Assessing Moats, Business Models, and Valuations in the Agentic Era
Agentic AI is reshaping the enterprise software stack, changing customer priorities and shifting investment budgets while software monetisation moves toward hybrid seat, consumption and outcome-based structures.
Agentic AI is reshaping the enterprise software stack, changing customer priorities, and shifting investment budgets at a pace faster than any previous technology transition. This shift has contributed to a broad sell-off in European software, with our coverage share prices declining on average by c.37% from their 52-week highs, as investors reassess incumbent value propositions against agile, well-funded AI challengers, particularly from the US.
Rather than causing full disruptive displacement, we expect AI to redistribute value across the software stack where incumbents are likely to retain control of base applications and mission-critical, regulated systems of record. However, in our view, incremental value will increasingly accrue to the orchestration, agentic, and contextual layers built on top of these systems. To capture this value, software monetisation models are transitioning from pure seat-based licensing to hybrid structures that incorporate consumption- and outcome-based pricing. While companies largely remain in the experimentation phase of AI, we expect the cost dynamics between LLMs to become a greater focus as companies seek to optimise AI spend contingent on achieving sufficient ROI.
Observing prior disruption cycles, we think it is likely to take a multi-year period for stocks to re-rate from trough multiples as the market digests the uptake of AI product offerings from incumbents, evidence of clear monetisation pathways and the transition to consumption-based models in determining more realistic long-term growth rates. Overall, in this note, we evaluate business models across our coverage and the defensible moats that should retain their core value, with the goal of revisiting this topic over the coming quarters as the agentic technology ecosystem evolves.
Agentic AI is reshaping the enterprise software stack, changing customer priorities, and shifting investment budgets at a pace faster than any previous technology transition. In that context, our European software coverage share prices are on average down c.37% from their 52-week highs as the market reassesses incumbent value propositions against fast-moving, increasingly enterprise-credible AI challengers, particularly in the US.
Our core view is that AI is unlikely to lead to disruptive displacement of application software, but we expect it to redistribute value across the stack. The base application and system-of-record layers are likely to remain anchored with incumbents, particularly where workflows are mission-critical, highly regulated or deeply embedded, but incremental value is likely to accrue within the agentic, orchestration and contextual layers above. We see the revenue models associated with incumbents as likely to undergo a transition as AI evolves toward hybrid monetisation, with seat-based licensing complemented by growing consumption and outcome-based elements. While companies largely remain in the experimentation phase of AI, we expect the cost dynamics between LLMs to become a greater focus as companies seek to optimise AI spend contingent on achieving sufficient ROI.
The European software debate remains focused on identifying the most resilient and adaptable incumbents, set against higher-growth, well-funded US peers that have shown faster initial deployment across models, agents, developer workflows and orchestration. While European incumbents are navigating this transition from different starting points, we note that some retain meaningful defensive value through embedded systems of record, domain expertise and workflow context. While it remains challenging to identify clear AI winners today, we differentiate based on which incumbents have sufficient defensibility, valuation support and credible AI monetisation paths to protect or re-expand value over time. As a result, we do not see any immediate catalyst for a re-rating until we see product roadmap delivery, rollout of agentic AI solutions, and adoption and monetisation sufficient to drive an upward re-rating.
Given the rapidly evolving nature of agentic AI, we remain open to feedback on these topics. We assess the debate across four key pillars: 1. Agentic AI Reshaping the Tech Stack in Enterprise Software We expect agentic AI to evolve software interaction from human-to-software toward agent-to-SaaS, elevating APIs as a larger component of the new interface. The fragmented landscape of mobile and desktop applications is giving way to a centralised, intent-driven orchestration model, where a primary AI agent routes tasks to specialised sub-agents, crossing information silos on the user’s behalf. We think this is likely to commoditise thin-wrapper apps and single-silo vendors where switching costs are low and data moats are shallow. Pure front-end vendors such as analytics and dashboard providers are similarly vulnerable. By contrast, cross-domain, deeply embedded platforms remain structurally harder to route around, as their workflow integration and data context are essential inputs to the orchestration layer. As a result, we believe power users are likely to retain incumbent UIs, albeit modernised with agentic capabilities, while the majority of interactions migrate toward third-party AI interfaces.
In the context of AI, what matters most is no longer data alone, but the contextual layer built around it: workflows, metadata, governance, auditability and years of accumulated operational knowledge. Systems of record remain critical, providing the grounding, controls and accountability that enterprise AI workflows require, and in many cases, we expect these systems to become more valuable as AI adoption scales. The most likely scenario in our view is therefore not wholesale displacement of systems of record, but value migration above them: incumbents retain the data and contextual workflow layer while third-party agents, AI-native providers or customers themselves contest the higher-value intelligence layer on top.
Agentic AI is simultaneously reshaping software economics. Traditional SaaS was built around seats and predictable costs; however, AI introduces variable economics driven by tokens, inference and workflow complexity. While token costs have fallen more than 90% over the past few years, usage is scaling faster, and the key debate remains whether consumption costs are structurally deflationary or whether leading LLM providers gain durable pricing power. Overall, business models are evolving toward hybrid monetisation: seat-based licensing complemented by growing consumption and outcome-based elements, with pricing power increasingly tied to contextual data, governance, and domain expertise rather than vendor lock-in. Historically, software vendors have had high visibility, strong pricing power and high lifetime value through the subscription model; however the shift to hybrid models combining seat- and consumption-based pricing is likely to reduce this advantage and increasingly drive dispersion between those enabling agentic workflows for mission-critical workflows versus less complex workflows. This is likely to result in potential shifts in IT budgets from buy to build.
On margins, we expect AI to represent a near-term headwind to gross margins given increased hardware costs, landing around 60-80% versus 75-85% in the SaaS period and 85-95% in the on-premise era. However, we expect AI to drive optimisation elsewhere within opex to help partially offset this pressure. Overall, as companies look to modernise their stacks, we expect AI budgets will be sourced through a combination of IT budget reallocation, labour-cost displacement and operational savings, favouring cloud/AI-native vendors and deeply embedded incumbents while compressing profit pools for legacy and commoditised providers.
Our European software coverage share prices are down c.37% from their 52-week highs, as investors question the software business model and reassess how value may be distributed across the AI stack. While current valuations already imply a degree of structural disruption risk, it is challenging to conclude whether the sector has reached a definitive trough, particularly if the market is repricing terminal value rather than near-term earnings. Historical technology cycles, such as the shift to the cloud, suggest that periods of business-model change and uncertainty around long-term value creation can take time to resolve. Re-ratings are contingent on moving beyond adoption rates and engagement toward observing clear proof points of monetisation, pricing power and margin durability, which largely remain unknown.
In this note, we highlight business models with clear, defensible moats that should retain their core value across an evolving Agentic tech stack. Within our coverage, we see SAP (Buy) as best positioned given its mission-critical enterprise workflows, underpinned by the contextual layer built over decades of cross-domain integration. Furthermore, we see NEKG (Buy) and Dassault Systemes (Neutral) as relatively well positioned given their deep domain expertise across construction and manufacturing, respectively, while Temenos (Neutral) benefits from operating in an inherently highly regulated sector. Nevertheless, we also see risk in certain vertical software companies where customers use LLMs and next gen data vendors to build new functionality, potentially limiting upsell and cross sell-opportunities. On the other hand, we see Sage (Neutral) as relatively more exposed, reflecting its horizontal software positioning and SMB-skewed customer base, although we highlight that the regulated nature of accounting software provides some defense.
Agentic AI Reshaping the Tech Stack in Enterprise Software
In our view, agentic AI is likely to progressively re-architect the user-interface and front-end layer that has traditionally been dominated by application vendors. Today’s landscape, where users navigate a fragmented set of applications across mobile and desktop platforms, will likely shift toward a more unified, intent-driven experience, in which a centralised prompt interface or primary agent serves as the orchestration layer, communicating with specialised sub-agents and crossing information silos to access relevant data on the user’s behalf.
Nonetheless, as the primary interface shifts toward AI-enabled generative UI, traditional human-to-software interaction may increasingly be replaced by agent-to-SaaS communication, with autonomous agents navigating APIs to achieve business outcomes, potentially commoditising the application layer itself. In effect, APIs become a large part of the new UI. The new application layer, in our view, will be dynamic, with traditional dashboards replaced by custom widgets and tiles that act as a shared canvas where end users can also view agent reasoning in real time, built from the ground up around tasks and workflows. The MCP underneath this front end acts as a standardised system integration layer.
“Wrapper” applications with thin value-add are likely the most susceptible to commoditisation, as agents increasingly bypass the UI entirely, a risk that is amplified for vendors operating within a single silo, where switching costs are lower and data moats shallower.
Firms with well-embedded application workflows spanning multiple domains and data silos are better positioned to retain relevance: their cross-domain data context and workflow integration remain key inputs to whichever orchestration layer prevails, making them materially harder to route around.
Importantly, we do not expect users to route everything through AI. Power users, particularly those already working in efficient, purpose-built application formats, are likely to retain incumbent UIs, albeit modernised to incorporate agentic AI workflows, while the majority migrate to third-party AI interfaces. In that sense, we expect the Ul to converge toward a centralised orchestration point rather than universal AI adoption across every interaction. We note that throughout the history of the application software industry, we have seen evolutions of the user interface, with new entrants pioneering more user-friendly interfaces. While the hold of the application vendor is likely to be challenged more than in prior evolutions, we still believe the key moat lies in application data and workflows, which are likely to withstand disruption. Meanwhile, software vendors that act as a pure front end, such as analytic and dashboard providers, are more likely to be disrupted.
The orchestration model The complexity of real-world enterprise tasks and the breadth of enterprise data underpin our view that the future of agentic AI will likely involve many specialised agents rather than a single all-encompassing frontier model. Each agent is likely to handle a specific domain or task, equipped with specialised data and tailored workflows, while communicating with a central orchestrator or with other agents as needed.
This modular architecture favours flexibility, scalability, resilience and cost optimisation. It also creates roles for small language models and open-source models in cost- or latency-sensitive parts of the stack, as well as deterministic automation or existing applications where they remain the most efficient solution. As open-source models continue to gain adoption, such as China’s DeepSeek, we expect enterprises to increasingly route tasks across different models based not only on cost, but also on
specialisation and performance requirements, rather than relying on a single centralised LLM for every use case. In our view, software incumbents, with established platforms and deep workflow integration, could be well positioned to act as the orchestration backbone given their deeply entrenched domain, cross domain and vertical expertise in enterprise business workflows.
Harness existing models by layering them with vertical and domain expertise, proprietary data, workflow logic, contextual metadata and governance frameworks without ceding their advantage to third-party agent layers.
The Microsoft Copilot-Anthropic collaboration is a useful example of the role software players could play: different LLM providers contribute model and interface strengths, while established software firms contribute domain-specific data, workflows, governance, and potentially the routing logic that determines where LLMs, SLMs, open-source models or conventional automation are best suited. Another example is SAP, which routes third-party agents through Joule, mediating access to the data and context beneath it. This represents the preferred posture for software incumbents seeking to remain in the value chain rather than be disintermediated by it (according to management, the SAP Joule + Claude collaboration reportedly reduces the error rate to 0%, vs c.60% for Joule alone on enterprise tasks). This model could reframe key software
moats: while data and governance may be back-end in nature, their value could be realised if incumbents are able to act not just as a SOR (system of record), but also as an orchestration layer that empowers and directs AI workflows. Whether that orchestration role can be priced, and how richly, remains to be proven, and is ultimately what we expect will partly determine how much of the AI value-add incumbents retain.
We expect more partnerships of this kind as customer expectations move beyond simple chat toward full agentic workflows. However, while AI usage has moved beyond pilot programmes in many enterprises, it remains experimental in terms of architecture, vendor selection, procurement and economics. A degree of switching and churn is therefore likely as customers trial different approaches and converge on optimal configurations.
Data to become more important, with the Contextual layer a key differentiator
The effectiveness of AI systems still depends heavily on data quality. In this context, while we acknowledge investor concerns that systems of record could be displaced by agents, we view this as unlikely. Systems of action, the applications that execute tasks and drive workflows, depend on systems of record for training, grounding, permissioning, and contextual execution. As AI usage expands, the importance of systems of record should therefore rise rather than fall.
The shift to AI also heightens the need for integrated data. Disorganised datasets and isolated silos limit model performance, particularly as agents increasingly operate across functions. This should sustain migration toward cloud and data-platform architectures, even if short-term churn rises as customers evaluate new options. Companies that consolidate, govern and enrich their data in AI-enabled environments should be best positioned to generate durable advantages.
We would also note that raw data is not the full moat where customers own and can extract it. In our view, the contextual layer built on top is a defensible moat in the AI era that contains the processes, workflows, metadata, permissioning, auditability, embedded business logic and cross-silo reconciliation. This layer allows AI use cases to move from broad intelligence to actionable, domain-specific business applications. It is also harder to replicate or extract because it reflects years of implementation, exceptions, controls and institutional knowledge.
We note this is gaining importance as the debate increasingly shifts from model capability to context and accuracy. In many enterprise workflows, the challenge is not generating an answer, but generating the right answer with sufficient accuracy, auditability and accountability. Recent examples, such as Ford rehiring hundreds of experienced engineers after AI-led quality systems failed to fully replicate veteran engineering judgement, highlight the continued importance of domain expertise and institutional knowledge alongside, and above, AI tools.
Against this backdrop, we expect incumbents to focus increasingly on building knowledge graphs: structured, interconnected representations of an enterprise’s data, metadata, relationships and business logic that AI agents can query and reason over. These encode institutional knowledge into a form agents can rely on, turning embedded domain expertise into a durable AI-era moat.
We note that vertical software generally offers specialised, industry-specific solutions and generally benefits from high switching costs, compounding domain expertise, proprietary data, regulatory compliance and embedded workflows. Our base case is that AI is more likely to enhance this type of core functionality than replace it, given the complexity and specialisation involved. At the same time, because the biggest historical constraint for vertical software has often been access to technology talent, AI productivity gains could accelerate adoption in verticals that have historically lagged, representing a potential tailwind for the segment.
That said, we note that we are seeing some shifts in the build-versus-buy calculus within the segment. There have been increasing examples of customers choosing to partner with LLM providers and systems integrators to develop their own AI applications on top of the base offerings they expect to maintain from incumbent software vendors. The Stellantis-Mistral-Accenture collaboration, as well as Palantir’s work with multiple automotive customers, Prometheus’ promise of a “digital engineer”, and Schindler’s use of AI for housing digital twins, are early but notable examples of large enterprises choosing to build rather than buy as they seek to capture AI value faster and more cost-effectively, and we expect this trend to continue.
IT services players are also positioning themselves as natural partners in this shift. Capgemini, for example, is offering vendor-agnostic AI solutions alongside its Agentic Control Plane, a platform that manages agent permissions across providers (such as Google, OpenAI, etc.), optimises token costs and redesigns processes for hybrid human-AI workforces. With decades of domain expertise, these players could occupy a potential midpoint between general-purpose LLMs and deeply embedded but slower-moving incumbents.
We do not see this as a direct threat to incumbents’ core customer relationships. Their workflows are deeply entrenched and the system-of-record position remains durable. However, as value migrates to the agentic workflow layer above the base application, incumbents without credible AI roadmaps may find their right to win in that layer increasingly contested. While build-your-own efforts have not yet materially displaced core vertical software, the growing availability of faster-moving and potentially cheaper alternatives could begin to pressure the segment’s pricing power and stickiness over time.
Horizontal software appears relatively less insulated from commoditisation, as AI agents can perform broad tasks more efficiently and cost-effectively, increasing competition and easing insourcing. The key distinction we highlight is single-lane versus mission-critical software: certain horizontal tools, notably Office of the CFO software, are better insulated given the low appetite for disruption where accuracy and accountability are paramount. More broadly, horizontal players with significant data moats retain an advantage, since agents replacing organisational tools still require deep domain knowledge and business context.
In general, we think this is difficult to do for complex processes and, where possible, would not be time efficient. Context is not simply the sum of smaller datasets but is the product of scale, cross-functional reconciliation, governance, and operational usage across business units. A finance unit and a supply-chain unit may each hold a different definition of the same business concept. The incumbent’s value lies in reconciling and operationalising that concept across the enterprise. That said, this advantage is not uniform: where vendor platforms remain poorly integrated, the contextual moat is weaker.
Yes, they can extract the data, but not the context and metadata, the intellectual property of which is owned by the incumbent software layer and which we view as the key moat. We expect regulated industries (such as finance and accounting, or verticals such as construction or healthcare) to remain particularly sticky, as the cost and risk of stripping out the contextual layer are materially higher, whereas CRM-type workflows and simpler tasks could be more exposed. Critically, extraction also removes the regulated, governed infrastructure around the data: auditability, lineage and compliance controls that are non-negotiable in enterprise contexts where even a 1% error rate can be prohibitive to adoption. Moving data outside secure, governed environments exposes it to ungoverned external sources, reintroducing risks that the incumbent platform was specifically designed to eliminate. SAP’s illustration of Claude moving from a c.60% error rate without context to c.0% when paired with Joule’s context is a good example of the build-vs-buy calculus in practice, and the offering’s ability to make all AI interactions 100% auditable reinforces the point. For most customers, the math could still favour staying with the incumbent for core and critical tasks, not because the data cannot leave, but because the trust architecture around it cannot easily be replicated.
We do not see this as proven yet, but it is where we see the most credible threat: incumbents retain the SoR (system of record) while ceding the incremental value layered on top of it. Players like Palantir, through their Ontology offering and forward-deployed engineer (FDE) model, represent the most credible vector for this erosion. If they begin building software directly around customer problems, as they appear to be doing, they can capture the value-add layer by going deep, industry by industry, at structurally higher margins. The other scenario we see is customers leveraging advanced coding tools and their own domain and vertical expertise to build solutions on top of the data within the existing system of record, a scenario we could see emerging across many vertical industries. These solutions are unlikely to displace the SoR, but they could win the right to operate on top of it unless software incumbents are able to fulfill the demand for AI offerings themselves. The broader build-versus-buy shift reinforces this dynamic: as enterprises increasingly partner with LLM providers and systems integrators to develop their own AI capabilities, the value-add layer above the SoR becomes more contestable, not just by dedicated challengers like Palantir, but also by customers themselves.
In that context, it is our view that the key question becomes whether software incumbents can earn the right to that value-add layer. Here, the incumbent’s defense is not the SoR itself but the quality of its moat: trust, data integrity and the reconciliation advantages that directly improve the efficiency and cost profile of its AI offering. These attributes could remain a barrier, placing a ceiling on how much incremental value a challenger can extract unless it can prove a superior position on them. This is where we see pricing and efficiency becoming decisive, and it shifts the debate to where we think it ultimately belongs: cost.
While data and context are the foundation, they are not the only durable advantages. We see several reinforcing moats that extend incumbents’ defensibility as the agentic era evolves:
Embedded adoption. Incumbents can integrate agentic capabilities directly into existing, widely used platforms, creating a natural adoption path within their installed base.
Domain and workflow expertise. Specialised knowledge of industry- and company-specific processes is essential for interpreting context, relevance and metadata, and therefore for building agents that deliver real business outcomes. This applies to both vertical and horizontal software.
Established customer base and lower CAC. Broad customer bases and entrenched workflows enable cross-sell and upsell of new agentic features at low acquisition cost. While AI may reduce opex and CAC for newcomers, incumbents benefit from the same dynamic.
Existing distribution and sales infrastructure. Decades of market presence have built sales forces and reseller networks, which are efficient channels for new AI-powered offerings and increasingly important if AI makes the underlying product less differentiated.
Most importantly, auditability, governance, and trust. Autonomous agents introduce meaningful security and judgement risks, and the lifecycle controls of many LLM ecosystems remain relatively opaque, with limited regulatory oversight. In enterprise contexts, where, unlike consumer use, 99% accuracy across large workflows and cash flows is not enough, we expect governance and security frameworks to become increasingly defensive moats. SAP, for instance, positions all of its AI offerings as 100% auditable, and this could potentially become a monetisation tool in its own right.
Economics of Agentic AI
3.1. Tokens are fundamentally changing the current SaaS cost structure Traditional SaaS economics were largely built around seats: revenue scaled with headcount and the marginal cost of supporting an incremental user was relatively predictable. AI introduces a different unit of work, whereby tokens represent the input, cached and output content processed by models across text, images, audio, video, tool calls and agentic workflows.
For enterprises, AI usage is billed, governed and optimised at the token or compute level, which translates infrastructure usage directly into software economics. This makes AI cost behaviour more volatile and non-linear than traditional per-seat software. A useful framework is to distinguish between a simple task and a complex task. A simple task is one where the model performs a narrow, mostly single action with limited context and short output, e.g., summarising an email, classifying a support ticket, extracting fields from a short document, or drafting a two-sentence reply. On the other hand, a complex task is one where the model must reason across multiple inputs, retrieve outside information, preserve state across several turns or generate long structured outputs, e.g., debugging code, performing multi-step financial analysis or running an agentic enterprise workflow.
Token consumption expands not only due to more complex inputs, but also because the model adds hidden token layers: retrieved passages, intermediate reasoning or planning steps, validation loops and repeated calls to different models, especially as output tokens typically carry a higher cost than input tokens for frontier models. In practice, token consumption is driven by six main variables: input length, output length, number of turns, retrieved context volume, number of model/tool calls per workflow and model behaviour itself, especially reasoning-oriented models that may use more compute-intensive generation patterns. As a result, costs can be highly variable, and so we believe the relevant lens to measure AI usage is cost per completed workflow or business outcome rather than simply price per million tokens.
This distinction matters because we expect token costs to create a dynamic where, as the cost per token declines, usage expands fast enough that aggregate AI spend can still rise materially. This is a demand-elasticity story in which cheaper inference expands the addressable use-case set and increases total compute consumption. The open question is whether this remains a structurally deflationary market that eventually outweighs the compute consumption or whether inference pricing could re-inflate over time if a smaller number of leading LLM providers gain durable pricing power, particularly in premium reasoning models or capacity-constrained periods.
Token costs continue to decline, but usage scales faster. Software vendors benefit from expanding AI-driven functionality and can absorb or pass through modest cost increases, preserving gross margins.
A small number of leading LLM providers gain durable pricing power, particularly in premium reasoning models or during capacity-constrained periods, and hyperscalers face pressure to demonstrate ROI on massive capex programmes. Token costs stabilise or re-inflate, leaving software vendors exposed to rising usage volumes against a less favourable supplier pricing backdrop. Margins compress unless vendors can reprice their own products accordingly.
Intensifying competition among frontier labs, including from Chinese entrants, materially lowers cost structures, while a broader industry shift away from token-heavy architectures drives downward pressure on inference pricing. However, visibility remains limited: early-stage pricing has been heavily subsidised, making the true cost floor difficult to assess. If realised, this scenario is the most margin-accretive for software vendors while eroding differentiation for frontier models.
In all scenarios, cost control becomes a key feature of AI offerings, not just an internal efficiency exercise, as enterprises are already showing signs of pushback as usage moves from experimentation to production. In our view, this further supports the strategic importance of orchestration: vendors that can reduce the cost per completed workflow should be better positioned to monetise AI sustainably.
We believe this uncertainty partly explains the recent trend away from locked-in vendor models toward multi-model architectures. Microsoft, which initially focused on an exclusive partnership with OpenAI, has since broadened its Azure AI Model Catalog to include models from other providers, including Meta’s Llama, Mistral, and DeepSeek, likely driven in part by enterprise demand for choice and the economics of usage-based pricing. As enterprises become more cost-conscious, we increasingly see workloads routed to cheaper open-source alternatives where performance is sufficiently comparable, reserving premium frontier models for the most complex reasoning tasks. Vendors with strong domain context can further improve AI economics by narrowing prompts, structuring data, reducing unnecessary token consumption and directing tasks toward the most efficient tool, whether a frontier model, SLM, classical ML, search engine or deterministic automation. Over time, this ability to reduce the cost per completed workflow may become an important source of margin protection and differentiation. We view this capability as a key source of cost advantage and a potential tailwind for incumbent software vendors, but note that more proof points are needed. Vendors like Microsoft and SAP have demonstrated efficiency gains in controlled environments, but the industry has not (yet) produced a critical mass of evidence showing these advantages translating into repeatable commercial wins at scale. Palantir stands as a partial exception, with a cost model tied directly to measurable savings and clear KPIs on efficiency gains, but it remains more an outlier than the norm. The gap between proof-of-concept and production-grade cost optimisation is meaningful, and we expect investors to demand greater visibility into real-world outcomes before underwriting revenue upside tied to orchestration.
Given these token costs and as AI penetrates enterprise workflows, we can see pricing models evolving from purely seat-based toward consumption-based and, in some cases, outcome-based structures. In fact, the seat-based pricing model is becoming associated with base usage rights that are bundled or packaged into seats, with any further consumption charged on top. Historically, software companies have benefited from the strong correlation between company and economic growth, increased headcount, higher demand for software seats and revenue growth. However, with AI agents, we expect this correlation to weaken, impacting the predictability of business models and profitability going forward. In that context, we expect software firms to move toward a hybrid pricing model: a base of seat-based licensing, which would likely see the strongest deflationary impact in terms of pricing, with its share of total software spend likely to plateau or compress as consumption-based elements grow, complemented by a growing consumption/usage-based share, potentially in a land-and-expand approach as token usage increases.
Software players have already begun experimenting with this kind of hybrid architecture as they scale their AI offerings. Microsoft remains mainly subscription-based with its Copilot suite, but has progressively introduced a pay-as-you-go credit system for agents and extended AI capabilities. SAP has adopted a similar tiered approach with its Business AI Units, embedding its Joule AI assistant within the base seat licence while gating access to advanced AI capabilities behind incremental AI unit consumption. Salesforce has moved toward a hybrid structure through its Flex Credits model, combining per-seat licensing for ongoing platform access with usage-based pricing for AI and data products.
We expect monetisation will increasingly focus on access to data and contextual layers, as well as security and policy frameworks, which are key differentiators that LLM providers may struggle to replicate. This could manifest either directly, through access to domain-specific data, workflows and governance frameworks, or indirectly through partnerships with LLM providers. Pricing power could therefore increasingly rely not merely on vendor lock-in, but on software firms demonstrating clear product differentiation, often by leveraging their deep domain expertise, business context and trusted data environments. While this may reduce pricing power at the application layer itself, the impact could be partly offset by a larger TAM unlocked by AI, especially as AI blurs the boundary between labour and software budgets (Exhibit 9), at least for companies able to expand their value proposition beyond the core application.
This creates a near-term dynamic where enterprise customers benefit from attractive pricing (subscriptions remain but are cheaper, with AI add-ons bundled in), but face a medium-term risk of rising consumption costs as providers seek to recoup infrastructure investment and usage scales.
In that context, we are also beginning to see a more pragmatic industry discussion emerge around AI economics. Several software vendors have acknowledged that inference and token costs remain meaningful, while some of the more aggressive early narratives around AI-driven productivity gains have been moderated. As a result, the debate is increasingly shifting from AI adoption alone toward the sustainability of AI economics, the value of domain expertise and whether productivity gains can be monetised at attractive margins.
We expect gross margins for enterprise software companies to come under pressure in the near-term, given higher inference costs and the prioritisation of customer adoption over immediate monetisation. We note that enterprise software gross margins have generally stepped down across each major architectural shift (Exhibit 10). In the on-premise model, gross margins were often in the 90% range, supported by large upfront licence fees, annual maintenance streams and relatively low incremental delivery costs once the software was built. In the cloud/SaaS period, the industry normalized toward roughly 75% to 85% gross margins, reflecting recurring hosting and support costs. In the AI embedded era, we see an incremental headwind from usage-based COGS associated with AI models, which may result in gross margins in the 60% to 80% range. Over the long-term, the durability of gross margins will largely depend on companies’ ability to defend them through pricing power, model orchestration and domain-specific workflow efficiency.
Having said that, while gross margins may face near-term pressure, we see meaningful efficiency gains across R&D, customer support, and go-to-market functions as AI reduces operating costs across enterprise workflows. As a result, we expect some gross margin compression to be offset by opex deflation elsewhere in the P&L. For example, SAP is aiming to achieve a run rate of EUR2bn in cost efficiencies by 2028 driven primarily by internal AI deployment.
First, IT budgets tend to evolve through structural reallocation rather than unlimited expansion. In the on-premise era, spend was concentrated in upfront licences, hardware, maintenance, customisation and integration. During the cloud transition, budgets shifted toward opex-based subscriptions. With AI, the mix should evolve again: maintenance spending may compress as automation reduces run-the-business costs, while growth spending expands toward AI capabilities, data consolidation and workflow transformation.
Second, AI may increasingly blur the boundary between labour and software budgets. While some of the more aggressive early narratives around AI-driven workforce reductions have retraced, such as at Klarna and Ford, the broader direction of travel still appears consistent with organisations seeking to achieve more with fewer hours worked. In this context, we believe the relevant question is not necessarily whether AI reduces headcount, but whether it enables a given workforce to become more productive. If so, a portion of the resulting efficiency gains could be reallocated toward software, AI tools and compute. We note that Goldman Sachs’ macro team estimates AI could save c.25% of hours worked over the next decade, creating a potentially significant pool of productivity gains that AI capabilities could capture.
Third, broader operational savings could fund AI where use cases improve resolution rates, reduce cycle times, prevent errors or automate repetitive workflows. This is particularly relevant where software vendors can link pricing to measurable outcomes.
Overall, budgets may remain stable, but the internal reallocation could create a divergence in outcomes across the software landscape:
Incumbents with domain-specific workflows that embed AI deeply into mission-critical applications, making their products the natural vehicle for AI-driven productivity gains.
Lower-value IT services firms, particularly those focused on routine implementation, testing and support tasks that are most susceptible to automation.
Valuation
Our European software coverage share prices are down an average of c.37% from their 52-week highs as the market reassesses incumbent value propositions against fast-moving, increasingly enterprise-credible AI challengers. We note that European software stock valuations are now depressed and approaching historical trough multiples. Our reverse DCF analysis suggests stocks are pricing in mid term growth rates close to nominal GDP growth, with the market questioning both mid-term growth assumptions and terminal value.
We have analysed prior periods of share price performance, valuation multiple evolution, top- and bottom-line growth rates and estimate revisions to help determine the potential path forward for software performance. We believe the cloud transition from 2015-2020 provides useful context to assess a period of disruption, both from a technology standpoint and in terms of business-model transformation and reassessment of terminal value. We observe the following:
During the cloud transition, forward P/E multiples for horizontal application software vendors contracted to the low- to mid-teens, while realised organic revenue growth troughed at low- to mid-single levels. We note that today these same companies are trading close to similar valuation multiples and the market is discounting medium-term revenue growth rates close to those observed at the time. Following the de-rating and estimate reset to incorporate the cloud transition, we observed an average 4-year period before stocks re-rated and the market regained confidence that software vendors could re-architect their product sets, transition their business models and, more importantly, demonstrate monetisation from the new technology cycle.
We also note that we are approaching the one-year mark since the current de-rating began, with similar concerns around disruption and growth. Despite fears of a growth slowdown, consensus estimates have not adjusted downward, as we remain in a phase where incumbent software companies are both launching AI roadmaps and rolling out agentic AI products. However, proof points around adoption, revenue monetisation and growth accretion remain limited, with some of the earliest expectations for meaningful evidence not expected until 2027. As a result, while valuations may find support at current levels, we believe it may still take some time before we see a positive re-rating in multiples. Going forward, we expect the debate to continue focusing on whether the downside case is already reflected in share prices, but a more durable re-rating will likely require several quarters of revenue-based evidence that incumbents are benefiting from the AI cycle rather than simply absorbing its costs.
When we compare the current business model transition with the cloud era, we note that the industry previously moved from a cyclical on premise license model to a more recurring and visible revenue model with slightly lower cloud gross margins. This resulted in a substantial reset to revenue growth and EBIT margins in years 1 and 2, followed by higher lifetime revenue value from year 3 onwards. More importantly, we saw an expansion in addressable seat counts as adoption broadened beyond the back office into the front-office and supply-chain functions. In the agentic AI era, we expect the revenue impact to be additive as the revenue model shifts from per seat to increasingly consumption-based structures, albeit with the risk of lower seat counts, which we believe could be offset by higher consumption revenues. However, we see similar pressure on gross margins as we observed during the cloud transition, given that compute and inference costs must be incorporated alongside hosting costs. As a result, while revenues may reaccelerate as AI products gain traction, the initial impact on gross margin and gross profit is likely to be a headwind until we see an optimisation in seat economics. Meanwhile, we also believe the perceived loss of pricing power will remain a headwind to investor sentiment until incumbent software companies can demonstrate otherwise.
4.2. Proof points to watch for that would change the current narrative We highlight a number of proof points that we believe investors should monitor closely.
On the positive side, we expect a key indicator will be whether incumbent software vendors can drive adoption and successfully monetize their products driving accelerating top line growth. While adoption of enterprise AI assistants has been encouraging, the key test will likely be whether this translates into sustainable revenue growth, pricing power and revenue acceleration, rather than simply higher usage. Evidence that software vendors can improve customer economics while generating attractive returns themselves would also strengthen the case that AI is additive to the software model rather than merely a new cost centre.
We also view the resilience of incumbent moats as an important area to monitor. Proprietary data, embedded workflows and deep domain expertise remain key advantages for software vendors today and if they can use this advantage to drive better consumption economics and drive new “killer apps” such as autonomous workflows would also drive a shift in the narrative. Continued customer retention, successful AI product launches within existing platforms, and productive partnerships with foundation model providers would support the view that incumbents can remain important participants in the AI value chain.
Finally, reliability may prove a more important differentiator than many investors currently appreciate. As enterprises continue to value deterministic, auditable and compliant outcomes, the embedded business logic within software applications could remain difficult for general-purpose models to replicate, reinforcing the role of the application layer even as model capabilities improve.
On the negative side, several developments would increase our concern around the bear case. Most notably, AI monetization could prove more challenging than expected, with enterprises reluctant to absorb meaningful seat-price increases while vendors face rising R&D, inference and infrastructure costs. In this scenario, software companies may struggle to convert strong adoption into incremental profit pools.
We would also watch closely for signs of moat erosion. This could include evidence that increasingly capable third-party AI tools can replicate incumbent functionality, that proprietary data advantages are becoming less relevant, or that customers are able to recreate workflows outside traditional software environments at lower cost. More structurally, a faster-than-expected shift towards build-over-buy could alter the economics of the sector. If agentic AI materially reduces the cost and complexity of building and maintaining software internally, enterprises may become more willing to develop solutions themselves rather than rely on third-party vendors. Similarly, if foundation model providers or new agentic AI native application software entrants increasingly capture the customer relationship and value creation, software companies could face a greater risk of disintermediation. Finally, a material improvement in LLM reliability, accuracy and efficiency (token usage) could reduce the importance of the controls, workflows and embedded expertise that have historically underpinned software incumbents’ competitive positions.
In our view, SAP maintains one of the more defensible positionings in our coverage, supported by its deep data moat, with systems touching over 74% of global GDP, as well as its focus on mission-critical enterprise workflows that carry structurally high switching costs. We believe SAP’s advantage is further underpinned by the contextual layer built over decades of cross-domain integration across ERP, supply chain, finance and procurement, including processes, metadata, business logic and governance frameworks.
SAP shares currently trade near historical lows on EV/recurring revenue and sit at a depressed percentile of the company’s 15-year EV/FCF distribution at c.13x for CY27. Furthermore, our reverse DCF implies the stock is discounting c.4% revenue growth over the next decade versus a GSe/consensus Visible Alpha FY27-30 CAGR of 14%/11%, suggesting the current valuation assumes meaningful disruption from AI.
Looking ahead, we note that SAP outlined its AI roadmap at its SAPPHIRE user conference in May, centred on its Business AI platform, which is designed to enable autonomous workflows and AI “suites”. Its Joule agent acts as the engagement and orchestration layer, leveraging Business Data Cloud alongside extensive partnerships with new AI entrants as well as at the data layer with next generation data platforms and hyperscalers. We note that these products and capabilities have only recently begun shipping in 2H 2026, and based on our discussions with systems integrators, meaningful customer traction is not expected until later this year, with monetisation proof points more likely to emerge in 2027.
In that context, we expect investor focus to remain on customer feedback from these product launches, as well as adoption, monetisation and reacceleration on the S4/HANA migration cycle.
We view a large part of Nemetschek’s business as relatively less exposed to AI-related disruption, given its deep vertical positioning in AEC, industry-specific data, likely to be enhanced by the HCSS acquisition, and strong business context and workflow integration. Given the highly regulated nature of the construction industry, Nemetschek’s accumulated domain knowledge around compliance, standards and governance also provides a layer of defensibility. We also note that construction remains among the least digitised industries globally and, when combined with persistent labour shortages and broader workforce challenges, creates structural opportunities. That said, we see some risk from enterprises choosing to build bespoke AI applications on top of Nemetschek’s core platform rather than purchasing incremental modules, which could limit the company’s ability to monetise the agentic workflow layer. We also note that its Media segment (albeit relatively small at c.10% of NEKG’s FY25 revenue) has been weighed down over the past few quarters by cautious consumer spending, which we believe may be partly driven by small businesses and freelancers using AI to develop animations and videos.
Shares are down c.40% YTD, a sharper decline than the rest of our EU coverage given the company’s relatively premium valuation, with the stock now trading between its mean and trough valuation levels. Our reverse DCF indicates that the market is discounting c.5% growth over the next decade vs the company’s medium-term expectation of closer to mid-teens growth, reflecting risks to terminal value.
In this context, we expect investors to remain focused on Nemetschek’s AI product roadmap. The company is developing its Nemetschek AI Assistant as an AI agent across key brands including Archicad, Vectorworks, and Allplan, and has recently launched Bluebeam Max, integrating Firmus AI and Anthropic’s Claude models. We await further evidence that these initiatives are translating into meaningful adoption, customer value, and incremental revenue growth. The main near-term risk remains the Media segment’s growth outlook, which could weigh on the stock given AI-related disruption risk, even though it remains a relatively small part of the group.
We view Dassault as relatively insulated from AI disruption given its deep domain expertise across manufacturing verticals and its data moat, which allows the company to leverage its proprietary 3DXPERIENCE platform, augmented by strategic partnerships such as Nvidia powering its Virtual Twin technologies. Having said that, we expect frontier AI labs to continue improving their functionalities, which will likely prompt more of Dassault’s industrial customers to experiment with external AI tools. We note prior commentary from Schindler regarding changes to certain contract scopes, albeit in areas that are not a strategic focus for Dassault. While Dassault’s workflows are deeply entrenched and its system-of-record position remains durable, we see a risk of capped upside as value migrates to the agentic workflow layer above the base application, where its right to win is increasingly contested.
Dassault’s shares trade near 15-year trough multiples (c.13x P/E and c.14x EV/FCF CY27E), and while our reverse DCF indicates that the market is ascribing only c.4% revenue growth (GSe: 2026-30E revenue growth of c.6%) with no margin expansion, we believe the current valuation reflects concerns around the path to medium-term revenue reacceleration.
Looking ahead, while Dassault has outlined its AI roadmap and identified a €1 billion incremental revenue opportunity, we await further clarity on strategy, product rollout plans and the monetisation pathway at the Capital Markets Day on November 17, 2026. Overall, we expect investor focus to remain on signs of top-line reacceleration, while also demonstrating consistent execution and evidence of uptake of its AI solutions, as well as addressing concerns around newer solutions built internally using LLMs that could displace Dassault’s point solutions for design, collaboration, simulation, and manufacturing.
In our view, Sage appears relatively more exposed to AI-related disruption within our coverage, reflecting its horizontal software positioning and SMB-skewed customer base, where switching costs are typically lower, especially at the low end of the market. We note that this may mean a less differentiated data and contextual moat relative to several peers. Having said that, Sage’s focus on Office of the CFO workflows provides a degree of defensibility given the mission-critical and highly governed nature of financial accounting processes, especially at the mid- and high-end of the small-business software market. We also note that while switching costs are lower for SMBs, many customers lack the resources and expertise required to build and maintain bespoke AI solutions, which should continue to support demand for integrated software platforms.
Shares are down c.20% YTD alongside the broader software sector, despite growth that has been in line or better than Street expectations, which the company has attributed to continued execution and some pricing upside driven by AI offerings. Our reverse DCF indicates the market is discounting c.5% growth over the next decade versus GSe’s FY27-31 CAGR of c.9%, suggesting investors remain cautious on the durability of long-term growth and pricing power given risks from AI disruption.
Looking ahead, Sage has taken a progressive approach to AI across its product offering and has launched products such as Sage Copilot, its generative AI assistant, which is being rolled out across major product lines including Sage Intacct, Sage X3, Sage Accounting, and Sage 50.
In that context, we expect investor focus to remain on the evolving competitive landscape, including new entrants such as Rillet, and Sage’s ability to sustain pricing power. Proof points around continued revenue acceleration, cross-sell and upsell execution, and AI-driven revenue contribution are likely to be key.
We view Temenos as somewhat insulated from AI disruption given that it operates in a sector characterised by high regulation and elevated customer risk aversion, driven primarily by the stringent requirements of mission-critical banking software. We believe Temenos’ industry specific data and understanding of the regulatory operating environment will likely protect Temenos from displacement by pure-play AI vendors in the near-term. In terms of pricing strategy, management views the company’s volume-based pricing model as more insulated from AI-driven workflow automation than traditional seat-based models, which may come under pressure.
Temenos has de-rated c.15% YTD, although the de-rating has been less pronounced than across broader EU Software, with the stock trading at CY27 c.17x P/E and c.18x EV/FCF. Our reverse DCF indicates that the market is discounting c.10% growth over the next decade, broadly in line with the company’s mid-term ARR growth expectations.
Looking ahead, we believe the deployment of AI within the global banking sector remains in the nascent stages, despite 75% of banks actively exploring generative AI and only 11% having fully implemented it in production. While Temenos has not yet outlined a comprehensive AI product roadmap and revenue opportunity, we await further clarity over the near to mid-term. In our view, the key debate is similar to that in other vertical software areas: whether corporates can build their own agentic native offerings using LLMs and thereby limit future revenue growth opportunities for incumbent software players. Furthermore, we expect investor focus to remain on Temenos’ execution of large deals amidst an uncertain macro backdrop and its ability to expand successfully in the US market as part of its overall strategic plan.
We view TeamViewer as relatively more exposed to AI-driven competitive pressures than many software peers, reflecting lower switching costs and a customer base skewed toward SMBs. While we do not expect AI to directly disrupt the company’s core remote connectivity offering, we believe it increases the importance of continued product innovation and execution to maintain differentiation. As a result, investors are likely to look for evidence that AI investments are translating into improved growth, retention, and competitive positioning, particularly against a backdrop of more modest growth and recent execution challenges.
TeamViewer shares are down c.40% in the last year. Our reverse DCF indicates that the market is discounting negative growth over the next decade, below the company’s medium-term growth expectations. We believe investor focus will remain on growth reacceleration, execution, and evidence that AI-related investments are contributing to commercial outcomes.
In our view, while Sinch has moats such as carrier relationships, communications infrastructure and compliance capabilities that are relevant to supporting enterprise communications demand, we believe competitive intensity may limit its ability to capture a disproportionate share of value creation as AI adoption scales. Management has highlighted AI-related initiatives including Voice Relay, Agentic Conversations and AI-enabled fraud detection, but we have yet to see clear evidence that these capabilities are translating into meaningful incremental growth, and therefore do not currently ascribe significant value to AI-related upside.
Sinch shares are up YTD, likely driven by stabilisation in organic gross profit growth over the past few quarters, along with a perception that the company is benefiting from AI-related tailwinds. Our reverse DCF indicates that the market is discounting low teens growth over the next decade, ahead of the company’s medium-term growth expectations. We believe investor focus will remain on evidence of sustainable growth acceleration, margin progression, and tangible commercial benefits from AI-related initiatives.
Price Target, Risks and Methodology
We are Buy rated on SAP. Our 12m PT of €230 and our ADR PT of US$265 are based on c.24x 2Q27-1Q28E P/E (including SBC). Key risks to our view and price target are as follows: (1) macro risks given broad-based exposure across geographies; (2) cloud and subscription risks, such as customer churn during the transition phase, would pose downside risks to our estimates; (3) difficulty in gaining traction in S/4 HANA/cloud and not being able to draw benefits from cross-selling synergies in its installed base could cause headwinds to overall cloud growth and margins; (4) increase in opex spending would result in stronger headwinds to margins, and (5) further management changes.
We are Buy rated on Nemetschek. Our 12-month price target of €90 is based on c.21x 2Q27-1Q28E EV/FCF EPS. Key downside risks to our view and price target are: (1) worse than expected macroeconomic and end-market recovery; (2) execution on the shift to a subscriptions-based model; (3) competition; (4) brand integration; (5) M&A; and (6) FX.
We are Neutral rated on Sage. Our 12-month price target is 1,050p, based on c.17x 3QFY27-2QFY28E P/E. Key risks to our view and price target: (1) positive/negative data points around small and medium-sized business momentum; (2) evidence around lower/higher new customer acquisition and migration; (3) more/less resilience in the renewal rate; and (4) increased strategic appeal.
We are Sell rated on Sinch. Our 12-month price target of SEK20 is based on c.11x 2Q27-1Q28 EV/FCF. Key upside risks to our thesis and price target include: (1) Macro environment improvement with better-than-expected structural growth; (2) Lower cost and pricing pressure; (3) Competition, and (4) Better-than-expected integration of acquisitions.
Appendix
Understanding AI’s implications for software requires understanding the three distinct, yet interconnected, layers of the tech stack in the context of AI. Each layer is led by different companies, at varying stages of maturity, and exhibits distinct growth profiles, with parallels to the typical evolution of previous technology cycles.
The Infrastructure Layer: forms the base of all AI operations, encompassing the physical and virtual resources needed for AI computation. This includes hardware such as Graphics Processing Units (GPUs), specialized servers, and the computational power provided by hyperscale cloud platforms like Amazon Web Services (AWS) and Microsoft Azure. The focus on this layer has been coupled with an intense phase of capital expenditure among tech companies, with hyperscalers now on pace for $1.4 trillion in capex over the 2025-2027 timeframe. This concentration on infrastructure as the anchor of AI has supported a strong re-rating among semiconductor stocks, as investors and tech leaders appear heavily focused on establishing the necessary computational and data foundations first.
The Data Layer sits between infrastructure and model development, underpinning the platform layer. Data management companies like Snowflake and Databricks specialize in providing secure, scalable storage, processing, and analytics capabilities for structured, unstructured, and semi-structured data, all essential for AI. The quality and accessibility of this data are key to transforming AI use cases into tangible workflow productivity gains, which helps explain the recent valuation rallies for these data-centric companies.
The Platform Layer is where the core intelligence of AI systems is designed, developed, and refined, and is often what is colloquially referred to as “AI.” Platform companies like OpenAI and Anthropic are responsible for building, training, and deploying AI models, including large language models (LLMs) and smaller, more compact alternatives. Within this layer, a key distinction exists between proprietary models (such as Anthropic’s Claude), which offer high-performance, out-of-the-box capabilities but limited customizability, and open-source models, which can be deployed locally with superior data security and high customizability but may present reliability and regulatory concerns. Proprietary models currently account for approximately 90% of use cases, though open-source alternatives have increasingly been cited by corporates as cost-effective approaches, particularly for insourcing projects. In line with historical technology shifts like the Internet, we expect a gradual shift of capital and product development focus from the Infrastructure Layer to the Platform Layer as infrastructure matures.
The Application Layer is the final and most visible layer, and is where AI capabilities can be embedded directly into end-user software applications, products, and services. This is where the bulk of our European software coverage sits, and consequently, where the core of the AI debate resides for European software investors. It is within this layer that computing power and models are meant to be translated into real-world capabilities, economic benefits, and productivity gains.