What kind of software will be eliminated by AI?

What kind of software will be eliminated by AI?

```

The software sector's pullback so far in 2026 is different from previous rounds of corrections driven by "slowing demand/rising interest rates": the market is more like discussing terminal value—can these companies still defend their profit pool ten years from now, and will their moats be cut open by "agentic" AI?

According to Trend Trading Desk, Gabriela Borges, an analyst at Goldman Sachs Global Investment Research, wrote directly in her report on the 16th: "The market is questioning software moats and business models." She broke down the seven most common bearish arguments from investors, gave them risk scores from 1 to 5, and distinguished whether they affect narrowly defined application software or spill over into broader infrastructure/security stacks, and even cloud vendors' capex-related ROI.

Interestingly, Goldman does not consider "system-level software being completely replaced by AI" to be the main risk (score 1). The sharper concerns lie in two directions: First, the shift of value from the System of Record (SoR) to the "agentic operating system/orchestration layer" (score 4); second, the rapid pace of technological iteration itself makes the endgame difficult to price (score 5)—it's hard to find a "valuation floor."

In this uncertainty, the report also gives clear handles: watch two types of signals—first, whether software companies can prove that "industry domain expertise" really delivers higher-quality agentic results; second, whether financial fundamentals can remain stable or even improve.

This round of software pullback: Market focuses on the "terminal value debate"

Goldman's view is: The 2026 downturn shifts the debate from "short-term growth curves" to "whether AI will erode moats." The focus is mainly on application software, but it's starting to affect infrastructure/security stacks and capex-related investments by cloud vendors.

The report is therefore written more like "debate deconstruction": lining up seven bearish arguments from "strawman" to "steelman," assigning risk scores, to answer the same question—what can still support terminal value.

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

  • A: The risk of SoR "being replaced" is very low (score 1)

The first bearish argument is "rip and replace": new players use AI to redo the system of record layer, making ERP/CRM/HR and other foundational systems obsolete. Goldman rates this as low risk, for a direct reason: generative AI is more like an analysis and generation engine, rather than a transactional engine; enterprise AI needs large amounts of high-quality, structured, traceable data, and SoR is the container and governance system for this data.

The report acknowledges that real replacement risks are not nonexistent: if someone builds a more modern, scalable, cheaper SoR architecture, migration could be triggered. SAP S/4HANA cloud upgrade is cited: large enterprise migration is typically an 18–36 month multi-phase project, expensive and lengthy, which leaves room for “cheaper, faster alternatives.”

Goldman's defensive moves focus on architecture: SoR must move from passive ledger to “system of reason,” shifting from “AI powered (add-on)” to “AI native (built-in).” Signals mentioned include Salesforce's 2024 replatforming and Workday's move from closed to open architecture.

Another key variable is enterprise data boundaries. If companies continue to “ring-fence” their data advantage in existing apps (e.g., Salesforce’s May 2025 Slack API changes, restricting LLM training and bulk exports), the SoR layer remains stable, but the profits above it could be siphoned off by new layers.

  • B: Value moves from SoR to "agentic OS/orchestration layer" (score 4)

Goldman sees the real risk not in SoR disappearing, but in SoR becoming a "compliant data substrate," with value concentrating on orchestration layers that can reason across systems, call APIs, and automate workflows. Agents can read/write/reconcile across SoRs, so users no longer need to access original system UIs; SoR moats built on UI, process ownership, and user habit are weakened.

The report uses "who sits atop whom" to describe this world: Sierra on Salesforce, Anthropic Cowork on Microsoft; incremental budgets may be captured more by the upper layers. Goldman notes the market is particularly sensitive to this line partly because a batch of application companies that expanded during low-rate cycles in 2020/2021 have weaker moats and are more prone to being “collapsed” by disintermediation narratives.

The remaining opportunity for traditional vendors is "domain expertise + context." The report quotes several companies to illustrate "why context is valuable":

  • Microsoft stresses staying in the same ecosystem reduces latency, ensures data freshness for LLMs, and that friction/cost/discontinuity of large-scale data migration are often underestimated;
  • HubSpot identifies the key shortfall of enterprise AI as “lack of context,” and SoR layers aggregate customer history/collaboration information, reducing repeated "AI teaching";
  • Datadog's 2/12 analyst day showed internally trained SLMs with lower costs yield higher accuracy, emphasizing that domain expertise translates into model and outcome differentiation.

Vertical software is more resilient short term, but "good enough" usability may alter pricing power (score 2)

The third bearish argument is "horizontal eats vertical": horizontal platforms use AI tools to let clients build industry workflows themselves, eroding vertical software's pricing power. Goldman gives this a risk score of 2; vertical software still has several barriers: proprietary industry data, deep workflow embedding giving it SoR attributes, long-term reputation, and compliance walls in highly regulated industries.

Guidewire is cited for data scale: in its customer base, about $775 billion in global P&C Insurance DWP is managed by at least one Guidewire product, and data accumulation is a hard-to-replicate moat. Goldman also emphasizes "clients give time": deeply embedded vertical software sees tolerance and switching measured in years, not months.

But Goldman doesn't dismiss the risk outright. The report also lists new shocks from horizontal/AI: Palantir's insurance use case cooperation with AIG/Anthropic; Intuit's GenOS lets users encode vertical workflows in horizontal accounting apps like Quickbooks. The key question: when horizontal platforms’ AI functions are merely “good enough” and not “markedly better,” will integration simplicity and reduced fragmentation still lure away customers—this directly affects vertical software's long-term pricing power.

Cheaper code brings more competition, but writing a product ≠ building a company (score 2)

The fourth bearish argument: "code costs drop." Goldman admits AI coding tools lower barriers and bring new entrants, but risk is 2, as software engineering is not just coding—lots of time is spent on design, debugging, risk assessment, review; efficiency gains do not mean developer jobs disappear.

Report provides "human still in the loop" data: Faros study of 10,000 developers shows high AI-using teams complete 21% more tasks, merge 98% more pull requests, but review time increases 91%. Efficiency gains push bottlenecks to new steps; in enterprise delivery, security, maintenance, integration, process orchestration, ecosystem building, and GTM remain hard work.

"Custom is the future" takes some budgets: Palantir turns customization into platform (score 3)

The fifth bearish argument: "companies prefer to build in-house." Goldman's conclusion is blended but clear: lower code costs won't universally change build vs buy, but companies will shift some budgets to internal builds in certain cases, risk score 3. Key reason: maintenance costs and responsibility compound over time—even if agentic efficiency lowers maintenance costs, professional vendors also lower theirs, so the "performance/cost frontier" is often with vendors.

The most likely to be grabbed by in-house builds, according to the report, are "the middle ground"—between traditional SoRs, requiring cross-department coordination, areas poorly connected by bundled software.

Palantir is seen as the customization paradigm: via AIP, it co-builds production-grade AI use cases with clients, emphasizing quantifiable ROI. Growth data: Palantir US Commercial business grew 109% in 2025, expects to accelerate over 115% in 2026E. Palantir leverages Forward Deployed Engineers (FDEs) to translate client intent into systems, then distills client-specific solutions into reusable capabilities; despite doubts about "software or services," this hybrid model maintains about 85% gross margin.

Goldman also notes in-house build enthusiasm may be peaking: SaaS vendors are catching up on AI, data governance/security protocols (e.g., A2A, MCP) are evolving, and IT teams are upskilling. ServiceNow has openly stated it's winning back budgets that previously shifted to "build in-house."

"LLM tax" suppresses gross margin: More relevant for the next 12–24 months, long term is about pricing power (score 3)

Sixth bearish argument: gross margin structure rewrite. Goldman projects industry-wide mild gross margin pressure for 12–24 months: vendors may swallow GPU inference costs and third-party model API fees to boost adoption. Because AI converts "usage intensity" directly to cost (tokens, model complexity, query frequency), SaaS economics shift from fixed cost leverage toward "pay-as-you-consume."

The report cites Bessemer: some fastest-growing AI-native companies (from 0 to $100M ARR) have gross margins around 25% or even negative; more mature AI-native companies have gross margins at about 60%, still below traditional SaaS.

But Goldman doesn't see this as permanent collapse: Epoch AI data shows LLM inference costs drop 9–900x annually; GPT-4 level MMLU performance price drops roughly 40x per year. Whether gross margin rebounds long term still comes down to "pricing power = differentiation." Goldman highlights Microsoft's structural advantage: vertical integration and its OpenAI relationship allow profit capture at multiple value chain layers, reducing third-party "LLM tax" drain.

The hardest to price is tech velocity: Uncertainty itself pressures valuation (score 5)

The seventh bearish argument is rated highest risk by Goldman: tech evolution is too fast, endgame unpredictable. The report lists updates since early this year—Anthropic (Cowork, Opus 4.6, vertical plug-ins), OpenAI (Frontier, OpenClaw), Google DeepMind (Deep Think), Meta (Avocado). It cites Bridgewater’s Nov 2025 white paper: pretraining scaling laws still apply, lists latest model updates and benchmarks (e.g., GPQA Diamond with multiple models >90%).

The report uses two "junctions brought by packaging" to illustrate unpredictability: ChatGPT brought capabilities to the public via a user-friendly interface; Cowork pushed capabilities to desktop GUI, allowing non-technical users to experiment. Looking forward, OpenClaw and self-hosted agent diffusion are described (in a conversation with Cloudflare CEO Matthew Prince) as likely to match ChatGPT’s rapid adoption in the next 3 years, with enterprise short-term constraints mainly in security.

Uncertainty may also bring new TAM. Microsoft MAI Superintelligence Team’s case: MAI-DxO’s success rate reached 85% in New England Journal of Medicine challenge; inputting Microsoft blog and initial metrics to ChatGPT yields TAM estimate at $50–100B/year (bull case $150–200B). But Goldman’s point isn’t "betting on endgame"—it admits: the unknown itself makes terminal value harder to anchor, and high uncertainty typically means low valuation multiples.

Signals to watch for “steadiness”: Domain expertise delivers, fundamentals don’t break

Goldman boils observable stability signals down to two: First, whether enterprise software companies can demonstrate via products and cases that domain expertise truly delivers higher-quality agentic results; second, whether financials can remain steady or even improve (especially verified by earnings seasons). Until then, it favors "architectural moats”—moats aren’t just at the app UI and workflow layer, but extend to technology and platform structure beneath.

 

~~~~~~~~~~~~~~~~~~~~~~~~

Content above courtesy of Trend Trading Desk.

For more in-depth interpretations, including real-time analysis and frontline research, join【Trend Trading Desk▪Annual Membership】

Risk warning and disclaimerThe market contains risks, investment should be cautious. This article does not constitute personal investment advice, nor does it consider individual users’ specific investment objectives, financial situation, or needs. Users should consider whether any opinions, views, or conclusions in this article suit their specific circumstances. Investing based on this content is at your own risk. ```