Goldman Sachs: AI Is Moving From Assistive Tools to Workflow Execution, with More Value Accruing to Data, Workflow Control, and Agent Infrastructure

Overview

This third annual Silicon Valley AI field-trip note compiles perspectives from startups, venture capital, and academia. Goldman Sachs' core conclusion is that AI commercial value is shifting from information retrieval and assistant functions toward verifiable, accountable workflow execution; the best-positioned platforms will combine proprietary information, industry knowledge, customer context, distribution capabilities, and authority to take action.

Core views

Goldman Sachs held its third annual Silicon Valley AI field trip on August 18-19, with participants including AI companies, venture-capital firms, and researchers from Stanford University, the University of California, Berkeley, and UCSF. The trip reinforced its view that model capabilities are still improving, agents are moving from assisting humans to completing workflows, and monetization will expand from seat-based pricing to charges for data usage, transactions, completed tasks, outcomes, and agent usage. The report argues that the strongest AI companies should possess proprietary data, domain expertise, customer context, workflow control, distribution channels, verification mechanisms, and permission to execute actions. For business and information services, the report believes AI will increase the utility of differentiated information assets: AI makes it easier to retrieve, connect, and apply data to customer decisions, but large datasets alone do not constitute a moat. Leading platforms also require authoritative content, validated models, specialized taxonomies, customer context, and experts who understand how information is used to provide more accurate, verifiable results in financial analysis, legal, insurance, risk management, healthcare, and other settings. Companies that historically monetized through subscriptions, reports, and labor-intensive services can use agents to organize, analyze, monitor, and act on unstructured information, upgrading from simple information-input providers to the "operating systems" of knowledge-intensive industries and capturing a larger share of customers' professional-services spend. The report further notes that financial analysis, legal research, insurance underwriting, and market research contain many high-cost professional-labor processes. Encoding the structured portions into reusable agents can both accelerate existing work and commercialize tasks that were previously uneconomic; lower delivery costs may also improve service accessibility, usage frequency, and the range of questions customers are willing to ask. The most suitable model is to automate standardized steps while retaining forecasting, material decisions, and exception handling for expert judgment. As agents independently complete more tasks and each professional can supervise more activity, seat-based pricing becomes less aligned with customer value; platforms can instead charge for incremental data consumption, completed analyses, processed documents, transactions, and agent usage. Daloopa said data consumption in some integrated agent workflows has increased by approximately 100x; Clio has subscription, payments, and other transaction-based revenue, while Moody's has begun charging incremental subscription fees for MCP access and separate fees for agent usage. In the competitive landscape, established platforms have advantages from trusted brands, embedded distribution, accumulated customer context, and access to systems where records, transactions, payments, and action triggers reside; these capabilities reduce the need for customers to move information across systems. AI-native companies can still gain share through substantially superior products, specialized workflows, or difficult-to-replicate infrastructure. For example, Corgi combines AI, insurance licenses, regulatory capital, and control over underwriting and claims, while Harvey combines legal expertise, workflow-specific agents, and evaluation capabilities. By contrast, providers with commoditized data, shallow workflow integration, or business models that primarily rely on human labor and user-count pricing are more vulnerable. Goldman Sachs emphasizes that whether enterprise AI can enter production depends on whether outputs are verifiable, erroneous actions are reversible, and accountability is clear; source-linked data, audit trails, confidence thresholds, independent evaluations, and targeted human review can reduce oversight burdens without eliminating responsibility. On the division of labor between open-source and frontier models, the report rejects an either-or view. The frontier-model camp argues that enterprise benchmarks do not fully capture real capability leaps and that, in production, the loss of accuracy may not justify model-cost savings; therefore, some AI-native applications continue to rely primarily on frontier providers even while discussing model diversification. On the other hand, industry participants believe most enterprise workflows do not require frontier-level intelligence, with one venture capitalist estimating that approximately 90% of tokens will flow to open-source models within 12 to 18 months; pricing pressure, higher usage limits, and customers reducing frontier-model consumption also indicate that competition has shifted toward economics. The report's overall view is that frontier models will retain the highest-value, most reliability-sensitive workloads, while open-source models will handle most usage volume. Independent software vendors such as Microsoft, Cloudflare, Databricks, and Vercel can capture value by helping enterprises orchestrate different models while keeping content private. The report also believes that world models may represent the next AI platform shift, potentially larger than large language models. Research is shifting from simple next-token prediction toward modeling environments, causal relationships, physical laws, and real-world interactions; the competitive moat may likewise shift from shared internet-scale data to proprietary data closely tied to physical systems, industries, and operating environments. Physical, industrial, scientific, and robotic systems represent a larger problem space than text generation, and many workloads are more compute-intensive. World-model token demand will add to Goldman Sachs' forecast that enterprise compute will support 24x more tokens over the next five years, potentially keeping compute supply-demand conditions tight for longer and benefiting Microsoft, Oracle, and CoreWeave. Durable differentiation at the software layer will come from workflow context rather than data alone. Simple AI wrapper layers are most vulnerable as models improve; horizontal platforms, while benefiting from scale, also face feature competition. The report places greater value on workflow-native software that combines industry expertise, embedded processes, and proprietary context that continuously generates, structures, and enriches data. As software development becomes faster and cheaper, this context may become one of the few remaining defensible sources of differentiation; the report cites Shopify and Samsara as examples of how an installed customer base or connected-asset footprint can continuously generate the operating context needed for differentiated AI. Agents will also become a new customer-acquisition channel: software must be discoverable, callable, and usable by agents acting on users' behalf. Vercel's view indicates that the internet's basic building blocks are shifting from users and applications to agents and tokens, and infrastructure is no longer optimized solely for human interaction. Vercel and ClickHouse note that agents will generate more requests, execute more actions and decisions, and connect more systems, shifting bottlenecks from application development to infrastructure supporting autonomous execution; the importance of real-time databases, low-latency compute, scalable data pipelines, and globally distributed systems is increasing. Cybersecurity is another important beneficiary. AI will expand the attack surface much like prior platform shifts such as the internet and cloud, and it will increase attack frequency and complexity as open-source models enhance attackers' capabilities, potentially driving near-term investment in patching, hygiene management, personnel processes, and product trials. In a long-term scenario, defenders that gain near-complete visibility into their environments may identify attackers' anomalous behavior earlier. AI can also alleviate cybersecurity talent shortages and vendor fragmentation, particularly in security operations: modernization built on SIEM data storage, continuous remediation, properly configured automated systems, and agent-native tools that can reason over data, execute controls, and take action could all represent opportunities. Company examples support the above logic. Daloopa serves research, valuation-model maintenance, and earnings analysis by extracting, standardizing, and validating financial-statement items, KPIs, guidance, and non-GAAP adjustments from regulatory filings, earnings materials, and other disclosures, with every data point linked to its original source. Its approximately 400-person verification team, proprietary data pipeline, and structured operating processes are designed to meet accuracy and timeliness requirements; the company open-sources standardized agents and analytical skills while keeping its data extraction, standardization, and verification infrastructure proprietary, and uses a continuously updated knowledge graph to preserve testable research memory, reduce repeated generation, and limit hallucinations. Divergent demonstrated how vertical AI can integrate the full design-to-production workflow: its DAPS connects AI engineering software, additive manufacturing, and automated assembly; before engineers select an approach, the software evaluates hundreds of structural alternatives, and the integration of its software, Model One printer, manufacturing processes, and factory operations enables production data to feed back into design and manufacturing decisions. The company said investment in this shared software backbone has exceeded $1 billion.

Analysis framework

Based on two days of company demonstrations, venture-capital interviews, and university research exchanges, the report first synthesizes common trends in models, agents, and monetization, then separately analyzes value transmission across business and information services, software, compute, physical AI, and cybersecurity. Its analysis focuses on how data, workflows, verification, and distribution capabilities shape competitive moats, using cases such as Daloopa and Divergent to illustrate the mechanisms.

Methodology notes

• Competition and Strategy FrameworkValue chain analysis

  Allocation of AI value among models, data, applications, workflows, infrastructure, and security layers

  The report compares the sources of value across different layers of the AI stack, arguing that models or data alone are insufficient to ensure an advantage; platforms that control customer workflows, verify outputs, and are embedded in action steps are more likely to capture durable value.

• Industry/Industrial Analysis FrameworkUpstream, Midstream, and Downstream Industry-Chain Transmission

  Transmission of model capabilities and token growth to software, compute infrastructure, physical AI, and cybersecurity

  The report links improved models and increased agent usage to changes in demand for real-time data, low-latency compute, data pipelines, security operations, and compute for physical systems.

• Competition and Strategy FrameworkMoat / competitive advantage

  Proprietary data, domain expertise, customer context, verification, and workflow control create AI moats

  The report uses this logic to distinguish established platforms, AI-native companies, and vulnerable simple wrapper layers or commoditized-data providers.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Business & Information Services (IRM, MCO, MSCI, SPGI, TRI)

  The report identifies this as a preferred AI-beneficiary area, arguing that differentiated data, industry expertise, and embedded workflows can expand share of customer spending.

  Strengths

  Trusted and continuously updated data assets, specialized expertise, customer context, verification capabilities, and workflow integration.

  Comparison

  Relative to providers offering only commoditized data or shallow integration, these companies have a greater opportunity to create verifiable, executable industry workflows.

  Risks

  Competitive positioning may come under pressure if data becomes commoditized, workflow integration is insufficient, or business models rely primarily on human labor and user counts.

• Software (MSFT, MDB, RBRK)

  The report identifies this as a preferred AI-beneficiary area; Microsoft is also viewed as a potential beneficiary if world models create more prolonged compute tightness.

  Strengths

  Can help enterprises orchestrate different models, keep content private, and capture agent-driven infrastructure and software demand.

  Comparison

  Relative to simple AI wrapper layers, platforms with enterprise distribution, workflow, and infrastructure capabilities are more defensible.

  Risks

  Rapid improvement in model capabilities may intensify feature competition and alter value allocation across software layers.

• Cybersecurity (CRWD, NET, PANW)

  The report reiterates constructive long-term views on these companies, believing AI will expand security demand.

  Strengths

  Security operations can benefit from automation, continuous remediation, data integration, and agent reasoning and actions based on SIEM data.

  Comparison

  The report believes security vendors with established capabilities may receive disproportionate benefits from AI-driven security investment.

  Risks

  AI will also increase attack frequency and complexity and expand the enterprise attack surface.

• SNOW

  The report reiterates a constructive long-term view and places it in the context of expanding demand for AI data and infrastructure.

Key data

• Silicon Valley AI field-trip datesAugust 18-19, 2026Goldman Sachs' third annual Silicon Valley AI field trip.
• View on open-source model token shareApproximately 90%A venture capitalist believes that most tokens may flow to open-source models within approximately 12 to 18 months.
• Enterprise compute token-demand forecastSupport for 24x more tokens over the next 5 yearsThe report believes world-model demand will be additive to this enterprise-compute forecast.
• Daloopa integrated-workflow data consumptionApproximately 100x increaseThe company said data consumption in some integrated agent workflows is far higher than in traditional analyst interfaces.
• Daloopa verification teamApproximately 400 peopleUsed alongside proprietary data pipelines and processes to meet accuracy and timeliness requirements.
• Divergent shared software-backbone investmentMore than $1 billionAn integrated software platform covering design, additive manufacturing, assembly, and quality processes.
• MSCI rating and target priceBuy, 12-month target price of $671Listed by the report as one of its key AI-beneficiary investment ideas.
• Microsoft rating and target priceBuy, 12-month target price of $640Listed by the report as one of its key AI-beneficiary investment ideas.

Impact & implications

The report believes AI will shift business and information services companies from delivering information to executing workflows, tying revenue more closely to usage, transactions, and outputs. Beneficiaries include not only platforms with trusted data and industry expertise, but also companies that help enterprises orchestrate models, handle agent-native workloads, provide compute, and strengthen security operations; meanwhile, commoditized data and simple AI-wrapper models without deep workflow integration face greater competitive pressure.

Risks

• Cost, performance, and reliability trade-offs between frontier and open-source models may intensify competition among model providers and at the application layer.
• Companies with commoditized data, shallow workflow integration, or business models reliant on human labor and seat counts are more vulnerable to disruption from AI-native competitors.
• In financial, legal, regulatory, and operational settings, the pace at which agents enter production may be constrained if verifiable outputs, audit trails, and clear accountability are lacking.
• AI expands the attack surface and increases attack frequency and complexity, particularly as open-source models enhance attackers' capabilities.