Software Revaluation in the AI Agent Era: From Seat-Based Software to the Responsibility Runtime Layer

AI Agent Software & SaaS Trend-Focused Deep Research Report

1.1 | One-Page Decision Dashboard

1.1.1 | One-Sentence Thesis

AI agent will not simply destroy the software industry; instead, it will migrate software value from “interfaces operated by humans” to a deeper layer where systems understand state, execute actions, assume responsibility, attribute outcomes, and charge based on value.

This statement does not directly imply that “all strong SaaS companies will become stronger.” The companies that truly benefit must simultaneously have high-value state changes, context control, write authority, closed-loop governance, responsibility transfer, and budget migration capabilities. High historical gross margins, subscription revenue, legacy customers, multifunction UI, or large amounts of data alone are not enough to guarantee continued success in the AI agent era.

The core of this report is not to judge “which company has the most AI features,” but to judge:

Which company can push AI from the feature layer to the execution layer,
then from the execution layer to the responsibility layer and outcome layer,
and finally convert this control point into ARR, RPO, gross margin, FCF/share, and per-share value.

1.1.2 | Conclusion Map of This Report

1.1.3 | Current Company Ranking

This ranking is not a valuation conclusion, nor is it a buy rating. It only indicates each company’s industry position, control-point depth, and evidence strength in the AI agent era.

1.1.4 | Three Most Important Judgments

First, state change matters more than interface. The most valuable software in the future will not be the interface users open most often, but the software best able to change business state, write back to systems, complete actions, assume responsibility, and accept accountability. If a product can only generate, summarize, query, or recommend, it remains in the tool layer; only when it enters tasks, approvals, write-backs, audits, and rollbacks does it begin to approach the runtime layer.

Second, seats will not disappear, but they will be downgraded. Seats will still carry identity, permissions, governance, supervision, and responsibility ownership; true incremental revenue will come more from AI workloads, quotas, tasks, usage, outcomes, and enterprise bundled packages. Pricing in the agent era does not jump in one step from seat to token; instead, it forms a hybrid stack of base seats + usage quotas + advanced workflows + governance modules + outcome-based fees.

Third, declining token costs are not automatically positive. Lower model costs are a necessary condition for gross margin recovery, not a sufficient condition. Who can retain the benefit of cost declines in its own income statement depends on pricing power, competitive intensity, customer bargaining power, model routing, caching, batch processing, productization capability, and outcome attribution. Low-responsibility tools may pass cost-saving benefits to customers; high-responsibility runtime layers are more likely to retain pricing power.

1.1.5 | Three Biggest Disconfirming Signals

1.1.6 | Reading Sequence for This Report

This report proceeds along an investment research chain:

Core mechanism
→ Winners and losers
→ Positioning of the seven companies
→ AI economics
→ Evidence of real agent transformation
→ Biggest counterexamples and abstraction risk
→ FCF/share and quarterly tracker

When reading, do not treat each chapter as parallel material. The true sequence of this report is: first judge where software value is migrating, then judge which companies own the control points, and then audit whether those control points have entered commercialization, gross margin, and per-share value.

1.1.7 | Next-Quarter Tracking Table

1.2 | Glossary: Aligning Abbreviations and Definitions First

This article uses English explanations for core concepts wherever possible, but some abbreviations in software and investment research cannot be fully avoided. After this section, the article will use these abbreviations directly by default.

1.2.1 | Investment and Financial Definitions

1.2.2 | Software and AI agent Definitions

1.2.3 | Custom Frameworks Used in This Article

1.2.4 | Company and Product Abbreviations

2.1 | Core Mechanism: Software Value Is Migrating from UI to the Responsibility Operating Layer

2.1.1 | Why AI agent Is Not an Ordinary Feature Update

The core assumption of traditional SaaS is that humans remain the execution subject: people log into software, review data, make judgments, and take actions, while software is responsible for records, collaboration, reporting, and workflows. The goal of AI agent is to hand part of the judgment and execution actions to the system itself. Software is no longer merely a "tool operated by people"; it increasingly resembles an "operating layer authorized to complete tasks."

This will rewrite the most important control points for software companies. In the past, strong software typically controlled the interface, data objects, team collaboration, and organizational training. In the future, stronger software will control business state changes, next actions, permission boundaries, cross-system writebacks, audit records, accountability, and outcome attribution.

Therefore, the research question must shift from "Does the company have AI agent?" to:

What state was the customer originally in?
What state does the software move the customer into?
Is the cost of error in that state change high?
Does the software have context and write authority?
When something goes wrong, can it be audited, rolled back, and assigned accountability?
Is the customer willing to migrate budget for this outcome?

2.1.2 | L0-L6 Software Value Layers

The most dangerous investment misjudgment is treating L0 / L1 AI features as L4 / L5 / L6 control points. A product's ability to generate text, images, query results, or summaries does not mean it can write back to systems, assume responsibility, or capture outcome budgets.

2.1.3 | Why L4 / L5 / L6 Are More Valuable

The L4 workflow execution layer is more valuable because it controls business state changes. A ticket moving from unresolved to resolved, a customer case moving from open to closed, a finance task moving from pending to completed, or an employee request moving from submitted to approved are all measurable state changes.

The L5 responsibility governance layer is more valuable because once agent obtains action authority, what customers care about most is no longer "Can it answer?" but "Can it be authorized, constrained, audited, and rolled back?" Enterprises will not easily hand high-responsibility processes to software that lacks a governance layer.

The L6 outcome delivery layer is more valuable because it may move from software budget into labor budget, services budget, BPO budget, expert budget, risk budget, and revenue growth budget. What customers buy is no longer software access, but task completion, error reduction, lower risk, labor savings, or revenue creation.

2.1.4 | Where the Seven Companies Currently Sit in the Layers

2.2 | Who Wins, Who Loses

2.2.1 | Three Types of Winners

First type of winner: high-responsibility vertical operating systems.

Representative companies are Intuit, Veeva, and parts of Workday, NICE, and industry clouds. Their shared traits are high cost of error, specialized business language, slow customer replacement, data that can trigger actions, and relatively attributable outcomes. In these scenarios, AI is not an ordinary assistant, but the productization of expert capability, compliance workflows, financial operations, industry approvals, and risk control.

Intuit's tax, accounting, small-business accounting, payroll, payments, credit, marketing, and cash flow management are naturally close to outcome delivery. Veeva's life sciences processes are slower, but the responsibility is deeper. Clinical, quality, commercial compliance, regulatory materials, and industry CRM are not low-responsibility tools that customers are willing to replace lightly.

Second type of winner: cross-system workflow operating layers.

The representative company is ServiceNow. ServiceNow's advantage is that it naturally handles state transitions within the enterprise: requests, tickets, approvals, incidents, services, assets, risk, and security. For an agent to execute safely in the enterprise, it needs context, permissions, tasks, auditability, rollback, and governance. If ServiceNow can combine Now Assist, the AI control tower, autonomous workforce, and workflow platform, it has a chance to become the operating layer for enterprise agents.

Third category of winners: companies with strong data context and the ability to move up into execution.

Representative companies include Snowflake's potential path, Salesforce's potential path, and Adobe's enterprise content potential path. These companies all have assets, but they still need to prove whether those assets can move up the stack.

Snowflake has data context, governance, and a model ecosystem, but data context is not the same as the right to act. Salesforce has customer records, sales and service processes, Data 360, Slack, and enterprise distribution, but whether Agentforce becomes an execution layer remains to be validated. Adobe has creative assets, documents, brand assets, and Marketing Cloud, but it must prove that it can migrate from creative tools to the enterprise content supply chain.

2.2.2 | Three Types of Losers

First category of losers: low-responsibility single-point tools.

These tools only solve low-responsibility tasks such as generating copy, generating images, summarizing meetings, auto-filling forms, and simple customer-service bots. Without proprietary context, write access, auditability, responsibility, distribution, and outcome attribution, they will be rapidly repriced downward by improvements in model capabilities, built-in platform features, and open source.

Second category of losers: software that has data but no right to act.

Data matters, but data is not the endpoint. Enterprises will not pay indefinitely for "more data"; they pay for better judgment and action. Companies that remain only in storage, cleaning, reporting, query, and analysis, and cannot enter the chain of business actions, will be abstracted away by upper-layer agents and business systems.

Third category of losers: software that bears AI costs but cannot charge for them.

Some software will be forced to add AI features because customers expect them to exist. But if AI cannot be charged for separately, legacy seats are compressed by efficiency gains, inference costs rise, competitors give it away for free, and customers also see AI as a basic feature, then AI is not a growth engine but a source of margin pressure.

2.2.3 | Three Fates of Tool-Type Software

Tool-type software will not all disappear, but it will split into three paths.

Adobe is the most typical example. Low-end creative generation is being repriced downward, but if the enterprise content supply chain can connect brand assets, copyright, approvals, publishing, distribution, and performance measurement, it may move up from the tool layer to the operating layer.

3.1 | Positioning Map of the Seven Companies

3.1.1 | Evidence Level Definitions

3.1.2 | Write-Depth Definitions

3.1.3 | Company Evidence Summary Table

3.1.4 | Weighted Trend Score

Note: This table measures only trend exposure and control-point strength in the AI agent era. It is not company-disclosed data, not newly added external data, not a valuation conclusion, and not a buy rating. Each individual score is a qualitative 5-point score assigned by the author based on the evidence discussed above; the composite trend range is converted to a 0-100 range according to 20 × Σ(individual score × weight). Because this is not the output of a financial model, the final result is presented as a range rather than a precise integer, to avoid misreading trend judgment as hard data.

The purpose of this table is to rank research priorities, not to score stocks. ServiceNow, Intuit, and Veeva occupy the top three tiers because they are closer to state change, write authority, accountability governance, and budget migration; Salesforce, Adobe, Workday, and Snowflake still need to prove they can move upward from records, tools, data, or mature system layers into execution, governance, accountability, or outcomes layers. The ranges for Workday and Salesforce are close and do not represent a precise ranking difference.

3.1.5 | ServiceNow: Closest to the enterprise agent operating layer

Fact. The original report disclosed ServiceNow Q1 FY2026 subscription revenue of $3.671 billion, up +22% year over year; current remaining performance obligations of $12.64 billion, up +22.5% year over year; remaining performance obligations of $27.7 billion, up +25%; and customers with Now Assist annual contract value above $1 million up +130% year over year. Knowledge 2026 further advanced the company's positioning toward a combination of AI control tower, autonomous workforce, data intelligence, and security.

Inference. This evidence is stronger than an ordinary AI launch event because it moves in the same direction as large-customer ACV, RPO, cRPO, and subscription revenue. What ServiceNow originally controlled was the state of enterprise requests, tickets, approvals, incidents, services, risks, and assets; the AI agent simply pushes this state-transition logic toward automation and governance.

Assessment. ServiceNow is the sample among the seven companies that is closest to becoming a horizontal workflow operating-layer winner. Its core question is not whether it can build an AI assistant, but whether it can become the control tower and default governance entry point for enterprise agents.

Disproof. If Microsoft, cloud vendors, model providers, or customer-built platforms take away the agent orchestration and governance entry point, and ServiceNow becomes merely a workflow system being called by others, the assessment should be revised downward.

What to watch over the next four quarters. The number of large-ACV Now Assist customers, production adoption of the AI control tower, workflow expansion across IT / HR / CRM / Security / Risk, whether RPO / cRPO continue moving in the same direction, and whether AI costs erode gross margin.

3.1.6 | Intuit: Vertical accountability operating system

Fact. The original report disclosed Intuit Q2 FY2026 total revenue of $4.651 billion, up +17% year over year; Online Ecosystem +21%; QuickBooks Online Accounting +24%; GAAP operating income +44%; and non-GAAP operating income +23%. Management defined the strategy as combining AI with expert intelligence to provide autonomous experiences that complete tasks on behalf of customers, and launched AI-native ERP for mid-market businesses.

Inference. Intuit's tasks are naturally close to outcomes: tax filing, bookkeeping, payroll, collections, payments, cash-flow forecasting, credit, lending, marketing, and customer acquisition. These tasks have clear error costs and are relatively easy to tie to customer value. For Intuit, the AI agent is not a "chat assistant"; it has the opportunity to turn tax, accounting, and SMB operating tasks performed by human experts into software.

Assessment. Intuit is the vertical accountability operating-system sample most worth studying in the AI agent era. Its opportunity is not to sell more software interfaces, but to move from software budgets into expert services, financial operations, and outcome-delivery budgets.

Disproof. IRS Direct File or other free tax-filing solutions pressure consumer tax pricing power; AI automation cannibalizes assisted tax / expert service revenue; banks, Stripe, Square, and Payroll platforms take away the SMB operating entry point; the AI + expert cost structure deteriorates.

What to watch over the next four quarters. QBO ARPU, add-on purchases in payments / payroll / marketing / financing, IES standalone customer and ARR disclosure, the revenue and cost structure of TurboTax AI + expert, and whether expert services become softwareized rather than service-heavy.

3.1.7 | Veeva: High-accountability industry cloud, slow but deep

Fact. The original report disclosed Veeva FY2026 total revenue of $3.1953 billion, up +16% year over year; subscription revenue of $2.6842 billion, up +17%; and non-GAAP operating income of $1.4338 billion. The company disclosed that more than 125 Vault CRM customers are live, including 2 top 20 biopharmas in major markets; 10 top 20 biopharmas have made global commitments to Vault CRM, and about 14 top 20 are expected to commit.

Inference. Veeva's AI transition will not explode like consumer AI, but its responsibility depth is high. Life sciences customers face high costs for errors, long replacement cycles, and strong standardization pressure across clinical, quality, commercial compliance, document approval, and regulatory submission workflows. The Vault CRM migration is a harder control point, while AI is an enhancing variable.

Judgment. Veeva is a classic high-responsibility industry cloud: not fast, but deep. Its long-term value comes from industry semantics, compliance workflows, customer trust, and platform migration, not the number of AI features.

Falsification. Vault CRM migration slows, the pace of top biopharma live / committed weakens; AI remains stuck at the recommendation layer for a long time and cannot enter core compliance workflows; large pharma companies build in-house or competing platforms weaken Veeva's industry-semantic authority.

What to watch over the next four quarters. Vault CRM live count, top 20 biopharma migration pace, cross-cloud expansion across R&D / Quality / Commercial, whether AI enters the core approval and compliance workflows, and whether non-GAAP operating margin remains high.

3.1.8 | Adobe: Tool Layer Under Pressure, with an Option on the Enterprise Content Supply Chain

Fact. The original text disclosed Adobe Q1 FY2026 revenue of $6.4 billion, +12% year over year; AI-first annual recurring revenue more than tripled year over year; subscription revenue was +13%; and operating cash flow was $2.96 billion.

Inference. Adobe is neither simply an AI victim nor an inevitable winner. Consumer and low-end creative generation is being hit by models, Canva, open-source tools, and new entry points; professional tools still have barriers in formats, habits, plugins, and team collaboration; what truly deserves a higher valuation is the enterprise content supply chain: brand, copyright, compliance, approval, delivery, multi-channel adaptation, and performance measurement.

Judgment. Adobe's core question is not whether Firefly can generate better images, but whether Firefly, GenStudio, Experience Cloud, Acrobat, and brand asset governance can upgrade Adobe from a creative tools company into the operating layer for enterprise content production.

Falsification. AI-first ARR is mainly a reclassification of existing subscriptions; Creative seats are eroded by AI-native tools; Firefly has high costs but is hard to monetize; GenStudio fails to enter enterprise content production and marketing budgets.

What to watch over the next four quarters. The definition of AI-first ARR, GenStudio enterprise customers, whether the Firefly Services API enters production pipelines, Creative Cloud seats and ARPU, and whether Acrobat / Document Cloud strengthens high-responsibility document flows.

3.1.9 | Salesforce: Strong Distribution and Record Layer, Execution Layer Still to Be Proven

Fact. The original text disclosed Salesforce FY2026 remaining performance obligations of more than $72 billion and operating cash flow of about $15 billion, while also disclosing cumulative 2.4 billion agent work units and 19T processed tokens.

Inference. Salesforce has the customer record layer, enterprise distribution, Data 360, Slack, and industry clouds. Agentforce usage signals are valuable, but work units and token counts are not revenue quality. The key question is whether this usage enters production accounts, ARR, renewals, Data 360 add-on purchases, and reacceleration in the core clouds.

Judgment. Salesforce is strong distribution + a strong record layer + an execution operating layer still to be proven. It has an opportunity, but high Agentforce usage alone cannot directly prove a successful move up the stack.

Falsification. Agentforce is merely an embedded AI feature inside CRM; work units are high but ARR is weak; customers put the agent orchestration and governance layer in ServiceNow, Microsoft, or self-built platforms; Data 360 does not become the required context for agents.

What to watch over the next four quarters. Agentforce ARR, production accounts, Data 360 attach, Slack and industry-cloud closed loops, whether core organic growth rises again, and whether AI affects gross margin.

3.1.10 | Snowflake: Strong Data Context, but Insufficient Execution Rights

Fact. The original text disclosed Snowflake Q4 FY2026 product revenue of $1.227 billion, +30% year over year; net revenue retention of 125%; remaining performance obligations of $9.77 billion, +42% year over year; and FY2027 product revenue guidance of +27% year over year.

Inference. Enterprise agents need trusted data, and Snowflake is a beneficiary. But "agents need data" does not mean Snowflake captures agent economics. If execution and business write-back happen inside Salesforce, ServiceNow, Workday, Veeva, Microsoft, or self-built systems, Snowflake may simply be a high-quality underlying consumption layer.

Judgment. Snowflake's core opportunity is the governed semantic context and data agent operating layer; its core risk is being abstracted by upper-layer business systems into a data interface.

Falsification. AI query efficiency reduces consumption; Cortex / Intelligence usage is strong but does not generate paid workloads; Databricks, cloud vendors, or business systems take the semantic layer; action rights for AI agent remain outside Snowflake for a long time.

What to watch over the next four quarters. Paid usage of Cortex / Intelligence, whether AI workloads increase consumption, whether the semantic layer becomes the default enterprise agent context, and whether Snowflake enters observability, data engineering, and the application layer.

3.1.11 | Workday: Mature SoR + an Option on the Agent Workforce

Fact. The original text disclosed Workday FY2026 total revenue +13.1%, subscription revenue +14.5%, operating cash flow +19.4%, free cash flow +26.7%, and $2.9 billion in repurchases. The company described itself as an enterprise AI platform for managing people, money, and agents, disclosed delivery of 1.7 billion AI actions in FY2026, and acquired Pipedream to strengthen AI agent integration.

Inference. Workday's HR / Finance SoR is very stable, and errors involving people and money are costly, so in theory it has a position in governing the agent workforce. But AI actions are not paid outcomes, and the Agent System of Record still requires proof of customer adoption, revenue attribution, and budget migration.

Judgment. Workday looks more like mature compounding + an agent option, not the most aggressive agent winner today.

Falsification. AI actions do not convert into revenue; agent workforce governance is only a concept; Microsoft, SAP, Oracle, or customer-built HR agents abstract Workday; mature growth cannot reaccelerate because of AI.

What to watch over the next four quarters. The relationship between AI actions and paid SKU, whether Pipedream integration expands agent workflows, whether HR / Finance customers buy agent governance, and FCF/share and repurchase efficiency.

4.1 | AI Economics: Budgets, Pricing, Gross Margin

4.1.1 | Three Budget Migration Curves

The investment question for AI agent is not "will the features become stronger," but "where does the budget come from?" If AI merely makes existing software easier to use, at most it supports renewals, price increases, or lower churn; if AI can replace labor, reduce outsourcing, lower errors, shorten delivery cycles, reduce compliance risk, or create revenue, only then does it open new budget pools.

First curve: migration from seat budgets to workload budgets.

The core unit of traditional SaaS is the seat. In the agent era, customers care not only about how many people have permission to use the software, but how much work the software completed: customer service conversations, IT tickets, financial reconciliations, payments, tax preparation, content generation and approval, lead allocation, and customer updates.

Seats will not disappear; they will recede into the layer of identity, permissions, governance, supervision, and basic access. Agent workload, credits, tasks, and outcomes will become the incremental revenue layer.

Second curve: migration from software budgets to labor and service budgets.

This is the larger opportunity. Intuit is the clearest example: what customers truly want to buy is not the QuickBooks interface, but accurate tax completion, automatic account reconciliation, smooth payroll payments and collections, cash flow risk identification, and improved small business operations. If these tasks were previously done by accountants, tax preparers, bookkeepers, consultants, or BPO providers, AI agent gives software companies an opportunity to enter service budgets.

Third curve: migration from infrastructure budgets to business operations budgets.

Snowflake, Databricks, cloud vendors, and model providers sit on this curve. AI agent needs data, models, vector retrieval, permissions, logs, inference cost optimization, and tool calls. But the problem for the infrastructure layer is this: customers may view it as a cost center, not an outcome center. To move from being "the source of AI costs" to "the source of AI control," infrastructure companies must own business semantics, governance, agent run logs, default calls, and application workloads.

4.1.2 | Hybrid Pricing Stack

Software pricing in the AI agent era will not be one-size-fits-all. A typical enterprise software contract may become:

Base seats / role access rights
+ agent usage quota
+ premium workflow packages
+ governance / audit value-added modules
+ outcome-based tiered pricing or success fees
+ enterprise commitment discounts

Seats sell access rights, identity, permissions, collaboration, and accountable parties. Usage sells workload, pass-through of inference costs, and a low initial threshold. Credits are a compromise between usage and budget controllability. Outcome-based pricing sells task completion, risk reduction, revenue creation, and labor substitution.

Salesforce's official Agentforce pricing page already shows multiple models coexisting, including usage-based pricing, flexible credits, per-conversation pricing, and per-user pricing; ServiceNow is also emphasizing new commercial models and an AI Control Tower. The industry direction is not that "seats disappear," but that software is entering an era of hybrid billing.

4.1.3 | Gross Margin J Curve

AI will put software gross margins through four stages.

Gross margin recovery does not happen naturally. It requires seven conditions: mature model routing, effective caching and batch processing, small models that can replace large models, reusable workflow templates, more precise context, fewer error retries, and pricing that matches value.

If AI functionality becomes commoditized, customers demand price reductions, model vendors and cloud providers capture the value, and software companies bear integration and responsibility, gross margins may decline for an extended period.

4.1.4 | Who Can Retain the Benefits of Cost Declines

4.2 | How to Identify a Real agent Transformation

4.2.1 | Anti-Launch-Event Narrative: Five Types of Evidence

In the future, nearly every software company will say it has AI agent. Researchers must divide the evidence into five categories.

Salesforce's 2.4 billion agent work units and 19T tokens are useful signals, but the next questions remain: how much net new ARR do they correspond to? Have they entered production customers? Have they improved renewals? Have they expanded Data 360 attach? Have they reaccelerated the core clouds?

ServiceNow's large-ACV Now Assist customers are closer to commercialization evidence because they are tied to contract value; but gross margin, net expansion, and whether the AI Control Tower becomes the default entry point still need to be watched.

4.2.2 | Leading, Concurrent, and Lagging Signals

The most valuable stage for long-term investment is before leading signals enter concurrent signals. But if there are only leading signals and no concurrent signals, they still should not be capitalized too early.

4.2.3 | Ten Questions for an AI agent Software Company

1. What core state change does it control?
2. Is the cost of errors in this state change high?
3. Does it own context, rather than merely calling external data?
4. Does it have write access, or does it only provide recommendations?
5. Can it close the loop: capture, judge, execute, write back, verify, and correct errors?
6. Can it assume responsibility or help customers transfer responsibility?
7. Can it attribute AI value to time, cost, risk, revenue, or cash flow?
8. Does it have evidence of standalone pricing, incremental usage, add-on purchases, NRR, or orders?
9. Does its unit cost decline with scale?
10. Can it prevent model vendors, cloud providers, systems integrators, or customer-built solutions from capturing the profits?

If a company can only answer "we have AI functionality," it does not enter investment-grade research.

4.3 | The Biggest Counterexamples and Abstraction Risk

4.3.1 | Model Vendors and Cloud Providers Capture the Profits

If application software merely wraps a model in a UI and lacks its own context, processes, data, write access, and responsibility, model vendors can provide the functionality directly, cloud providers can platformize it, and customers can also build it themselves. To prevent value leakage, the application layer must own proprietary context, high-responsibility workflows, write access, outcome attribution, the customer's default entry point, and an auditable governance layer.

4.3.2 | Microsoft and Platform Bundling

Microsoft's threat to enterprise software is not a single feature, but the combination of distribution, identity, Office / Teams entry points, Azure, Power Platform, and Copilot. Salesforce, ServiceNow, Workday, and Adobe all must prove they are not backend systems abstracted by the Microsoft entry point.

4.3.3 | Customers Build Their Own Execution Layer

Large enterprises may build general-purpose agent, the data layer, permissions, logs, audit, and tool calls on internal platforms, then call external SaaS APIs. In that case, SaaS companies remain important, but the control point is migrated away. The standard for judgment is: whether customers hand the agent's next action, permission boundaries, audit entry point, and outcome attribution to the vendor, or keep them in a self-built platform.

4.3.4 | AI Usage Does Not Convert Into Revenue

Usage is a demand signal, but not revenue quality. Metrics such as AI actions, tokens, queries, agents launched, and pilots have investment-grade meaning only when they enter paid SKU, ARR, RPO, NRR, gross margin, and FCF/share.

4.3.5 | Long-Term Gross Margin Deterioration

AI may cause software to deteriorate from high-gross-margin subscriptions into hybrid delivery with high service costs, inference costs, and customer success costs. The most dangerous path is: customers view AI as basic functionality and are unwilling to pay; model and cloud costs rise; agent errors require more human backstops; implementation and governance complexity increases; software companies bear the costs but lack pricing power.

4.3.6 | agent Errors Lead to Permission Contraction

The biggest risk for high-responsibility agent is error. Erroneous payments, erroneous filings, erroneous approvals, erroneous customer communications, and erroneous security actions can all cause enterprises to shrink permissions. The governance layer is not a nice-to-have, but software infrastructure in the agent era.

4.4 | From Trend to Per-Share Value: Investment Conclusion and Quarterly tracker

4.4.1 | Why the Trend Must Land in FCF/share

A strong industry trend does not equal strong stock returns. For the AI agent trend to enter long-term investment judgment, it must cross this bridge:

Control point
→ Orders / ARR / RPO / NRR
→ Gross margin and AI costs
→ SBC / dilution
→ FCF/share
→ What the current valuation implies
→ Whose odds are more attractive

If a company can only prove AI usage but cannot prove revenue, gross margin, and cash per share, the trend can only remain at the level of industry judgment. Only when a company can turn AI control points into FCF/share, while valuation has not prepaid too much future success, does it approach an investment-grade conclusion.

4.4.2 | Company-Level Per-Share Value Bridge

4.4.3 | Current Investment Research Conclusion

The companies currently most worth entering deep single-stock research are ServiceNow, Intuit, and Veeva.

ServiceNow's advantage is that it is closest to the horizontal workflow runtime layer, and there is already evidence that large Now Assist ACV, RPO / cRPO, and subscription revenue are moving in the same direction. The biggest question is whether it will be abstracted away by Microsoft, cloud vendors, or customers' self-built agent platforms.

Intuit's advantage is that it is closest to vertical outcome delivery, and AI + expert has the opportunity to enter labor and service budgets. The biggest questions are the service-model cost structure, pressure from free tax filing, and competition for the SMB operating entry point.

Veeva's advantage is that it has high-accountability industry semantics, compliance workflows, and hard evidence of Vault CRM migration. The biggest questions are slow growth, a long AI monetization cycle, and the need not to conflate the AI narrative with Vault CRM migration.

Adobe, Salesforce, Snowflake, and Workday are all worth tracking, but they need more evidence: Adobe needs to prove the enterprise content supply chain, Salesforce needs to prove Agentforce ARR and production customers, Snowflake needs to prove that AI increases consumption and enters the action layer, and Workday needs to prove that agent workforce governance becomes a budget pool.

4.4.4 | Upgrade / Downgrade Decision Conditions

4.4.5 | Unified Tracker for the Next Four Quarters

5.1 | Appendix A: Complete Software Control Point Framework

Research on any AI software company can proceed through 13 questions:

1. Core state change: What state was the customer originally in, and what state does the software move the customer into?
2. Cost of error and accountability: How high is the cost if this state change goes wrong?
3. Context control: Does the company own the context, or is it merely calling external context?
4. Write access: Can the company write back to the system, trigger actions, and close tasks?
5. Closed-loop control: Can it capture state, judge, execute, write back, verify, and correct errors?
6. Business clock: What are the event frequency, value per instance, wallet expansion surface, and sales friction?
7. Pricing migration: Can it move from seats into usage, tasks, outcomes, or services budgets?
8. Unit economics: Is the unit cost of completing each high-value state change declining?
9. Gross margin J-curve: Does short-term cost pressure have a path to long-term recovery?
10. Competitive capture: Are profits taken away by model providers, cloud vendors, platforms, or customer self-build?
11. Outcome attribution: Does the customer believe the outcome comes from the software?
12. Per-share value: Does the trend flow into FCF/share rather than only into non-GAAP metrics?
13. Counterfactuals and falsification: What evidence would overturn the thesis?

5.2 | Appendix B: Detailed Model of Pricing Migration

Seats still matter because they define access rights, identity, permissions, collaboration, and accountable parties. After agent replaces some operations, enterprises still need to know on whose behalf agent is acting, who approved it, and who is responsible.

Usage matters because it covers inference costs and workload growth. Credits matter because enterprises need budget control. Outcome-based payment matters because customers ultimately want to buy completed tasks, reduced risk, revenue creation, and labor substitution.

The most likely end state is neither pure seats nor pure token, but hybrid pricing: base platform fee + agent usage credits + advanced workflow packages + governance / audit modules + outcome-based tiered billing.

The scenarios best suited to outcome-based pricing include tax filing, customer service issue resolution, fraud interception, security threat blocking, compliance approval, and sales lead conversion. The scenarios least suited to outcome-based pricing include creative exploration, general knowledge queries, low-accountability writing, and brand advertising with long attribution chains.

5.3 | Appendix C: Detailed Breakdown of the Gross Margin J-Curve

AI gross margin pressure comes from four sources: model inference costs, R&D costs, enterprise deployment and data governance costs, and customer success and human fallback costs.

Gross margin recovery requires seven conditions: model routing, substitution with smaller models, cached input, batch, prompt compression, more precise retrieval, workflow templating, and fewer error retries.

OpenAI official API pricing shows model capability and price tiers, with cached input priced below ordinary input, and the Batch API can offer discounts. This indicates that inference costs can be engineered down, but whether cost declines turn into software company profits depends on control points and pricing power.

5.4 | Appendix D: Complexity Migration and Game-Theory Model

AI reduces single-point functional complexity and UI friction; it does not eliminate business complexity. The business world is not a static system that maximizes efficiency, but a game system in which multiple parties compete for profit pools, entry points, standards, and boundaries of responsibility.

Generating a marketing email has become cheaper, but deciding whom to send it to, when to send it, whether it is compliant, whether it fits the brand, whether it affects customer lifetime value, whether it writes back to CRM, whether it triggers sales follow-up, and whether revenue can be attributed to it has instead become more complex.

The new moats come from five positions: context, action permissions, accountability governance, closed-loop learning, and organizational embedding. The old complexity was people learning to use software; the new complexity is enterprises trusting software to deliver business outcomes.

5.5 | Appendix E: Complete Screening Funnel

Gate 1: Whether the state change is high value.
Gate 2: Whether the context and data are proprietary and can trigger actions.
Gate 3: Whether write access has been upgraded from D0-D2 to D3-D5.
Gate 4: Whether the cycle of accountability and trust formation is deep enough.
Gate 5: Whether closed-loop control is complete.
Gate 6: Whether the budget moves from software into labor, services, risk, or outcomes.
Gate 7: Whether unit economics can scale.
Gate 8: Whether competitive capture can withstand model providers, cloud vendors, platforms, and customer self-build.
Gate 9: Whether per-share value improves.

If a company cannot pass Gate 1 through Gate 5, it is merely an AI feature company. If it passes Gate 1 through Gate 7 but per-share value does not improve, it is a good trend but not necessarily a good stock. Only if it passes all Gates does it enter the long-term compounding candidate research pool.

5.6 | Appendix F: Key Sources

1. ServiceNow Q1 FY2026 Results
   https://investor.servicenow.com/news/news-details/2026/ServiceNow-Reports-First-Quarter-2026-Financial-Results/default.aspx

2. ServiceNow Knowledge 2026 AI Control Tower / Autonomous Work
   https://newsroom.servicenow.com/press-releases/details/2026/ServiceNow-turns-enterprise-AI-chaos-into-control-with-the-platform-for-governed-autonomous-work/default.aspx

3. Intuit Q2 FY2026 Results
   https://investors.intuit.com/news-events/press-releases/detail/1307/intuit-reports-strong-second-quarter-results-and-reiterates-full-year-guidance

4. Intuit Virtual Team of AI Agents
   https://investors.intuit.com/news-events/press-releases/detail/1258/intuit-introduces-ground-breaking-virtual-team-of-ai-agents-to-fuel-growth-for-businesses

5. Veeva FY2026 Q4 and Full-Year Results
   https://ir.veeva.com/news/news-details/2026/Veeva-Announces-Fourth-Quarter-and-Fiscal-Year-2026-Results/default.aspx

6. Adobe Q1 FY2026 Results
   https://www.adobe.com/cc-shared/assets/investor-relations/pdfs/21306202/ay45th643t5y46.pdf

7. Salesforce FY2026 Q4 Results
   https://investor.salesforce.com/news/news-details/2026/Salesforce-Delivers-Record-Fourth-Quarter-Fiscal-2026-Results/default.aspx

8. Salesforce Agentforce Pricing
   https://www.salesforce.com/agentforce/pricing/

9. Snowflake FY2026 Q4 and Full-Year Results
   https://www.snowflake.com/en/news/press-releases/snowflake-reports-financial-results-for-the-fourth-quarter-and-full-year-of-fiscal-2026/

10. Workday FY2026 Q4 and Full-Year Results
    https://investor.workday.com/news-and-events/press-releases/news-details/2026/Workday-Announces-Fiscal-2026-Fourth-Quarter-and-Full-Year-Financial-Results/default.aspx

11. OpenAI API Pricing
    https://openai.com/api/pricing/