us how they embarked on building their own proprietary claims management system. After working on it for over a decade, they finally realized the output was wholly inadequate and that they could never quite keep up with the best third-party technology. This person argued that insurers are very unlikely to replace core vendors such as CCC, Mitchell, Guidewire, or Duck Creek with internal AI builds.

We also spoke with one of the Mag-7 tech companies. Even with all of its tech talent and resources, the agentic enablement leader did not think it made sense to replace core vendors. The sunk cost was highlighted as simply too large.

Of all the experts we spoke with, only one predicted a potential full replacement of the CRM system. They were a large specialized healthcare services business and wanted more control over their IT as healthcare regulations change. They were frustrated by waiting for Salesforce to implement such updates. They also have an outsized tech spend versus typical enterprises.

We also arrived at our own epiphany regarding the “AI-build versus buy” debate. Even if an enterprise successfully builds and operates a leading internal system today, maintaining that edge requires constant, compounding innovation, a process that does not exist natively within most non-software companies. Leading SaaS incumbents do not stand still. Attempting to match their R&D velocity would create a massive distraction in the core business and innovation processes. Further, if many companies did this across the economy, it would be an incredibly inefficient use of economic resources. There are economies of scale for third parties that amortize these efforts across many customers.

We maintain conviction that strategic software vendors will not lose to internal corporate builds using AI.

AI Point Solutions

Even if incumbents hold their ground against broad platform replacements, they face the risk of losing incremental revenue as customers adopt best-of-breed AI point solutions. However, this dynamic is not new. Strategic vendors have always faced threats from point solutions and have never captured 100% of their clients’ wallet shares in their respective markets. The more critical issue is whether incumbents lose the agentic layer entirely, which would relegate them to a back-end database.

During our interviews, we asked IT leaders how they plan to allocate agentic spending. The majority consensus was clear: when a specific use case is highly related to a strategic vendor, such as a sales-related agent workflow with CRM or a finance-related agent workflow with ERP, customers plan to use the incumbent’s native agentic offering. The reasons cited included easier, faster time-to-market, out-of-the-box enterprise-grade governance, lower integration burden, existing permissions, and direct access to richer native metadata and workflow context.

Furthermore, customers told us that building highly secure, compliant agents internally using foundation models like OpenAI or Anthropic is incredibly complex. For example, a leader in agentic enablement at a Mag-7 company said that, even with its vast technical skills and resources, it did not dismiss Agentforce as a nonstarter because the hard part is not simply querying Salesforce data, but recreating Salesforce’s embedded metadata, permissions, workflows, business logic, trust, and governance layer. Custom MCP-based agents can work, but only with strong internal engineering, careful permissioning, and ongoing agent governance. Otherwise, autonomous agents may access or act on data incorrectly, creating serious compliance and security risks. In the end, this company chose to continue internally but felt this would not be a good decision for non-tech companies. Still, the expert believed Salesforce remains highly valuable as the operational system of record because of its installed base, user familiarity, customer/vendor interoperability, workflow history, and training investment. Another large payments company cited that even in the system-of-record role, there is significant value in all of the embedded workflows and processes. In this sense, the choice between using a software vendors’ AI tools and building something custom with Anthropic/OpenAI has many of the same hallmarks as the broader build-versus-buy question.

In contrast, customers prefer to build their own custom agents for cross-platform workflows or areas central to their core proprietary business logic (e.g. proprietary risk modelling for an insurer). That is, each strategic software vendor has no right to win outside its own domain. We believe a company like ServiceNow could face slightly elevated risk here due to its cross-platform orchestration focus, although experts also noted using ServiceNow’s Now Assist for native workflow execution, even if the core cross-platform orchestration layer was being built internally.

We can also see customers’ agentic spending plans in broader surveys. Here, the strategic vendors are capturing their fair share (as compared to their segment’s share of overall IT spend).

AI-Native Competitors

Another critical question is whether strategic SaaS incumbents are structurally impaired from offering the same agentic AI solutions as AI-native entrants. That is, they might hold ground in the short-term but die a slow death long-term. We analyse this through two lenses: technical debt and business model counterpositioning.

Technical Debt

There are significant technical challenges to offering robust enterprise-grade agentic products. However, many are faced by both incumbents and new entrants. Some favour the new entrant, while others favour the incumbent. The biggest technical challenges in our view are more fundamental.

First, traditional software was designed around human workflows, often assuming human clicks, judgement, or implicit understanding. When humans are removed from the workflow, the broader process may also change.

Consider the “job to be done” in a sales pipeline. The immediate task may be to prioritise prospects, but the ultimate objective is to convert prospects into sales. In a fully agentic environment, where agents are always active and can engage many prospects simultaneously, prioritisation may become less important. If prospective customers also use agents, the need for interpersonal relationships may diminish.

This remains somewhat speculative, but it illustrates how individual workflows—and entire business processes—could change. Retooling existing codebases for fundamentally different workflows with limited or no human involvement may be a significant undertaking in certain areas. However, we expect humans to remain in the loop for some time, giving incumbent vendors time to adapt.

Second, incumbents are slow. An undeniable technical challenge is the simple fact that the pace of innovation for incumbents will always be slower than smaller, more nimble start-ups. This has always been the case, but perhaps matters more in a moment of potential regime change.

In contrast, AI-native entrants face their own key technical weaknesses. New entrants lack the control over client data, business objects, workflows, historic context, and auditing rules that incumbents have built over many years. In addition, their clients have their own significant technical debt and disparate data issues that need to be solved before agents can add value. Coordinating complex retooling from outside the system of record is a big disadvantage. If the incumbent owns the workflow engine, permission model, business objects, and audit trail, it may have the safer action surface.

Overall, we see technical debt as a real risk, but not an insurmountable one.

Counterpositioning

Technical challenges are not the only risk. We think Hamilton Helmer’s framework on counterpositioning is a key consideration when assessing AI-native entrants. For those unfamiliar, counterpositioning is a business model challenge where incumbents choose not to compete with new entrants because either (i) doing so would destroy more value in their existing business than they would gain by winning against the new competitor; or (ii) they suffer from institutional bias or agency issues, even if competing would create value.

Various commentators argue this dynamic is currently playing out, suggesting large software companies are structurally paralyzed as they seek to protect their high-margin, seat-based licensing revenues. However, we simply do not see this in the evidence. Incumbent firms are not paralysed and are aggressively racing to offer agentic solutions, frequently driving significant platform shifts in the process. Salesforce, ServiceNow, and SAP, for example, are all heavily investing in agentic layers.

This does not mean incumbent high-margin seat revenue is insulated. It simply means incumbents recognize the shift and calculate that they can compete effectively with their own agentic offerings, and that doing so is better than the alternative. We agree with this assessment, particularly for strategic software vendors that control authoritative data and workflow context and possess deeply embedded customer relationships.

As noted in our previous letter, whether these efforts can offset losses in seat-based revenue depends on whether the incumbent itself provides the automation that enables these seat reductions. If so, enabling seat reductions is a significant value proposition to the customer, and we have conviction that delivering higher value will ultimately lead to higher revenue and value in the long-term, regardless of the pricing model, despite AI revenue being lower margin as they leak value to the model providers. If not, then the business will likely suffer. This is why it is so important for incumbent vendors to have a solid AI roadmap.

That said, organisation-wide vendors remain at risk because their billing is heavily indexed to general knowledge worker headcounts. For example, if Salesforce enables headcount reduction in sales teams, that is good for Salesforce longer-term but negative for all companies that earned revenue from those same people, such as security companies and general productivity tools. This is a negative we consider in our portfolio companies Microsoft 365 and Okta.

The Necessity of a Credible AI Roadmap

We demand that our portfolio software companies demonstrate a credible, aggressive AI roadmap. Without it, they carry only risk and no opportunity. If they are not adding value on top of the LLM models then they do not deserve to win.

• Salesforce (CRM): Despite some market critiques of Agentforce, the company’s platform shift fortunately started before the ChatGPT moment. It is not perfect, but it is rapidly improving and there is real revenue generation. Customers signalled to us that they like the product overall and can see more use cases being deployed over time. Many customers said that it makes sense to use Agentforce when the workflow was related to their Salesforce products.
• CCC Intelligent Solutions (CCC): Customers consistently view CCC as an innovator, recognizing that CCC can develop AI tools (withing CCC’s markets) far better than insurers can internally. The company has a strong engineering culture from the top down. Customers cite real progress in AI efficacy, particularly with its automatic estimation tools, even though there is a long way to go.
• Okta: Okta is paving the way for non-human identity control, developing new open protocols for agent authorization recently adopted by Anthropic. While it remains early days and revenue is limited, it is good progress.
• Microsoft: While Microsoft has made some missteps regarding the positioning of its proprietary models and Copilot, we are considering the AI risk to its SaaS business here, namely Microsoft 365. The company has executed well in integrating agnostic models into the Office suite. Copilot for Excel is finally good, even if powered by Claude. Remember, Microsoft’s is often the second-mover, relying on its incumbency to catch up (Cloud, Teams, Identity, AI).
• Guidewire: Guidewire is building the governed platform layer for agentic AI across core insurance workflows, data, and actions. Its advantage is system-of-record depth in its markets. It has also released newer products that are AI-native. However, overall, we see it as a bit behind what say Salesforce has done.

Agentic Revenues Have a Long Ramp

The final key insight from our channel checks is the time required to implement and scale enterprise agentic use cases.

First, many enterprises simply do not have their underlying data and systems organized in a way that allows for effective AI deployment. Preparing fragmented, legacy data architectures for agentic consumption is a monumental, ongoing task. Agentic AI touches permissions, data residency, auditability, compliance, and human approvals. In the old world, data gaps, errors etc were smoothed over with embedded human knowledge. This needs fixing in the agent world. It takes time.

Second, it is taking time for less tech-forward enterprises to build trust with AI.

Third, the technology to power true autonomous agentic use cases has not existed for very long. Customers described having only a handful of AI use cases that have exited pilot testing and entered production, but noted hundreds more sitting in the pipeline. Customers could see their agentic spend rise as these are rolled out, benefiting all of their agentic vendors. This is highlighted in the survey data shown above.

The market is failing to appreciate this multi-year ramp in enterprise workflow agentification. We have heard sell-side analysts ask software management teams on earnings calls why their AI revenues have not scaled at the same vertical trajectory as Anthropic’s. This fundamentally ignores the reality that Anthropic likely has very limited revenue stemming from true end-customer enterprise agentic workflows as they too face this slow ramp.

Updated AI Defence Framework

To survive and thrive in the AI era, our incumbent software holdings must possess specific characteristics:

• Mission-Critical with Deep Switching Costs: Typical SaaS moats rely on switching costs and embedded sales and distribution. We demand switching costs that go beyond technical integration friction, which is a weaker moat, and embed deeply into the customer’s core business processes, where a platform change causes substantial operational disruption. That is, the software must run mission-critical functions and possess strategic vendor status.
• Moats Beyond the Norm: Ideally, these companies possess additional layered moats that are structurally difficult for AI to attack.
• CCC: CCC possesses a powerful three-sided network connecting repair shops, insurers, and parts suppliers. It is also legally and reputationally beneficial for insurers to use an independent third-party interface rather than decide their own estimates internally.
• Salesforce: A massive global ecosystem of certified implementation partners and trained administrators.
• Microsoft 365: An entrenched protocol and file-format network effect.
• All of our software companies have deep trust with their clients affording them a distribution edge.
• Ownership and Control of Data: The platform must serve as the system of record with strong ownership and/or control over that data. There is even more protection in highly regulated industries or data that cannot accept even small error rates (e.g. general ledger).
• A Credible AI Roadmap: Demonstrable track record of early agentic AI development and revenue success.
• Limited Exposure to Unrelated Seat Reductions: Seat-based models that are organisation-wide, such as general productivity tools and security tools, could lose seats through no fault of their own. If this is the case, there must be other factors that can overcome this risk.
• Competitive Price to Value Equation: The vendor cannot operate under an unsustainable pricing umbrella, as this makes it easier for new entrants to attack. Salesforce, for example, faces some risk here, as enterprise buyers voice frustration over its pricing. We see this in Commerce Cloud, where customers / system integrators tell us that a combination of price and slower innovation has ceded significant ground to Shopify and Adobe.

Final Thoughts

We are not underwriting “AI does not matter.” But the “SaaS is dead” narrative is too reductive. We are underwriting a more specific view:

• The user interface moat weakens.
• Seat economics are pressured.
• Systems of record remain valuable and the more so if also a system of action.
• Workflow/action surfaces become more valuable.
• Governance, identity, audit, and permissions remain or become more important.
• AI-native entrants win where incumbents sell expensive, opaque modules around the core.
• Incumbents are at risk in cross-platform tasks and where it relates to the core business of the client (i.e. where businesses should differentiate themselves).
• Incumbents that execute well win where they control data, workflow state, permissions, and regulated execution.
• The moat does not disappear. It migrates deeper into the execution layer.

The software companies we want to own must pass a higher bar than before. We do not discount the risk. Risk has increased. Terminal values should be lower. But valuations are much lower. Present valuations factor in only the risk, without any of the opportunity that AI creates for well-placed software vendors.

Software Disruption Update 4 Nov 2026

We presented our thoughts on Software AI Disruption Risk in our last letter. Since then, we have been speaking with software engineering experts to deepen our understanding of technical areas that might create advantages or disadvantages for incumbents or new entrants. Our focus on this continued work ignored the business model challenges and instead looked only at the technical challenges. We wanted to know if there were any software engineering-related reasons that would make it difficult for incumbents to compete, akin to the on-premise vendors during the public cloud transition.

In short, while the AI transition is much more transformational to society and as a technology than was the public cloud, the technical differences between incumbent SaaS vendors and new AI-native-software entrants seem less extreme (note, we are talking about software new entrants here, not the frontier labs).

The move from on-premise software to cloud SaaS changed several things simultaneously. There was a physical infrastructure change, with the vendor taking responsibility for running the servers. The vendor also took over software maintenance and updates (i.e. software went from a product to a service). Upgrade cycles, maintenance, availability, security and compliance also became continuous responsibilities for the vendor. For many applications, there was also a shift from single-tenant to multi-tenant architecture, where customers shared infrastructure. Software designed around one customer’s installation had to work safely and efficiently across many customers. Simply hosting the old product in the cloud did not achieve this, and therefore incumbent on-premise businesses also suffered a deep technical disadvantage.

By contrast, an established SaaS vendor moving towards agentic AI need not undergo a comparable wholesale architectural conversion. It already has many of the foundations that AI-native entrants still need to build. Both incumbents and new entrants must add agentic layers, such as a centralised data/reasoning engine, an ontology layer that maps data and workflows (i.e. context), agentic reasoning logic, connections to deterministic process automation tools, MCP access, and agentic governance. These additions remain significant undertakings.

In various areas, the incumbents have a credible advantage since they own or control the data, context, and client relationships. Good context is a bottleneck. Even in a headless model, the incumbent isn’t reduced to a dumb database; it remains a critical system of record and context layer. Further, for various applications, integrating the agents with the system of record reduces the number of steps between reasoning about an action and taking it, with lower latency and token consumption as well as higher accuracy. A native agent may also have access to context and capabilities that are not exposed externally, allowing more actions to be authorised and more workflows to be automated within defined controls.

That said, incumbents still face real architectural disadvantages. Older SaaS incumbents often have monolithic software architectures (everything in one big codebase) versus more modern microservices stacks. Building, testing, and releasing changes across a monolithic core takes time, although the controlled release process also protects reliability. Incumbents must preserve existing APIs, customer customisations and automation, which complicates agent development and testing. A new entrant can design a narrower product around agents from the start and change it more quickly. Further, as mentioned in our last letter, re-tooling old human workflows for agentic execution may be harder and take longer than simply building new workflows from scratch. Finally, while incumbents have a context advantage, they still do not capture all the context in customer decision-making processes. Much of this lies in things like emails, Slack channels, and in people’s heads. Incumbents and new entrants need to capture this too.

Overall, our further work on the technical areas led to the same view presented in our prior letter. Mission-critical, enterprise software vendors with credible AI roadmaps have significant potential in the AI era.