AI Application Layer, Copilot, Agents, and Vertical SaaS: A Deep Dive

Core Conclusions

• The AI application layer has evolved from "a front-end interface for calling models" into "a system layer that embeds models into high-frequency workflows, carrying data permissions and executable actions." What truly generates standalone revenue is not one-off generation capability, but applications that get written into business processes, own a permissions system, and can write back to the system of record. McKinsey's 2025 survey shows that workflow redesign is the single most important factor in whether enterprises see EBIT impact from generative AI; Microsoft's Work Trend Index likewise shows enterprises shifting from "using AI to assist tasks" toward "managing a set of agents to complete tasks."

• Public data shows that the tracks that reach revenue first are not general office work, but AI coding, AI customer service, legal and professional information, design and content production, enterprise knowledge search, and parts of medical documentation and data applications. These tracks already show clear ARR, ACV, paying customers, resolution/automation rates, or large-account contract evidence.

• There are three key paths by which the AI application layer goes from "feature enhancement" to "standalone revenue source": first, becoming a new paid SKU or add-on seat; second, introducing a new billing layer charged per task, per outcome, or per credits/agent workload; third, when AI revenue is not disclosed separately, showing it through ARPU, ACV, RPO, RPO/cRPO, NRR, gross margin, or labor-efficiency gains. Microsoft 365 Copilot, Intercom Fin, Microsoft Copilot Studio, the OpenAI API, and the Anthropic API all demonstrate that billing is moving from pure seat toward a "seat + usage" hybrid model.

• The financial reality of general office AI: there is revenue, but for now it is more defensive growth than large-scale profit realization. Microsoft has disclosed that Microsoft 365 Copilot paid seats exceed 20 million, with ARPU growth once again driven by E5 and Copilot; but its commercial paid-seat base also already exceeds 450 million, which means the Copilot attach rate is still early. Google's Gemini Enterprise/Agentspace has also taken product and pricing shape, but Google has not yet separately disclosed Workspace AI's revenue contribution.

• AI coding is currently the clearest, most direct, and most elastic AI application revenue pool. GitHub Copilot reached 20 million users by mid-2025, and 90% of the Fortune 100 use GitHub Copilot; Cursor surpassed 2 billion dollars in annualized revenue by February 2026; OpenAI disclosed that Codex has grown more than 5x since the start of the year. The shared traits of this track are high frequency, strong ROI, high willingness to pay, and the ability to bind deeply into enterprise codebases, CI/CD, security, and testing workflows.

• AI customer service and call centers are the second track that has clearly proven a revenue model. NICE disclosed in Q1 2026 that AI ARR grew 66% year over year, with AI and self-service ARR reaching 345 million dollars; Five9 disclosed in Q1 2026 that AI revenue grew 68% year over year, with an annualized run rate above 125 million dollars; more than 7,000 teams use Intercom Fin, with the average resolution rate rising to 67% and outcome-based billing at 0.99 dollars per outcome. This track has moved beyond "assisting agents" to replacing manual ticket handling in some scenarios.

• Highly regulated verticals such as legal, tax, risk control, and medicine are among the best tracks for margins and moats in AI commercialization. Thomson Reuters noted in Q1 2026 that recurring revenue growth was driven by Westlaw, CoCounsel, Practical Law, and others; Reuters reported in the same period that CoCounsel users had reached 1 million and that the share of AI-related contract value had risen to 30%; RELX and Wolters Kluwer build moats through proprietary content libraries, citation systems, expert review, and compliance workflows, with Wolters Kluwer even disclosing that about 70% of its digital revenue already comes from AI-powered solutions.

• The software companies that truly benefit from AI usually share six traits at once: a system of record; permissions and governance; high-frequency workflows; proprietary/private data; high switching costs; and provable ROI. Microsoft, ServiceNow, Salesforce, Intuit, Oracle, Thomson Reuters, RELX, Wolters Kluwer, Tempus, and Veeva fit this framework better. By contrast, horizontal SaaS that offers only lightweight collaboration interfaces and lacks proprietary data or execution authority is more easily squeezed on pricing power by platform-style Copilots or AI-native tools.

• The way traditional SaaS raises prices and expands ARPU through AI is bifurcating noticeably. Microsoft and Oracle have explicitly cast AI as an ARPU driver; Atlassian disclosed that Rovo customers grow ARR roughly 2x faster than non-Rovo customers; but Asana's AI Studio reached only just over 6 million dollars in ARR by the end of FY2026, showing that for many collaboration-style SaaS, AI is still stuck at the stage of "shipping plenty of features, but with revenue scale still small."

• AI Agents are changing the software industry's business model, but will not immediately replace seat-based pricing entirely. What is more likely is a "three-layer billing" model: a base layer of seat/platform subscription serving as the gateway for identity, permissions, and governance; a middle layer of usage/credits billed by task volume or agent workload; and a top layer of outcome/priced results, charging by outcome in measurable scenarios such as customer service, contracts, and sales execution. Intercom, Microsoft Copilot Studio, the Google Gemini Enterprise Agent Platform, and the OpenAI API are already this kind of hybrid model.

• It is also becoming increasingly clear which companies have "a strong AI narrative but insufficient financial benefit." Typical traits include: many AI features, but no separate disclosure of revenue or ARR; high customer adoption, but unclear paid penetration; AI mainly bundled for free, more for retaining customers and defending against competitors; or AI revenue still extremely small relative to the core business. General office work, parts of collaboration/project management, and some "AI wrapper" applications mostly sit at this stage today. Asana's AI ARR scale is one of the clearest negative comparisons.

• What is most prone to bubble-like valuation is not "any company with AI," but "companies that grow fast, have a strong narrative, have proven a bit of revenue, but whose valuation already prices in distant success in full." The most typical right now are Palantir, Databricks, OpenAI, Anthropic, Cursor, Harvey, Sierra, and Decagon; some of these companies do have real revenue, but the implied long-term success rate the market assigns them is already extremely high.

• The most important catalysts over the next 12-24 months are not how many more agents get launched, but four "hard metrics": whether AI-related ARR/ACV continues to be disclosed; whether the AI attach rate moves from pilot to scale; whether RPO/cRPO and NRR are being driven by AI; and whether inference cost declines enough to improve gross margin rather than being offset by a usage surge. The biggest risks are: enterprise ROI falling short of expectations, AI features becoming free, model platforms moving down into the application layer, agent security incidents, and what Gartner warns of as "agent washing" and project cancellations.

Industry Chain Panorama

The AI industry chain is being rebuilt from the linear structure of "model - interface - application" into a closed loop of "model - orchestration/governance - workflow application - system of record - outcome delivery." The companies that occupy the center of the profit pool are usually not the layer closest to the model, but the layer best able to control enterprise permissions, data semantics, workflow nodes, and outcome acceptance. McKinsey's survey and Microsoft's enterprise-trend data point to the same thing: enterprises are no longer satisfied with "generating content," and are instead looking for AI that "redesigns processes, replaces steps, and validates outcomes."

Conclusive read on the industry chain: The profit pool will not stay long at the spot "closest to the model"; it is more likely to settle in the application and platform layers that best handle task orchestration, best understand industry data, and sit closest to the execution workflow. Model vendors earn "compute rent," but workflow-software vendors that hold the task entry point and the outcome loop can earn the more durable software premium.

Business Model and Scenario Analysis

The business model of AI applications has clearly departed from the single-seat model of traditional SaaS, but it will not fully replace seats in the near term. The more realistic conclusion is: seat handles identity, permissions, budget attribution, and governance; usage/credits handles covering variable compute cost; and outcome/task billing handles monetizing ROI directly. Microsoft Copilot Studio already uses Copilot Credits to meter an agent's actions and responses; Google's Gemini Enterprise Agent Platform bills by Agent Runtime; the OpenAI API charges simultaneously for tokens, file search tool calls, and containers/sessions; Anthropic has published token, prompt caching, and dual Team/Enterprise pricing; and Intercom defines customer-service automation directly as "0.99 dollars per outcome."

This means the key question raised by users -- "will AI Agents shift software from per-seat billing toward per-task, per-outcome, and per-agent-workload billing" -- has the answer: yes, but not as a replacement, as an overlay. Enterprises will still keep a seat base as the master contract and permissions foundation, because the CIO/CFO need budget certainty; but in high-usage scenarios, usage or outcome billing will become increasingly common, especially in customer service, development, contract processing, marketing generation, and enterprise agent orchestration. HubSpot's market messaging has also begun pushing AI credits and outcome-oriented pricing to the foreground; Salesforce has introduced Agentic Work Units as a new usage metric.

The gross margin of AI applications is not necessarily compressed permanently by model cost. What determines margin is the pace of inference-cost decline, model-routing capability, cache hit rate, batching, and workflow value density. OpenAI's published Batch API can save 50% on input/output cost; Anthropic has disclosed differential pricing for prompt caching; and Salesforce's published production-deployment research shows that with good inference architecture and autoscaling design for a compound AI system, tail latency can drop by more than 50%, throughput can rise by up to 3.9x, and cost can be cut by 30%-40%. This shows that high-quality application vendors are not merely "passively bearing model cost"; they are turning inference optimization itself into a competitive advantage.

AI will also change retention and ARPU, but only on the condition that it is embedded in a verifiable workflow. Microsoft has explicitly said M365 commercial-cloud ARPU is once again driven by E5 and Copilot; Atlassian disclosed that Rovo customers grow ARR roughly 2x faster than non-Rovo customers; while Asana, after making AI Studio a standalone product, still saw AI ARR only just exceed 6 million dollars by the end of FY2026, proving that not every "SaaS with AI added" can capture meaningful revenue leverage.

The table below answers the question "which application scenarios can already convert into revenue":

Scenario Analysis

The essential difference among these three scenarios lies not in "whether the model is good," but in whether enterprises are willing to redesign processes, whether there is an acceptable unit of workload, and whether they are willing to reallocate spending from existing software budgets. This is also why Gartner warns that a large share of agentic-AI projects may be canceled before 2027: without a process loop, permissions governance, and value validation, agents will only become demo products.

Deep Dive on Key Tracks

The table below compresses the 25 tracks listed by users into the dimensions "most critical to investment judgment": track logic, revenue conversion, commercialization stage, pricing, margin/cost, moat, catalysts, risks, and investment appeal. Scores reflect the author's overall judgment, out of 10.

Track priority ranking: If we look only at "revenue elasticity + profit elasticity + moat + verifiability over the next 12-24 months," I would put AI coding, AI customer service, AI legal, AI healthcare, and the enterprise AI Agent platform at the very front. If we look only at "near-term landing rather than long-range narrative," AI customer service and AI coding have the highest priority; if we look only at "high margin and long-term moat," legal/tax/professional information and healthcare are better.

Investment Targets Master Table and Tiering

The table below prioritizes representative names that already have clear revenue, ARR, ACV, RPO, customer cases, or product-pricing evidence; for companies that do not disclose AI revenue separately, I explicitly label them "defensive/indirect beneficiary" or "needs further verification."

Representative unlisted names are as follows:

Company Tiering and Investment Priority

Scoring Model and Total-Score Ranking

Scoring weights follow the user's suggestion: AI revenue direct exposure 25%, product/data/workflow moat 20%, customer quality and revenue certainty 15%, financial quality and margin 15%, growth elasticity 10%, valuation reasonableness 10%, catalysts 5%. Reverse risk model: risk of core function being replaced by the model 30%, seat pricing disrupted by Agents 20%, lack of moat 20%, budget squeezed by platform-style AI 15%, valuation too high 15%.

Key Listed and Unlisted Companies

Given length and the quality of public disclosure, the following prioritizes the 15 listed companies most worth entering the next round of deep-dive modeling and the 10 unlisted companies most worth tracking long-term. The conclusions still hold to two principles: distinguish "AI launch" from "AI landing," and distinguish "revenue certainty" from "valuation appeal."

The 10 unlisted companies most worth tracking long-term are: OpenAI, Anthropic, Cursor, Glean, Harvey, Sierra, Intercom, Synthesia, ElevenLabs, Abridge. Among them, Cursor, Glean, Intercom, ElevenLabs, and Abridge are closer to "AI application companies that can already be understood through an enterprise-software framework"; OpenAI and Anthropic are closer to foundation-model platforms; and Harvey, Sierra, and Synthesia are the names closest to the IPO-pipeline reservoir between the vertical and application layers.

Valuation Risk and Final Judgment

From the standpoint of market expectations, what is most worth distinguishing right now is three types of company.

The first type is "companies whose AI revenue is genuinely growing, but whose valuation already fully reflects expectations." Typical examples include Palantir, Databricks, Cursor, OpenAI, Anthropic, Sierra, and Harvey. They are not without fundamentals; on the contrary, the revenue growth of many of them is real. The problem is that valuation has already capitalized years of high growth, strong retention, and potential platformization in advance. As long as production adoption, gross margin, or contract cycles come in slightly below expectations, a re-rating could be triggered.

The second type is "companies whose AI revenue is genuinely growing, but which the market is not valuing on a pure AI story." I place more weight on Thomson Reuters, Wolters Kluwer, NICE, Docusign, Intuit, and parts of Oracle. Their common trait is that AI is more about reinforcing an existing high-quality recurring-revenue machine than betting on a brand-new story; the market therefore usually does not assign them extreme AI multiples, yet their AI revenue and profit realization are actually more verifiable.

The third type is "companies with a strong AI narrative but insufficient financial validation." These companies typically have plenty of AI launches but lack public evidence of AI-related ARR, ACV, RPO, customer expansion, or margin improvement. Parts of horizontal collaboration, general office add-ons, light knowledge management, and low-code wrappers all sit here. Asana is one of the clearest samples: AI Studio is indeed growing fast, but as of the end of FY2026 ARR was still only over 6 million dollars, still small relative to its core revenue scale.

On "which traditional software companies could be disrupted by AI rather than enhanced by it," I worry more about the following types than about any single company:

• Software with light workflow, a weak system of record, and weak proprietary data.

• Software that mainly sells an interface and information retrieval, rather than an execution loop and compliance liability.

• Software that charges per seat, but whose seat no longer corresponds to real workload.

• Software that users are willing to bypass directly to reach the Copilot/ChatGPT/enterprise-search layer.

By this standard, project management, light collaboration, parts of knowledge bases, parts of marketing-content tools, traditional BPO, and "utility software" without unique data assets carry greater risk; while ERP, fintax, legal, healthcare, core customer-service systems, development platforms, and security platforms carry relatively lower risk, because they more easily upgrade from "tool" to "work-execution platform." This is also why I lean toward understanding the future software industry as a migration from system of record to system of action, but the winners usually still need to first own record, permissions, governance, and data lineage.

Final Judgment

The importance of the AI application layer within the AI industry chain comes not from being closest to the user, but from having the best chance to turn model capability into enterprise-software revenue that is billable, governable, auditable, write-back-capable, and renewable. The five sub-tracks most worth watching going forward are: AI coding, AI customer service, legal/professional-information AI, medical-workflow AI, and the enterprise AI Agent platform. They have either already proven revenue elasticity or simultaneously possess a high moat and high long-term margins.

For the 10 listed companies most worth deep research, I would prioritize: Microsoft, ServiceNow, Salesforce, Intuit, Oracle, Thomson Reuters, Wolters Kluwer, NICE, GitLab, Tempus. For the 10 unlisted companies most worth tracking, I would prioritize: OpenAI, Anthropic, Cursor, Glean, Harvey, Sierra, Intercom, Synthesia, ElevenLabs, Abridge. These names cover seven directions where clear commercialization signals have already appeared: platform, coding, customer service, legal, healthcare, design/video, and voice.

The five points the market most easily misreads are: first, launching an AI feature does not equal landing AI revenue; second, high usage does not equal a high paid rate; third, free AI bundling is often defense, not growth; fourth, a decline in model cost does not automatically equal improved application-layer margin; the key lies in routing/caching/workflow value density; fifth, Agents will not kill all seats overnight, but they will first compress low-moat seats and amplify the usage-billing rights of high-moat platforms.

The metrics most worth tracking over the next 6-12 months are: AI-related ARR/ACV, paid seat/attach rate, RPO/cRPO, NRR, AI outcome/resolution rate, AWU/credits consumption, $100k/$1m large-account count, inference cost/gross margin, and AI's impact on sales and support labor efficiency. Only when these hard metrics keep improving will the market switch from an "AI narrative" to an "AI valuation anchor."

If I had to give a single, narrower direction for follow-up research, I think the one most worth drilling into is: the intersection of the enterprise AI Agent platform and AI customer service. The reason is direct: this is where the five most critical variables already appear together -- real revenue, outcome pricing, workflow control, customer stickiness, and the seat-to-outcome business-model migration; and it is most likely to be the first to decide which traditional SaaS names are "AI-enhanced software companies" and which get re-rated as "back-office systems" or eroded by AI-native challengers.

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