AI Agent Platforms, Enterprise Automation, Agent Operating Systems, and the Workflow Rebuild

Core Conclusions

• An AI Agent is not "a smarter chat box," but "a unit of software labor that is policy-constrained, can call tools, can access permissions, can retain state, and can execute actions." OpenAI has explicitly positioned its Agents SDK as a runtime where orchestration, tool execution, state, and approval sit on the application side; Anthropic and Amazon Bedrock likewise define an Agent as a system that can call tools, access knowledge bases, execute API/desktop actions, and complete tasks in a loop—rather than a one-shot answering machine.

• What fundamentally separates an Agent from a traditional Copilot, chatbot, RPA, or workflow automation is "whether it owns a goal-driven, multi-step execution loop." Chatbots are mainly Q&A; Copilots in most scenarios still center on "assisted generation/assisted retrieval"; RPA is strong at scripted execution with clear rules and stable interfaces; workflow automation is strong at deterministic processes; an Agent strings "plan—retrieve—call tools—manage state—approve—execute—review" into one chain and handles exceptions in uncertain environments. Workday even compares the Agent directly to a chatbot, stressing that the former can understand across context, plan, take action within boundaries, and learn from experience.

• What enterprises will genuinely pay an Agent for is not "answering more like a human," but "completing the work that would otherwise require people, SaaS seats, RPA bots, or outsourced teams." That is exactly why the first commercialized landing spots cluster in customer service, ITSM, sales follow-up, knowledge retrieval, software development, and contract and finance processing—scenarios with quantifiable ROI. HubSpot puts Breeze Agents on customer service, sales, research, and data tasks; Workday's public cases already land on leave, payroll, contract review, and business-process optimization; Zoom puts agentic features into post-meeting follow-up, document generation, and task execution.

• Enterprises need an Agent platform, not just direct calls to a large-model API, and the reason is not the model itself but permissions, connectors, approval, workflow state, traceability, cost control, and governance. Google Gemini Enterprise emphasizes permissions-aware enterprise search, SSO, user-level access control, and cross-SaaS connectors; ServiceNow stresses an "Autonomous Workforce with business context and permissions"; Workday stresses an Agent System of Record; Atlassian binds Rovo to the Teamwork Graph and a large set of connectors; Microsoft Copilot Studio stresses instructions, knowledge, tools, triggers, embedding, and multi-channel publishing.

• The long-term profit pool more likely stays in the "system entry point + workflow + permissions + data graph + governance" layer, rather than the model layer alone. The model layer is capital-intensive, crowded with rivals, and falling fast on price; what is genuinely hard at the enterprise Agent level is wiring the model into SaaS / ERP / CRM / ITSM / document stores / identity systems and executing auditably. Workday's Agent System of Record, Atlassian's Teamwork Graph, Google's permissions-aware enterprise search, Microsoft Graph and Entra, and ServiceNow's platform-style workflow data / AI control tower positioning all show that enterprise value is migrating from "system of record" toward "system of action"—but that migration still has to be built on top of existing systems and permissions.

• The clearest direct financial beneficiaries today are Microsoft, Salesforce, ServiceNow, and Palantir; Oracle leans more toward AI infrastructure and database/cloud-contract upside. Microsoft's Q3 FY26 disclosure shows its AI business annualized revenue run rate already exceeds 37 billion dollars, with commercial cRPO reaching 627 billion dollars; Salesforce has commoditized Agentforce into a blended per-conversation, per-Flex-Credits, per-user model, and its latest public disclosures point to a substantive signal in Agentforce deal size and Data Cloud+AI ARR; ServiceNow has publicly said 2026 AI-related contract value could top 1.5 billion dollars, with Now Assist large customers up more than 130% year over year; Palantir's Q1 2026 total revenue grew 85% year over year, with U.S. commercial revenue up 133% year over year, and AIP is the core growth narrative.

• Atlassian, HubSpot, and Zoom look more like "Agent-enhanced software companies," where the near-term story is more about defense and ARPU restructuring than a proven standalone Agent revenue engine. Atlassian folds Rovo into its cloud subscription and charges separately for Rovo Dev via a developer subscription plus overage credits; HubSpot makes part of Breeze free, embeds part into bundles, and routes some Agents through HubSpot Credits; Zoom keeps AI Companion as the default enhancement to Zoom Workplace and charges extra through Custom AI Companion. Such a strategy aids retention and defense, but in the near term it dilutes the visibility of "new AI revenue."

• The business model is migrating from pure per-seat toward a "seat + consumption + outcome" blend, rather than abandoning the seat model in one step. Microsoft keeps the $30/user/month Microsoft 365 Copilot while introducing per-message agentic consumption pricing in Copilot Chat; Salesforce has both $2/conversation and Flex Credits plus $5/user/month; Atlassian Rovo Dev is $20/developer/month plus credits overage; HubSpot is starting to route some Agent features into credits. The conclusion: Agents aimed at employee productivity sit closer to the seat model, while Agents aimed at process execution and external service sit closer to usage / outcome pricing.

• The track most prone to being overhyped and bubbly is "orchestration frameworks / Agent builders / memory wrappers / general-purpose multi-Agent middleware." This layer faces very strong open-source substitutes: AutoGen, Semantic Kernel, DSPy, CrewAI, and LlamaIndex all stress development frameworks, plugins, extensibility, multi-Agent, or modular design; LangChain, by contrast, more explicitly places its commercial center of gravity on LangSmith's observability, evaluation, deployment, and operations, rather than the "build an Agent" framework itself. In other words, a framework can get hot without being the best at making money.

• The strongest moat layer is not the model, nor the prompt, but "permissions-aware data access + business-system entry points + approval/audit/observability." An open standard like MCP will erode the premium on closed connectors, but it will not eliminate the complexity of enterprise-grade identity, permissions, auditing, cost monitoring, and HITL; on the contrary, the more open MCP becomes, the more an enterprise needs a single control plane to govern it all.

• The real boundaries of the enterprise Agent are still hard, and "Agent washing" is clearly visible in today's market. Gartner expects that by 2027, more than 40% of agentic AI projects will be cancelled over rising costs or unclear business value, and expects that by 2028, 33% of enterprise software will contain agentic AI and 15% of day-to-day business decisions will be made autonomously by agentic AI; meanwhile, MIT's 2025 AI Agent Index notes that most agent developers still offer limited transparency on safety, evaluation, and societal impact.

• Enterprise workflows are far harder than public demos, and the real bottleneck is hidden state, cascading side effects, and reliable cross-system execution. ServiceNow's World of Workflows benchmark shows that the implicit processes and invisible state of enterprise systems give frontier models "dynamics blindness"; Salesforce's SCUBA benchmark shows that zero-shot, open-model-driven computer-use agents succeed on real CRM workflows less than 5% of the time, with closed-source methods reaching only 39%, rising to roughly 50% once demonstrations are added; in APEX-Agents, the best Pass@1 on cross-application, long-chain knowledge-work tasks is still only 24%. This means the core investment variable for the next 12–24 months is not "who shipped an Agent first," but who can turn an Agent from a demo into a reliable production system.

• From an investment standpoint, the most worth watching is not "every beneficiary of the Agent concept," but three kinds of companies: first, platform-type direct beneficiaries that can already convert Agents into new contracts, new usage, and new backlog; second, governance / observability / permission-layer beneficiaries, because the deeper productionization goes, the more governance becomes a hard requirement; third, vertical Agent applications with high workflow barriers, especially in customer service, ITSM, code, legal, finance, and research. Conversely, companies with only Agent marketing and no revenue evidence are more likely to be just a defensive narrative in the near term.

Industry Chain Panorama and Architecture

The conclusion first: the AI Agent value chain is not a two-layer "model companies — application companies" stack, but a seven-layer stack of "model/tool capability — orchestration and execution — connectors and permissions — data retrieval and memory — governance and observability — business-system entry points — vertical execution applications." The highest enterprise value usually does not sit at the bottom layer, but where it is closest to real work and auditable execution. OpenAI, Anthropic, and Amazon Bedrock have already turned the agent loop, tool calling, MCP, guardrails, human approval, and state management into standard capabilities; Google, Microsoft, ServiceNow, Workday, and Atlassian then connect these capabilities, within their enterprise platforms, on top of SSO, data sources, user permissions, and business workflows.

A typical enterprise Agent system today is broadly built from fourteen key components: the model layer, orchestration layer, tool-calling layer, data-retrieval layer, memory layer, permission layer, identity-authentication layer, workflow layer, approval layer, observability layer, security-governance layer, cost-control layer, business-system-integration layer, and user-interface layer. From a moat standpoint, the steadiest is not the model layer but the permission layer, the business-system-integration layer, the workflow layer, and the governance layer; from the angle of what is easiest to replace with open source, the most fragile are the orchestration framework, the prompt/agent wrapper, and the general-purpose memory abstraction; from the angle of what cloud/model vendors are most likely to build in, basic tool calling, the Agent builder, simple RAG, and basic observability are the most at risk.

If forced to answer "which layer most deserves a high valuation," my answer is: not the model at the bottom, nor the chat UI at the surface, but the layer "closest to real work responsibility"—the permissions-aware data and business-operations plane. Because what truly earns money, retains customers, and forms a customer moat is not "whether it can generate," but "whether it can act on the enterprise's behalf, and when something goes wrong, whether it can be traced, intercepted, and corrected." The product designs of ServiceNow, Workday, Google, Microsoft, and Atlassian are nearly all converging in this direction.

Business Model and Financial Elasticity

How do AI Agent platforms charge? Public materials already give a clear answer: the future mainstream will not be a single billing model, but seat fees, consumption fees, and outcome fees coexisting. Microsoft keeps the $30/user/month Microsoft 365 Copilot on one side while extending Copilot Chat's Agent into per-message / Azure-meter pay-as-you-go on the other; Salesforce offers $2/conversation, $500/100,000 Flex Credits, and a $5/user/month Agentforce User License all at once; Atlassian folds Rovo's base capabilities into its cloud subscription, then sells Rovo Dev at $20/developer/month plus overage credits; HubSpot makes part of Breeze free, embeds part into bundles, and settles some Agents in HubSpot Credits; Zoom's AI Companion stays "bundled by default," with Custom AI Companion as the add-on charge.

The advantage of per-seat pricing is simple procurement, certain budgets, and a good fit for employee-collaboration assistants; the drawback is that once an Agent replaces human operation, customers start asking "why is a bot still being charged per head." The advantage of per-task/per-process/per-outcome pricing is that ROI is easier to align, fitting customer service, sales outbound, ticket processing, contract review, and ticket routing; the drawback is that budgets are harder to forecast and customers are initially reluctant to sign a "blank check." So the most common form over the next 2–3 years is foreseeable: the foundation keeps the seat, while the automated-execution layer switches to credits / conversations / actions / usage.

Will Agents raise SaaS ARPU or reduce seats? The answer is both will happen, but at different companies and in different scenarios. If a company holds a strong system-of-record and workflow-control, Agents tend to raise ARPU first, bring usage growth, and strengthen customer stickiness; if a company is only a single-point tool built around human users, the Agent will compress seats and reduce seat expansion. HubSpot is the public market's most typical case of "investors worrying that AI efficiency cuts headcount and thereby hits the seat model"; UiPath faces a different pressure—RPA must prove it can upgrade into an agentic-automation orchestration layer, or it will be swallowed from above by platform-type software and the model-tool layer.

How does model-call cost affect gross margin? This will be one of the biggest margin-divergence variables in enterprise software over the next 12–24 months. Anthropic's official docs already make clear that server-side tools carry additional usage-based pricing; the product pricing of Microsoft, Salesforce, Atlassian, and HubSpot is also migrating toward credits/meters, which in essence externalizes part of the token and tool-call cost to customers. More crucially, Salesforce's public compound-AI-system deployment research shows that through modular reasoning architecture, dynamic scaling, and MLOps, P95 latency can be cut by more than 50%, throughput raised to 3.9x, and cost reduced 30%–40%. The conclusion is direct: the core of future AI-software gross margin is not just cheaper models, but who can turn the agentic workload into an optimizable, routable, observable system.

The anchors for these three scenarios are not pulled from thin air: on one hand, Gartner has publicly forecast that by 2028, 33% of enterprise software will contain agentic AI and 15% of day-to-day business decisions will be made autonomously by agentic AI; on the other hand, it explicitly warns that by 2027 more than 40% of projects may be abandoned over unclear ROI. So the most reasonable judgment is not "Agents will surely swallow software fast," but commercialization will land first in high-ROI workflows and penetrate through the installed base, rather than fully rebuilding everything at once.

Tracks and Competitive Landscape

The enterprise Agent platform is the major track most worth taking seriously today, because it is closest to the budget entry point and the easiest to turn an Agent directly into a sustainable contract. The winning logic here is not "who has the best model," but "who is closest to the user entry point, enterprise data, permission systems, and execution workflows." Microsoft leans on Microsoft 365, Graph, Entra, and Copilot Studio; Salesforce leans on CRM, Data Cloud, Flow, Slack, and Agentforce; ServiceNow leans on its workflow platform and AI Control Tower/Autonomous Workforce; Google leans on enterprise search, connectors, and permissions-aware Q&A; Workday tries to use the Agent System of Record to seize "digital employee management," a new system definition.

Agent orchestration frameworks / SDKs are the most technically active and the easiest to spin a story around in capital markets, but also the easiest to open-source and platform-ize. AutoGen, Semantic Kernel, DSPy, CrewAI, LlamaIndex, Mistral Agents, and Cohere tool use are all lowering the bar to "build an Agent"; the reason LangChain's commercial value is steadier is that it has shifted its focus to LangSmith's tracing, evaluation, deployment, and fleet. In other words, "able to build" will soon be worthless, while "able to test, manage, and ship" is where value starts.

Connectors / MCP / iPaaS may be the most underestimated layer in the future profit pool. Because once an Agent moves from "answering" into "executing," it must connect CRM, ITSM, ERP, email, calendar, documents, browser, database, and internal APIs. MCP has been described as an open protocol across clients and servers, supported by ChatGPT, Claude, VS Code, Cursor, and others; OpenAI also lists remote MCP directly as a tool type. An open standard erodes the format barriers of closed platforms, but it does not erode the enterprise's willingness to pay for authentication, auditing, permission mapping, and connection stability. So connectors will be open-sourced; connector operation and governance will not.

Governance, security, and observability is the track that most resembles "the next-generation middleware." Its value is not at the demo stage, but in the fact that "once a project is going live, the company's legal, security, audit, and IT teams suddenly have to buy." OpenAI's guardrails and human review, Anthropic's strict tool use, Google's access controls and user-level access, ServiceNow's AI Control Tower, and Salesforce's Digital Wallet with usage monitoring all show that once an Agent executes real actions, governance turns from optional into a commercial necessity. From an investment logic standpoint, this layer usually has high margins, high repeat purchase, and high switching costs—only its near-term market size will not be as large as that of the frontline application.

The workflow Agents that land first are not the coolest general-purpose assistants but the most "boring," high-frequency processes: IT tickets, service tickets, sales follow-up, meeting follow-ups, document research, contract analysis, expense and leave/attendance queries, coding, and code review. The reason is not technical preference but that these scenarios connect directly to financial metrics like "reduce labor," "shorten cycle time," "raise throughput," and "raise first-contact resolution." The public product roadmaps of Workday, HubSpot, Zoom, ServiceNow, and Salesforce all bear this out.

The impact on traditional software and services will happen first in four kinds of places. The first is seat-intensive, single-point-function SaaS; the second is traditional RPA reliant on static scripts and templates; the third is low-end call centers and BPO; the fourth is low-end software-development outsourcing and simple reporting/research outsourcing. Enterprises will gradually migrate large amounts of budget from "buying software for one person to use" to "buying a system that completes the task." But note that incumbent software vendors with a system of record, domain data, and process control are not necessarily disrupted; they are more likely forced to upgrade from a system of record into a system of action. This is exactly why ServiceNow, Workday, Microsoft, Salesforce, and Atlassian deserve more attention than pure "AI-assistant shells."

From a competitive-landscape view, my ranking is: Microsoft and ServiceNow are the strongest candidates for the "enterprise Agent control plane"; Salesforce is the strongest candidate for "customer workflow and external service Agents"; Palantir is the high-elasticity candidate for "high-value complex workflows and government/industrial system Agents"; Google and Workday are potential strong platforms for "search/knowledge and digital-labor management"; Oracle leans more toward a supporting beneficiary of infrastructure and database/industry SaaS. But viewed through "whether the market has already fully priced in expectations," Palantir is the most expensive, Microsoft the steadiest, Salesforce/ServiceNow more balanced on risk/reward, and HubSpot/UiPath carry the larger expectation gap. This is not a buy/sell recommendation, but a relative judgment along the three axes of "certainty—elasticity—valuation."

Target Tiering and Scoring

The tiering below applies only to the companies with the strongest evidence in current public materials; it does not force every regional company into a bucket, because many companies have only shipped Agent features without disclosing Agent revenue, ARR, usage, RPO, or customer-conversion evidence.

The scoring model I suggest still follows the logic you provided, but in this round of public-material validation it fits a qualitative + semi-quantitative score better:

• Direct Agent-revenue exposure: 25%

• Product/data/workflow barriers: 20%

• Customer quality and revenue certainty: 15%

• Platform ecosystem and connector capability: 15%

• Financial quality and margin: 10%

• Valuation reasonableness: 10%

• Future catalysts: 5%

Based on current evidence, my priority research ranking is: Microsoft, ServiceNow, Salesforce, Palantir, Oracle, Atlassian, HubSpot, Workday, UiPath, Zoom. This ranking does not mean "most worth buying," but "the clearest Agent payoff path most worth continuous tracking." Looking only at valuation and expectation gap, I would put Microsoft/ServiceNow/Salesforce on the "higher-certainty, more manageable-expectations" side, Palantir on the "strongest payoff but hottest valuation" side, and HubSpot/UiPath on the "possible expectation gap but accompanied by model disruption" side.

For the corresponding reverse score of Agent-disruption risk, I would put the highest-risk targets on companies and service providers that have "no system of record, no permission control, per-seat billing, and functions that an Agent can replace directly"; while companies with process orchestration, identity systems, organizational graphs, and ticket/CRM/ERP/collaboration entry points—even if pressured in the near term—are more likely to transition successfully.

Profiles of Key Listed Companies

Microsoft: The strongest point is not just model distribution but embedding Agents into Microsoft 365, Graph, Entra, and Copilot Studio, holding the work scenario, permissions, and office entry point together. The company's official Q3 FY26 disclosure shows an AI-business annualized revenue run rate above 37 billion dollars, with commercial cRPO reaching 627 billion dollars; Copilot Studio supports instructions, triggers, knowledge, tools, and embedded deployment, and the business model keeps the $30/seat while introducing per-message metered automation usage. The biggest advantage is installed base and channel; the biggest risk is that the valuation already reflects AI quite fully, and free/low-cost tiers may suppress the pace of paid conversion.

Salesforce: This is the "enterprise Agent usage commercialization template" most worth watching today. Agentforce is no longer a vague keynote concept but a fully billable model combining conversations, actions, credits, and seats; the public market has already seen its cumulative deal count and Data Cloud+AI ARR move toward substantive scale. Its core moat is not the model but CRM data, Flow, Slack, MuleSoft, Data Cloud, industry clouds, and the customer-service entry point. The risk: the market still questions whether Agentforce's increment this year can pull overall growth fast enough, while Salesforce itself is buying model tokens heavily, so inference cost and product gross margin need continued validation.

ServiceNow: If the Agent era is likened to "software moving from a database to an execution system," ServiceNow is one of the most natural beneficiaries, because it is already the brain of tickets, approvals, processes, and cross-department workflows. AI Agents/Autonomous Workforce explicitly stress role, business context, and permissions; the company has publicly said in 2026 that AI contract value could top 1.5 billion dollars, with Q1 2026 large customers whose Now Assist annual ACV exceeds 1 million dollars up more than 130% year over year, and RPO continuing to expand. The biggest advantage is the workflow moat and enterprise-grade governance position; the biggest risk is the market's worry that AI may ultimately erode the necessity of some traditional workflow seats.

Palantir: AIP is one of the listed-software companies closest to "Agents actually executing in complex real systems." In Q1 2026, Palantir's total revenue grew 85% year over year and U.S. commercial revenue grew 133% year over year, and management describes AIP as a "no-slop zone" where "every agent action is governed, attributed and auditable." This phrasing matters, because it captures exactly the core of Agent commercialization—reliable execution rather than flashy demos. Its moat lies in high-complexity system integration, frontline engineering teams, and government/industrial scenarios; the risk is that the valuation already implies extremely high sustained high-growth expectations.

Oracle: If the discussion is only about "enterprise Agent platforms," Oracle is not the sexiest name; but if the discussion is "who will see meaningful revenue gains from enterprise Agent inference volume, database integration, industry SaaS, and AI cloud contracts," Oracle is well worth studying. Reuters disclosed that its FY26 Q3 RPO grew 325% year over year to 553 billion dollars, with growth driven mainly by large-scale AI contracts; meanwhile, Oracle officially keeps stressing the value of AI agents integrating with high-quality enterprise data and ERP/HCM systems. Its benefit path leans more toward "underlying infra + database + app suite" than the most front-of-house Agent UX.

Atlassian: Rovo's strategic value exceeds its near-term revenue scale. Its most critical asset is the Teamwork Graph, which links collaboration, projects, meetings, documents, code, and goal context, then extends that context to external tools through broad connectors and MCP; Rovo Dev is its Agent product with the best chance to charge separately and lift ARPU. For investors, Atlassian looks more like a "platform-type long-term beneficiary" than a pure Agent stock with a near-term earnings burst.

HubSpot: Breeze proves HubSpot is not just turning AI into a copywriting button but trying to turn customer service, sales prospecting, company research, and data Q&A into executable agent teammates, while introducing a marketplace and credits. The problem is exactly here: HubSpot's seat-based growth model may be backfired by AI efficiency. The market has begun to scrutinize it with greater skepticism, and Barron's has explicitly noted investor worry that AI may reduce human users and thereby weaken the company's traditional seat growth. It is a textbook "Agent-enhanced company" and at the same time a "sample of Agents disrupting their own legacy model."

Workday: Workday may be the most forward-looking in product definition; it does not just build Agents but proposes the Agent System of Record, trying to manage employees and digital employees in a single system, with analytics, ROI, and skill-level controls. If this set is adopted by large enterprises, Workday would evolve from an HR/Finance system-of-record into a digital-labor console. The problem is that current public financial disclosure still lacks clear enough evidence of standalone Agent revenue, so it fits the priority-tracking list better than being placed directly among the already-realized.

Zoom: Zoom's AI Companion path shows that not every company's best strategy is "new SKU + high-priced add-on." It embeds agentic features into meetings, tasks, documents, and video workflows, and keeps stabilizing suite value through "default enhancement," then uses Custom AI Companion for modest price increases. This model helps lift retention and curb product commoditization, but in the near term it leans defensive rather than explosive.

UiPath: UiPath is the most typical "double-edged sword." On one hand, it has automation, orchestration, robotic execution, and process assets, and looks like it ought to be an "RPA 2.0" beneficiary; on the other hand, the market most worries that LLM Agents will eat the script-type value of traditional RPA. Barron's reaction to the slower FY2027 guidance and market worry about AI competition show that UiPath must prove it is not just "old RPA plus some GenAI" but truly becomes an Agent orchestration / governance platform. It is worth tracking, but the risk is markedly higher than Microsoft / Salesforce / ServiceNow.

Google / Alphabet: Gemini Enterprise's product definition is very clear—it is not a single assistant but a fusion of intranet search, AI assistant, and agentic platform, with permissions-aware access, prebuilt connectors, Agent Gallery, SSO, and user-level access control. The problem is that, for Alphabet, the Agent's gains are diluted by the enormous scale of advertising, cloud, and search, so it fits being viewed as "strategically important but with impure investment beta."

Amazon: Amazon Bedrock Agents host prompt engineering, memory, monitoring, encryption, user permissions, and API invocation in one place, and the product direction is very right; but on stock mapping this looks more like part of AWS's AI infrastructure and platform capability than a separately valuable Agent business. For investors, AWS's benefit is clear, while the Agent business's attribution is not.

Reserve watch list: To keep expanding to more listed software companies, I would prioritize putting SAP, GitLab, Pegasystems, NICE, Five9, Datadog, CrowdStrike, Okta on the next-round tracking list; not because they will all become core winners, but because they sit at the key Agent-era nodes of ERP/development/process/customer-service/observability/security/identity respectively. But this round's validation of these companies' "Agent revenue landing" is insufficient, so it is unwise to give overly firm financial judgments on them in this report.

Risks, Misconceptions, and Follow-up Research

The biggest misconception is making Agent research about "who shipped the most features." Only three things truly matter: first, whether the Agent enters production; second, whether the Agent enters billing; third, whether the Agent changes the customer's budget structure. Gartner has publicly warned that many agentic projects may be cut, and MIT's Agent Index flags low industry transparency; while benchmarks like ServiceNow WoW, Salesforce SCUBA, and APEX-Agents further show that a vast reliability gap separates public demos from enterprise production.

On a risk breakdown, of the eighteen most critical risks, six deserve investors' ongoing attention: adoption below expectations, paid-conversion rate below expectations, permission/data leakage, security incidents, inference cost compressing gross margin, and platforms squeezing the space for standalone software. Public platform docs have already written these into the product structure: OpenAI stresses human review and guardrails, Anthropic gives strict tool use and sandboxed computer use, Google stresses access control and user-level access, Workday stresses human oversight and analytics, ServiceNow stresses the AI Control Tower, and Salesforce uses the Digital Wallet to visualize usage. The market is giving a higher quality premium to companies that can "turn risk into tooling."

The final research conclusion can be compressed into ten sentences.

First, the Agent's place in the value chain is not "another AI application," but the intermediate operating layer that turns base models into executable software labor.

Second, the five sub-tracks most worth watching are: the enterprise Agent platform, the governance/permission/observability layer, customer-service and ITSM Agents, AI coding agents, and the connector/MCP/iPaaS layer.

Third, the listed companies most worth deep research are, in priority order: Microsoft, Salesforce, ServiceNow, Palantir, Oracle, Atlassian, HubSpot, Workday, UiPath, Zoom. The first four have "the strongest payoff evidence," the middle three have "excellent platform positions but still need financial-breakout validation," and the last two are "having their models rebuilt, so the expectation gap is larger."

Fourth, the most worthwhile private-market directions to track should no longer be "build yet another general-purpose Agent shell," but AI coding, AI customer service, Agent governance/observability, enterprise connectors/MCP, document intelligence / memory, and vertical legal/medical/financial Agents. Much of this layer's revenue is still not fully public and must be screened by customer lists, payment models, and usage depth, not by demo hype.

Fifth, the point the market most easily misreads is treating the Agent as a "model upgrade"; more precisely, the Agent is a software-business-model upgrade. Revenue attribution will migrate from seat toward seat+usage+outcome.

Sixth, the metrics most worth tracking over the next 6–12 months are not MAU, but: standalone Agent revenue, credits consumption, conversation/action counts, RPO/cRPO, the number of large AI customers, NRR, inference cost, automated-resolution rate, human-transfer rate, and the number of security incidents. These are what determine whether an Agent can move from product launch to profit realization.

Seventh, Agent-enhanced software companies are those that embed Agents into existing products to raise retention, expansion, and differentiation—for example Atlassian, HubSpot, Zoom, and Workday.

Eighth, platform-type Agent companies are those that truly hold enterprise identity, permissions, data entry points, and workflows—for example Microsoft, Salesforce, ServiceNow, Google, and Workday.

Ninth, Agent-native challengers will appear mainly in code, customer service, research, legal, and vertical professional processes; they will grab budget first rather than seats. The listed-market mapping for this layer is still scarce, and primary-market opportunities clearly outnumber secondary-market ones.

Tenth, the traditional software or service companies most likely to be impacted are not all legacy software companies, but the point-solution tools and low-end outsourcing services with no data barrier, no permission barrier, no workflow barrier, and primarily per-seat billing.

Open questions and limitations: This report covers the North American companies with the fullest public disclosure most deeply; a large number of companies in China, Europe, Japan, Korea, and India currently lack clear enough public data on Agent revenue breakdown, so they fit better in a next round of dedicated research in local language and local-capital-market terms. Another limitation is that many companies do not disclose, item by item, Agent-related ARR, NRR, RPO/cRPO, precise valuation multiples, and model cost in public materials, so for these fields I have deliberately avoided filling beyond public evidence.

A narrower follow-up direction: If only one is to be pursued further, I suggest prioritizing enterprise Agent governance and permission control. The reason: this layer is the easiest to underestimate, the hardest for open source to fully replace, the best at spanning every model and every Agent platform, and the most directly tied to real budgets right before enterprises go live. The second dedicated direction is AI customer-service Agents, because it is the frontline application where real billing and process replacement are seen first.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.