What kind of software will be replaced by AI?

Since 2026, the software sector’s pullback differs from previous rounds driven by “demand slowdown/rate hikes”: the market is more like debating terminal value—whether these companies can maintain their profit pools in ten years, and whether their moats will be reshaped by agentic AI.

According to Wind Trading Desk, Goldman Sachs Global Investment Research analyst Gabriela Borges straightforwardly wrote in a report on the 16th: “The market is questioning the moats and business models of software companies.” She dissected seven common bearish arguments from investors, assigned risk scores from 1 to 5, and distinguished whether they impact narrow application software or spill over into broader infrastructure/security stacks and cloud vendor capital expenditures related to ROI.

Interestingly, Goldman Sachs does not consider “system-level software being completely replaced by AI” as a primary risk (score 1). More acute concerns are twofold: first, the value shifting from the system of record (SoR) layer toward “agentic operating/ orchestration layers” (score 4); second, the speed of technological iteration making the ultimate valuation difficult to price (score 5)—valuation floors are hard to establish.

In this environment of uncertainty, the report provides clear signals to watch: first, whether software companies can demonstrate that “industry domain experience” truly leads to higher-quality agentic outcomes; second, whether their financial fundamentals can remain stable or improve.

This round of software correction is centered on the “terminal value debate”

Goldman Sachs believes: The 2026 decline shifted the debate from “short-term growth curve” to “whether moats will be eroded by AI.” The focus is mainly on application software, but it is beginning to affect infrastructure/security stacks and the ROI of cloud vendor capital expenditures.

Thus, the report reads more like “debunking arguments”: lining up the seven bearish points from “straw man” to “steel man,” assigning risk scores, and trying to answer the same question—what can still support terminal value.

SoR is unlikely to be overturned, but “value migration” is more dangerous

• A: Low risk of SoR being “replaced” (score 1)

The first bearish argument is “rip and replace”: new players using AI to overhaul the system of record layer, making ERP/CRM/HR and similar foundational systems obsolete. Goldman Sachs rates this as low risk, with a straightforward reason: generative AI is more like an analysis and generation engine rather than a transaction engine; enterprise AI requires large amounts of high-quality, structured, traceable data, and SoR is precisely the container and governance system for these data.

The report also acknowledges that genuine replacement risks are not absent: if someone rebuilds a more modern, scalable, and lower TCO architecture at the SoR layer, it could trigger migration. For example, SAP S/4HANA cloud upgrades: large enterprises typically undertake multi-phase migrations over 18–36 months, which are costly and lengthy, leaving room for “cheaper, faster alternatives.”

Goldman Sachs’s defensive moves focus on architecture: SoR should evolve from passive ledgers to “system of reason,” from “AI-powered (with plugins)” to “AI-native (built-in architecture).” Signals include Salesforce’s replatforming in 2024 and Workday’s shift from closed to open systems.

Another key variable is enterprise data boundaries. If companies continue to “circle” their data advantages within existing applications (e.g., Salesforce’s adjustment of Slack API terms in May 2025 to restrict LLM training and batch exports), then SoR as the foundation remains stable, but its profit pools could be siphoned off by new layers.

• B: Value shifting from SoR to “agentic operating/orchestration layers” (score 4)

Goldman Sachs sees a more realistic risk not as SoR disappearing but as SoR becoming a “compliance data foundation,” with value concentrated in orchestration layers capable of cross-system reasoning, API calls, and workflow automation. Agents can read/write across multiple SoRs, reconcile data, and users no longer need to directly access original systems. The moats built on UI dependence, process lock-in, and user habits will weaken.

The report depicts this world as “who sits on whom”: Sierra layered over Salesforce, Anthropic Cowork over Microsoft, with incremental budgets more likely captured by upper layers. Goldman Sachs also warns that the market’s sensitivity to this stems partly from the weak moats of some application companies that expanded during the low-interest-rate cycle of 2020–2021, making them more vulnerable to “disintermediation” narratives.

Opportunities for traditional vendors hinge on “domain experience + context.” The report cites several companies illustrating “why context is valuable”:

• Microsoft emphasizes that staying within the same ecosystem reduces latency, ensures data freshness, and provides more context for LLMs, while large-scale data migration friction, costs, and “breaks” are often underestimated;
• HubSpot attributes enterprise AI’s key shortcoming to “lack of context,” which system of record layers can aggregate—customer history and collaboration info—reducing repetitive “teaching AI”;
• Datadog showcased at the 2/12 analyst day that internally trained SLM models, with lower costs, deliver higher accuracy, emphasizing that “domain experience” translates into differentiation at the model and results level.

Vertical software is more resilient in the short term, but “being good enough” could shift pricing power (score 2)

• The third bearish argument is “horizontal platforms eating into verticals”: horizontal platforms using AI tools enable customers to build industry workflows themselves, eroding vertical software’s pricing power. Goldman Sachs rates this risk as 2, citing that vertical software still faces several barriers: proprietary industry data, deep embedding in workflows forming SoR attributes, long-standing reputation, and regulatory compliance barriers in heavily regulated industries.

The report uses Guidewire as an example: among its clients, about $775 billion in P&C insurance premiums are managed with at least one Guidewire product, and this historical data creates a moat difficult for outsiders to replicate. Goldman Sachs emphasizes “time” as a factor: deeply embedded vertical software often sees customer switching measured in years, not months.

However, the report also highlights new impacts from horizontal/AI-driven solutions: Palantir’s collaborations with AIG and Anthropic in insurance use cases; Intuit’s GenOS enabling easier vertical workflow coding within horizontal accounting software like QuickBooks. The key question: when AI features on horizontal platforms are only “good enough” rather than “significantly better,” will they still attract customers due to easier integration and less fragmentation—directly affecting long-term vertical software pricing power.

Cheaper code means more competition, but building products ≠ building companies (score 2)

• The fourth bearish argument is “declining code costs.” Goldman Sachs admits AI coding tools will lower development barriers and attract more entrants, but rates this risk as 2 because: software engineering is not just coding; engineers spend significant time on design, debugging, risk assessment, and reviews. Efficiency gains do not eliminate the need for development roles.

The report cites a “people still in the loop” statistic: Faros’s study of 10,000 developers shows that teams using high AI adoption complete tasks 21% faster and merge 98% more pull requests, but review time for pull requests increased by 91%. Efficiency improvements will shift bottlenecks to new stages, especially in enterprise delivery, where security, maintenance, integration, workflow orchestration, ecosystem building, and GTM remain hard tasks.

“Future is customization” could divert some budgets; Palantir’s platform for custom solutions (score 3)

The fifth bearish argument is “companies prefer to build in-house.” Goldman Sachs’s conclusion is more of a nuanced compromise: declining code costs won’t universally shift from build vs. buy, but companies will allocate budgets to internal development in certain scenarios, rated as 3. The main reason: maintenance costs and responsibilities will compound over time— even if agentic efficiency reduces maintenance costs, professional vendors’ maintenance costs will also decline, and the “performance/cost frontier” often remains with vendors.

The report suggests that the “middle ground”—between traditional SoR and new AI-enabled solutions—may be most vulnerable to in-house development, especially where cross-department coordination is needed and past software integrations were poor.

Palantir is cited as an example of a custom platform: co-developing production-level AI use cases with clients, emphasizing measurable ROI. Its growth data: Palantir’s US commercial revenue grew 109% in 2025, with an expected acceleration to over 115% in 2026. Palantir relies on frontline deployment engineers (FDEs) to translate client intent into operational systems and to turn client-specific solutions into reusable capabilities; despite questions about “software vs. services,” the company maintains about 85% gross margin.

Goldman Sachs also warns that the in-house build wave may be nearing a “local peak”: SaaS vendors are adding AI capabilities, data governance, and security protocols (e.g., A2A, MCP), and IT teams are climbing the learning curve. ServiceNow has publicly discussed regaining budgets that previously went toward “building in-house.”

“LLM tax” will pressure margins: more realistic in the short term (12–24 months), but long-term remains about pricing power (score 3)

• The sixth bearish argument is that gross margin structures are being rewritten. Goldman Sachs expects the industry to experience 12–24 months of moderate margin pressure: vendors may initially absorb GPU inference costs and third-party model API fees to boost adoption. Because AI makes “usage intensity” a direct cost (tokens, model complexity, query frequency), SaaS shifts from fixed cost leverage toward a “pay-as-you-go” economics.

The report cites Bessemer’s observations: some of the fastest AI-native companies to reach $100 million ARR have gross margins around 25%, with many even negative; more mature AI-native firms often hover around 60%, still below traditional SaaS.

However, Goldman Sachs does not see this as a permanent collapse: citing Epoch AI data, LLM inference costs are decreasing 9 to 900 times annually; prices for models approaching GPT-4’s performance (e.g., MMLU benchmarks) are dropping about 40 times per year. Long-term margin recovery depends on “pricing power = differentiation.” The report highlights Microsoft’s structural advantage: vertical integration combined with its relationship with OpenAI allows capturing profits at multiple levels of the value chain and reducing “LLM tax” leakage.

The hardest to price is the pace of technological change: uncertainty itself depresses valuations (score 5)

• The seventh bearish argument is rated as the highest risk by Goldman Sachs: rapid technological evolution makes the ultimate outcome unpredictable. The report lists recent updates—Anthropic (Cowork, Opus 4.6, vertical plugins), OpenAI (Frontier, OpenClaw), Google DeepMind (Deep Think), Meta (Avocado). It cites Bridgewater’s November 2025 white paper: the pretraining scaling laws are still in effect; recent model updates and benchmark scores (e.g., GPQA Diamond with multiple models >90%) support ongoing progress.

It uses two “inflection points” from “wrapping” to illustrate unpredictability: ChatGPT’s ability to democratize by packaging capabilities into accessible interfaces; Cowork pushing capabilities to desktop GUIs, enabling non-technical users to experiment. Looking further ahead, the diffusion of self-hosted agents like OpenClaw, as described in the report and in conversations with Cloudflare CEO Matthew Prince, could replicate the speed of ChatGPT’s spread over the next three years, with enterprise adoption primarily constrained by security in the short term.

Uncertainty may also create new TAM: the Microsoft MAI Superintelligence Team’s case: MAI-DxO’s 85% success rate in NEJM case challenges, and the TAM estimate based on inputting initial metrics into ChatGPT, reaching $50–$100 billion annually (up to $150–$200 billion in optimistic scenarios). But Goldman Sachs’s point isn’t to bet on a specific outcome; rather, that the unknown makes terminal value harder to anchor, and high uncertainty often leads to undervaluation multiples.

Key signals to “stay steady”: domain experience validation and fundamental stability

Goldman Sachs condenses observable stable signals into two: First, whether enterprise software companies can demonstrate through products and case studies that domain experience truly yields higher-quality agentic results; second, whether their financial fundamentals can remain stable or improve (especially verified through quarterly earnings). Prior to that, it favors “architectural moats”—moats not only at the application interface and workflow level but extending into deeper technical and platform structures.

All the insightful content above is from Wind Trading Desk.

For more detailed analysis, real-time insights, and frontline research, join the 【**Wind Trading Desk▪Annual Membership**】

![](https://img-cdn.gateio.im/social/moments-405e5113af-27450a9e61-8b7abd-d8d215)

Risk Disclaimer and Terms of Use

Market risks exist; investments should be cautious. This article does not constitute personal investment advice and does not consider individual user’s specific investment goals, financial situation, or needs. Users should consider whether any opinions, viewpoints, or conclusions herein are suitable for their circumstances. Investment is at their own risk.