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.