AI Agents: A Deep Dive into the Enterprise Execution Layer, Governance, and Monetization

Core Conclusions

• The place AI Agents occupy in the value chain is not "a model that chats better," but the assembly of base models, enterprise data, tool invocation, permission systems, approval flows, monitoring, and auditing into a deliverable "system of action" execution layer. The durable profit pool does not sit in single-shot inference itself; it sits with whoever controls the workflow entry point, permission management, enterprise data context, and the connector network.

• The fundamental difference between Agents and Copilots, chatbots, and RPA is this: Copilots mainly raise human productivity, bots/RPA mainly execute business along fixed scripts, while an Agent can autonomously plan under goal constraints, call tools across multiple steps, retain state, handle exceptions, and enter a human-in-the-loop when needed. It moves from "answering" to "acting."

• Enterprises do not just buy model APIs directly, and the reason is no mystery: a model API solves "can it reason," whereas what enterprises actually need to solve is "can it connect to systems within permissions, call tools, audit behavior, control costs, trace failures, satisfy compliance, and keep improving." That is why MCP, connectors, identity, governance, observability, and the workflow layer have rapidly become table stakes.

• The scenarios showing commercial validation first are not "big, do-everything general Agents," but settings with clear workflows and easily quantified ROI: AI coding, customer service / contact centers, CRM sales processes, ITSM / enterprise service desks, knowledge retrieval and enterprise search, plus HR / finance / legal subprocesses with strong rule boundaries.

• As of now, the public companies with the "clearest Agent revenue on the ground" in public materials are few, and the group with the strongest evidence is Salesforce, Microsoft, NICE, HubSpot, Palantir, UiPath, and Twilio. Salesforce has disclosed that Agentforce ARR has reached 800 million USD with 169% year-over-year growth and 29,000 cumulative Agentforce deals closed. Microsoft has disclosed that Microsoft 365 Copilot paid seats now exceed 20 million, with ARPU growth once again driven by Copilot and E5. NICE has disclosed AI ARR growing 66% year over year. HubSpot has shifted its Customer Agent and Prospecting Agent to billing by "completed outcome." Palantir's Q1 2026 revenue grew 85% year over year, with US commercial revenue up 133%. UiPath ARR reached 1.853 billion USD. Twilio's Q1 2026 revenue grew 20% year over year, and its earnings materials explicitly cite AI Agents for customer service and sales processes.

• Another group of large companies looks more like "defensive Agent beneficiaries" than near-term new growth engines: Oracle, SAP, Workday, and parts of Atlassian. Their greatest value is not selling Agents standalone at high ARPU, but using Agents to reinforce their existing systems of record—lifting retention, expansion, cloud backlog / RPO, and platform stickiness. Oracle has made AI Agent Studio available to Fusion Applications customers "at no additional charge." SAP's Joule already spans 35 solutions but has not separately disclosed Agent revenue. Workday is advancing around an "Agent System of Record" and role-based Agents, emphasizing managing an entire agent fleet on the customer's behalf.

• On the moat, the layer most likely to form a durable barrier is not the model layer but "permission-aware data access + enterprise workflow entry point + cross-system connectors + audit and governance + a distribution ecosystem." The model layer, under multi-model competition, open source, and falling prices, more easily trends toward intensifying competition. What is genuinely hard to displace is an enterprise's organizational structure, approval logic, historical process state, and security policy.

• Agents will significantly reshape software pricing models, but not as an "overnight, complete shift from seat to usage." The more realistic path is: a seat + credits hybrid model appears first, then expands in high-ROI scenarios toward charging by task, by outcome, by automated process, and by Agent workload. HubSpot already charges by resolution and recommended lead. Intercom Fin uses each resolution as the billing unit. Copilot Studio bills in Copilot Credits. Atlassian Rovo Dev uses seat + credits + overage billing. Zoom uses a bundled-free + premium-feature surcharge model.

• The segments with the highest revenue elasticity look strongest today in: enterprise Agent platforms, customer-service Agents, coding Agents, Agent security / identity governance, the iPaaS / MCP / connector layer, and enterprise search / knowledge Agents. They either directly replace labor / outsourcing costs, or lift existing SaaS ARPU and expansion rates, or become the mandatory control point every Agent must pass through before execution.

• The companies with the best margins are not necessarily the "sexiest" Agent companies. Better gross-margin / margin prospects usually belong to SaaS platforms that already have distribution channels and ready-made workflows, identity and security control layers, and high-value vertical-industry Agents. The most exposed to inference-cost margin erosion are usually general Agents that are heavily real-time interactive, low priced, high frequency in calls, and lacking workflow pricing power. Microsoft's CFO has stated plainly that the business model for AI applications will increasingly take the form of usage / consumption, while the company hedges inference cost through its hardware stack and software-efficiency optimization.

• The segments most prone to valuation bubbles are the Agent-native hot narratives: coding Agents, the general enterprise agent layer, and high-valuation private model companies that lack clear revenue disclosure. Cursor's parent Anysphere reached a valuation near 29.3 billion USD after its November 2025 round. Glean, in its early-2026 official disclosure, had ARR above 200 million USD at a 7.2 billion USD valuation. Harvey's 2026 round reached an 11 billion USD valuation. The valuation expectations for OpenAI and Anthropic run far above traditional software multiples. In public markets, Palantir's current trailing PE already exceeds 150x.

• The companies most clearly disrupted by Agents are not all SaaS, but "weak-workflow, weak-permission, weak-data-moat, weak-distribution" seat-based tool software, along with low-value outsourcing, traditional script-based RPA, basic customer-service seats, and parts of low-end development outsourcing. Conversely, companies that own the system of record have a chance to upgrade from a "record system" to an "execution system."

• The most important catalysts over the next 12–24 months are not "another Agent launch," but five hard metrics: separately disclosed AI / Agent ARR, acceleration in RPO / cRPO, ARPU uplift from usage-based pricing, customer conversion rates from pilot to production, and the impact of security / governance incidents on purchasing decisions. Gartner has warned that by 2028, 25% of enterprise generative-AI applications will experience at least 5 minor security incidents per year. The Infosys-HFS 2026 survey shows only 14% of enterprises have truly scaled Agentic AI into production.

• The biggest risk is not that the technology is "completely unusable," but that enterprise adoption lags narrative expansion: governance, identity, data sovereignty, prompt injection, tool misuse, auditing, and ROI proof will more often become the procurement bottleneck than model capability itself. McKinsey's 2026 AI Trust Maturity Survey likewise notes that enterprises still have clear gaps in governance, risk management, and trust maturity.

Value-Chain Landscape and Technical Architecture

From an investment standpoint, the AI Agent value chain splits into a "capability supply layer" and an "execution control layer." The former includes models, APIs, SDKs, and open-source frameworks; the latter includes connectors, permissions, data retrieval, memory, workflow, governance, security, monitoring, and the final business application entry point. The former more easily faces intensifying competition; the latter more easily forms sustainable customer lock-in and a profit barrier.

In enterprise Agent architecture, what deserves buy-side analysts' attention most is not "whose model is best," but the following layers. First, the permission and identity layer, because once an Agent starts to "execute," permissions become the product. Second, the workflow and system-of-record entry point, because this decides whether the loop can close. Third, the connector and data layer, because this decides whether the Agent is truly embedded in the enterprise's real systems. Fourth, governance / security / observability, because this decides whether a pilot can reach production. By contrast, pure framework-layer and pure observability-layer products are very good, but without a control plane they easily fall into "great product, hard to monetize." WhyLabs has announced it is ceasing operations, a cautionary counterexample worth noting.

Business Model and Scenario Analysis

Judging from public products and pricing, the Agent business model has already begun departing from traditional seat-only SaaS. The clearest trend is a "seat + credits + outcome" hybrid model: enterprises first buy a platform seat or main product, then pay for tasks actually completed, autonomous run counts, inference credits, resolutions, recommended leads, developer credits, and so on. This model maps more closely to customer ROI than pure token billing, and has a better chance of lifting ARPU, NRR, and retention.

On financial transmission, the Agent impact on SaaS typically comes in three stages. The first stage is "defensive bundling," showing up mainly in win rate and retention rather than new revenue. The second stage is "hybrid monetization," where usage/credits begin to drive ARPU. The third stage is when "system of action" takes hold, meaning enterprises migrate part of their manual-process budget from seats, outsourcing, and person-days to the Agent platform. Salesforce, HubSpot, and NICE have entered the second stage; Microsoft has clearly entered the second stage in personal and developer productivity scenarios; Oracle, SAP, and Workday lean more toward the transition from the first stage to the second.

Inference cost will still compress gross margin, but leading platforms are already hedging along three lines. First, finer-grained routing, pushing simple tasks down to cheaper models. Second, reducing context overhead—using caching, tool search, programmatic tool calling, permission-aware retrieval, and context management to cut token waste. Third, control-plane products (governance, identity, security, workflow) capturing more of total contract value. OpenAI's pricing page already distinguishes input, cached-input, and output costs. Anthropic has launched a Tool Search Tool and Programmatic Tool Calling to reduce context consumption. Google offers simulation, scoring, and optimization in its Agent Platform. Microsoft management has stated clearly that the usage-based business model plus hardware / software efficiency gains will help sustain better AI gross margin over the long term.

Below is a 24-month scenario analysis grounded in current public evidence. It is not an industry fact but a research framework: a way to identify which companies most need "adoption acceleration" to support their current valuation, and which can benefit even with mild adoption.

Across these three scenarios, the key watershed is not model capability but two variables. First, whether enterprises are willing to connect Agents to systems that are "writable." Second, whether vendors can move pricing from "AI features given away free" to "charging by completed business outcome." Only when the second variable holds will Agents truly change the SaaS business model.

Segment Breakdown and Competitive Landscape

Below, the focus is on the segments with the most investment value today, rather than every Agent direction that is conceptually plausible. Scores are capped at 5 and weigh commercialization evidence, moat, revenue elasticity, and competitive intensity together.

On the competitive landscape, Microsoft, Salesforce, and ServiceNow are currently the three strongest in the enterprise Agent platform layer. Microsoft has the desktop entry, the collaboration entry, and the developer entry, and has already moved Copilot from seat to credits. Salesforce is the first large SaaS company to articulate Agent ARR clearly in public financial terms. ServiceNow's edge is its cross-department process and governance control plane. Google is no weaker in tech stack and Agent Platform completeness—even strong in IAM agent identity, simulation, and governed connectivity—but its control over the office / CRM / ITSM main entry points still trails Microsoft, Salesforce, and ServiceNow. OpenAI and Anthropic have extremely strong influence over models, APIs, MCP, and developer standards, but without deeper workflow and permission entry points, their long-term profit pool more likely stays at "capability supplier" rather than "enterprise operating system." Oracle, SAP, and Workday derive their win rate from owning the system of record: they may not be the first to see visible incremental revenue, but once customers connect Agents directly into ERP/HCM/finance record systems, their stickiness strengthens markedly.

Public-Company Watchlist and Investment Tiering

The table below prioritizes coverage of public companies "most likely to show trackable financial validation over the next 12–24 months." The classification rules: Tier A are core direct beneficiaries; Tier B are clear beneficiaries with valuation, model-cost, or competitive risk; Tier C lean defensive; Tier D have narrative stronger than financial validation; Tier E are potential disruption targets.

Scoring Model and Priority

The total scores below are a researcher-model score using user-given weights: Agent revenue direct exposure 25%, product/data/workflow moat 20%, customer quality and revenue certainty 15%, platform ecosystem and connector capability 15%, financial quality and margin 10%, valuation reasonableness 10%, future catalysts 5%. This is not a market-price forecast but a research-priority ranking.

Global Watchlist

Beyond the focus list above, the following companies are suggested for a secondary research pool. They are either important indirect beneficiaries or geographically notable potential platform-type assets.

Private Companies and Disrupted Industries

Important Private Companies

The table below keeps only the non-public companies "most worth continuing to track." The standard is not fame but whether they hold potential platform power, whether real ARR / paid validation has appeared, and whether they can capture budget.

Traditional Software and Services Disrupted by Agents

What Agents disrupt first is not all software, but products and services whose value proposition has always rested on "clicking, copying, transcribing, answering repetitive questions, and doing weak analysis on a person's behalf inside an interface." The budget-migration path is roughly: low-end service seats / BPO → AI support agent; fixed-rule RPA → agentic orchestration; low-end development outsourcing → coding agents + internal platform teams; weak-workflow seat tools → system-of-action platforms.

This means that the system of record will not necessarily be replaced, but many front-end UIs will be rewritten by Agents. If a SaaS vendor owns the core records, permissions, and approval flow, it can upgrade from a back-end record system into an execution system; if it has only a thin interface and a weak data moat, it is more likely to be absorbed by a platform-type Agent. Workday launched its Agent System of Record precisely because it is trying to keep this trend in its own hands.

Risks, Open Questions, and Final Conclusion

The biggest uncertainty for enterprise Agents does not lie in "whether the model is smart enough," but in "whether enterprises are willing to hand it real permissions and real processes." This brings four core risks. First, adoption below expectations: Infosys-HFS data shows only 14% of enterprises have truly deployed agentic AI at scale, meaning the market is still far from broad production. Second, security and governance incidents: Gartner has predicted that by 2028, 25% of enterprise GenAI applications will face at least 5 minor security incidents per year; the fact that Palo Alto, Okta, Cloudflare, and CrowdStrike are all treating Agent security as a new control point shows precisely that this problem will not vanish on its own. Third, model cost and supply changes: if enterprises keep directing budget toward infrastructure and models rather than the application layer, many software companies will face the awkward position of AI capability with "usage but no profit." Fourth, open source and standards: the spread of MCP and open-source observability/frameworks will raise customer optionality and erode the expected premium of certain closed platforms.

Final Judgment

The most important significance of AI Agents within the AI value chain is not creating another "smarter chat box," but pushing enterprise software from the system-of-record era toward the system-of-action era. The long-term profit pool will therefore most likely land in five places: enterprise Agent platforms, the workflow control plane, connectors/MCP/iPaaS, identity and governance, and security and runtime protection. The model layer matters, of course, but it looks more like the upstream compute and capability layer; the application layer can only make real money on the premise of controlling the enterprise's real work.

The five segments most worth attention, I would converge to: enterprise Agent platforms, customer-service / contact-center Agents, AI coding Agents, Agent identity governance/security, and connectors/MCP/iPaaS. These five either already have a clear billing model, or hold the future control plane, or carry the most obvious seat/budget-migration logic.

The ten public companies most worth further research are: Microsoft, Salesforce, ServiceNow, NICE, HubSpot, Palantir, UiPath, Pegasystems, Appian, SAP. Among them, Microsoft, Salesforce, and ServiceNow hold the platform core; NICE and HubSpot are the application layer with the clearest pricing-model change; Palantir is high elasticity but high valuation; UiPath, Pega, and Appian are the group with the most expectation gap in the "process/automation transformation" direction; SAP represents a high-quality defensive beneficiary underestimated by the market.

The ten private companies most worth tracking are: OpenAI, Anthropic, Glean, Cursor/Anysphere, Harvey, Decagon, Sierra, Abridge, LangChain, Workato. If choosing a single narrower thread, I would prioritize AI customer-service Agents and the Agent governance/identity layer: the former commercializes fastest, and the latter is most likely to become a production must-have and form an independent profit pool.

The five points most easily misread by the market are: First, launching an Agent feature ≠ Agent revenue on the ground; Second, strong model capability ≠ the enterprise platform will win; Third, seats may not disappear immediately, but Agents will first lift ARPU and then compress low-end seats; Fourth, security, identity, and audit are not accessories but a procurement threshold; Fifth, open-source frameworks will compress mid-layer pricing, yet they raise the value of control-plane and system-of-record vendors.

The metrics most worth tracking over the next 6–12 months are not "who held another launch event," but: whether AI/Agent ARR is disclosed separately, whether cRPO/RPO accelerates because of Agents, whether the seat + credits structure significantly lifts ARPU, whether the production-customer count grows, whether security/governance incidents change the procurement path, and whether falling model cost improves gross margin.

Suggested narrower follow-on research directions, two as priorities: First, "AI customer-service Agents and contact-center transformation," because the business model has already migrated from human seats toward billing by outcome / by AI ARR, with the clearest financial validation. Second, "Agent governance, identity, and security," because it decides whether enterprises can turn a PoC into a production system, and has the most potential to become an independent new control-layer market.

Open Questions and Limitations

This report prioritizes public materials available as of May 18, 2026, but three categories of limitation must be clearly noted. First, many companies have launched Agent products but have not separately disclosed Agent revenue, ARR, NRR, RPO, or gross margin, so "feature launch" and "financial benefit" must be strictly distinguished. Second, some non-US-listed and private companies disclose only products, funding, or cases, with insufficient revenue and valuation transparency, marked as "undisclosed" or "needs further validation" in the tables. Third, some private model companies and Agent-native companies currently at extremely high valuations have public figures drawn mostly from funding news rather than audited financials, so they are suited as a sector thermometer rather than a basis for high-confidence earnings forecasts.

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