What Really Happened With Agents in the Q2 2026 Earnings Season for US and A-Shares

Author: qinbafrank

I've been digging through the earnings reports of agent-related companies in both the US and A-share markets over the past couple of days, and the impression is strong: agents are starting to move from launch events into financial statements.

Over the past year, the market's view of agents focused mainly on whether models could invoke tools, whether they could complete multi-step tasks continuously, and how many agents a company had released.

By this earnings season, the metrics have clearly changed. Now the focus is on how much ARR and ACV agents are contributing, how many new orders they're bringing in, how many customers are in production environments, and how much actual work is being completed each quarter.

This shift is significant.

A product that works only proves the technical direction is viable. Revenue, contracts, collections, and profits starting to appear show that enterprises are willing to pay for this capability on an ongoing basis.

So I'm more inclined to define 2026 as the "starting point and inaugural year of financial validation" for enterprise-grade agents. This is the validation year—there's still a long way to go before the industry matures—but at least the first batch of real data that can be plugged into financial models has emerged.

I. US Stocks Have Already Provided the First Round of Answers

The companies that are clearest right now basically all control enterprise data, permissions, and workflows.

1. ServiceNow's Q2 subscription revenue reached $3.877 billion, up 24.5% year-over-year; AI business ACV surpassed $1 billion, and agent deployments grew ninefold in nine months.

Many people used to see ServiceNow as IT ticketing and workflow software, but it's increasingly looking like the control center for enterprise agents.

In the future, enterprises may run Claude, Gemini, OpenAI, and various vertical models simultaneously. Which agent can access which data, who it can act on behalf of, which actions require approval, and how to roll back after errors—all these issues need a unified governance layer.

ServiceNow is already in that position. What it sells is gradually extending from workflow software to the identity, permissions, monitoring, and execution system for agents.

2. Salesforce's validation is also very direct

Agentforce ARR has exceeded $1.5 billion, up more than 240% year-over-year; Agentforce and Data 360 combined ARR is close to $3.9 billion. In Q2, Agentforce and Slack completed 3.2 billion Agentic Work Units, up 97% quarter-over-quarter.

In the past, the market worried that general-purpose large models would bypass CRM and directly take over sales and customer service. Now it seems that for large models to truly complete enterprise tasks, they still need customer data, historical orders, contract status, sales leads, and permission systems.

Claude can handle intent understanding, while Salesforce holds the business context and execution track.

Of course, the 240% annual growth rate needs a slight discount. Starting this quarter, Salesforce expanded the statistical scope of Agentforce ARR to include other AI products, Slackbot, and Headless 360. The direction is clear, but the specific growth rate still needs to be judged in light of the change in scope.

3. Workday's earnings are also very representative

AI has already contributed more than 25% of new ACV, and over 5,500 customers are using at least one Workday-built agent.

This data shows that in sensitive areas like HR, finance, and auditing, enterprises are not simply bypassing their existing software platforms. Employee data, compensation rules, financial systems, and organizational permissions are all embedded within Workday, and agents need to work along these deterministic tracks.

For many traditional SaaS companies, AI will indeed compress some pages, buttons, and low-value seats. But platforms that own core data and business rules may actually gain new upsell opportunities because of agents.

4. Palantir represents another form

Q2 revenue grew 93% year-over-year, and US commercial revenue grew 149%. Palantir's strength has always been connecting enterprise data, business objects, permissions, and actual production processes, then using software to complete decisions and execution.

Many agent products stop at "helping you generate an answer," while Palantir is closer to "helping the enterprise complete an outcome."

Manufacturing companies care about how much equipment downtime has been reduced, supply chain companies care about how much inventory turnover has improved, and sales teams care about how much conversion rates have increased. What enterprises are ultimately willing to pay a premium for are usually these quantifiable business results.

Putting these companies together, a clear trend emerges:

The first to benefit from the agent dividend are generally companies that were already in the core business systems of enterprises.

Model capability is certainly important, but deploying agents in enterprises also requires data, permissions, processes, auditing, and customer entry points. Simply connecting a model and adding a chat interface will quickly see its barriers lowered. This is also what I discussed in the long article about AI adoption entering the engineering era: SaaS companies that truly own workflows and core proprietary data hold the work entry point, business objects, business semantics, user permissions, historical operation data, system records, industry rules, final action execution, and customer outcome feedback.

II. The Most Important Change This Earnings Season Is Actually the Measurement Method

In the past, when judging AI products, the market often looked at registered users, trial customers, and token call volume. These metrics can prove interest, but they can hardly prove that the business model works.

Now more and more companies are disclosing: how much new ACV AI contributes, how much ARR agents generate, how many customers have moved from trial to production, and how many tasks have actually been completed.

Salesforce starting to use Agentic Work Units to measure workload is a very important signal.

In the past, SaaS mainly charged by seat: an enterprise with 10,000 employees buys 10,000 accounts. In the future, as agents complete a large number of tasks for enterprises, the pricing model may gradually shift to:

Base seat fee + AI usage + task completion volume;

Further down the road, some high-value scenarios may even charge by outcome.

For example, how many tickets a customer service agent resolves, how many leads a sales agent screens, how many invoices a finance and tax agent processes, and how many risk events a security agent closes.

Once the billing unit shifts from "how many people are using the software" to "how much work the software has completed," the revenue ceiling for SaaS will also change.

It starts to tap into enterprise labor costs, outsourcing expenses, and operational budgets, rather than just reallocating within the IT budget.

III. Why Improvements in Model Intelligence Will Make Agents a Long-Term Industry

There is a clear capability threshold in the agent space: suppose an agent has a 95% success rate for a single-step task, after executing ten steps consecutively, the probability of the entire task being completed successfully is only about 60%.

So early agents often had a problem: each step looked smart, but when placed in a long process, errors started to appear. Enterprises still had to assign employees to check, correct, and rework, and in the end, not much time was saved.

As model reasoning, context length, tool invocation, and memory capabilities improve, this situation will gradually get better: the more steps an agent can complete consecutively, the fewer nodes require human intervention, and work that previously couldn't be automated will cross the commercialization threshold.

This is not a linear process.

A slight improvement in model capability may only make answers more fluent; when reliability crosses a certain threshold, an entire workflow has the chance to be automated end-to-end.

So agent revenue is likely to show strong nonlinearity in the future.

Today it can only organize data, next it can fill in systems, and later it can submit approvals, track progress, and handle exceptions. Each additional link completed significantly increases business value.

This is also why agents don't need to achieve so-called artificial general intelligence first.

Finance and tax, customer service, IT operations, sales lead screening, and document processing all have a large number of tasks with clear boundaries, high repetition frequency, and checkable results. This is also the second scenario I discussed after malicious coding:

As long as agents can reliably complete most of these tasks, customer ROI is already established.

IV. A-Shares Are Also Starting to Validate, but the Path Differs from US Stocks

What US stocks are currently monetizing first are enterprise platform and workflow companies, while in A-shares, vertical scenarios deserve more attention.

1. Kingsoft Office's H1 revenue was 3.313 billion yuan, up 24.69% year-over-year, and WPS 365 has maintained growth of over 60% for six consecutive quarters.

Kingsoft Office's advantage is intuitive. It owns the document entry point, format capabilities, collaboration relationships, and enterprise office data. Future office agents can directly generate, modify, review, and deliver documents, which is more valuable than placing a chat window next to documents.

But Kingsoft Office's H1 net profit attributable to the parent grew more than 200%, largely from investment income. When looking at the main business, these one-time items still need to be stripped out.

2. Servyou is a relatively typical case among A-share vertical agents

H1 revenue was 1.012 billion yuan, up 9.72% year-over-year; non-GAAP net profit grew 44.04%. AI product and business collections already account for 33% of digital finance and tax business collections, up from 28% for full-year 2025.

The finance and tax scenario is very suitable for agents. Rules are relatively clear, data is highly structured, results can be reviewed, and customers are willing to pay to reduce labor costs and tax risks.

This type of business doesn't require a model that does everything. As long as it is reliable enough in a narrow scenario, it can generate real revenue.

3. Hand Enterprise Solutions

H1 AI intelligent business revenue was about 220 million yuan, up more than 100% year-over-year, already accounting for a double-digit percentage of company revenue.

Hand has long been engaged in ERP, supply chain, manufacturing, and financial system implementation. The biggest difficulty in enterprise agent deployment often lies in legacy system connections, non-standard data, and complex approval processes. Hand is familiar with these links and has the opportunity to become the deployment and integration layer for domestic enterprise AI.

But in A-shares, special attention must be paid to financial quality, so while the opportunity for A-share agents is large, the analysis difficulty may be higher than for US stocks.

Many companies still rely mainly on project-based revenue. Whether AI business can be standardized, whether gross margins can improve, and whether they can shift from one-time implementation to subscription and continuous usage-based billing still need further observation.

V. The Agent Industry Chain May Form a Four-Layer Structure

The first layer is foundation models and computing power, which determine the intelligence ceiling and inference cost of agents.

The second layer is data, identity, permissions, and governance. ServiceNow, Salesforce, and Workday are currently competing for this layer.

The third layer is workflow orchestration and execution. Palantir, UiPath, and some enterprise digital service providers compete here.

The fourth layer is vertical industry agents, including finance and tax, office, legal, healthcare, manufacturing, customer service, and sales.

Currently, the clearest financial validation is in the second layer.

Enterprises dare not let an agent without permission boundaries and auditability directly operate core systems, so data and governance platforms have a strong moat.

The layer with more long-term elasticity may be the fourth layer. Vertical agents can directly correspond to specific workloads, ROI is easy to calculate, and it is easier to enter labor and outsourcing budgets.

The third layer determines whether agents can truly enter production environments. It's not enough for models to think; they also need to connect to legacy systems, invoke tools, handle exceptions, and complete the entire process.

UiPath will be a very important observation sample. The company plans to release its Q2 earnings after market close on September 3, 2026, US Eastern Time, at which point we can further judge the speed of traditional RPA upgrading to the agent execution layer.

VI. How to Evaluate Agent Companies Going Forward

In the future, when looking at these companies, I think several metrics will become increasingly important:

1. First look at AI-related ARR, ACV, and real revenue, then look at the number of production customers; the reference value of trial customers will become lower and lower;

2. Then look at AI's contribution to new contracts. Workday's disclosure that "AI accounts for more than 25% of new ACV" is more meaningful than how many agents have been released;

3. Also look at task invocation volume, complete task success rate, renewal rate, and customer expansion rate;

4. Finally, return to the financial statements and observe gross margin after deducting inference costs, implementation cycles, accounts receivable, and operating cash flow.

If every agent deployment requires a large number of engineers to spend six months on customization, it is closer to traditional IT services.

If agents can be replicated quickly, and the more customers use them, the lower the unit cost, then the operating leverage of software companies will truly be unleashed.

Finally, my understanding of the agent space:

1) Improvements in model intelligence determine how much agents can do;

2) Enterprise data, permissions, and workflows determine whether this intelligence can safely enter production systems;

3) Real task outcomes determine how much customers are willing to pay.

Every model upgrade pushes the boundary of automatable tasks outward.

Today it's programming, customer service, and documents; later it will enter finance, sales, supply chain, healthcare, and more professional scenarios. As reliability improves, manual checks decrease, and unit economics will continue to improve.

So agents are likely to be a new efficiency-driven industry that can last for many years.

However, this round of opportunities will not be evenly distributed.

The first batch of winners will most likely be platforms that already control enterprise data and business processes. The second batch will be execution-layer companies that can deploy AI into production environments and truly close the task loop. The elasticity of A-shares comes more from vertical scenarios, but financial quality and revenue standardization need to be watched more closely.

2026 is more like the inaugural year of financial validation for agents. In the past, the market discussed how smart models are; going forward, it will increasingly care about: how much work they actually completed, how much money they saved customers, and how much value ultimately entered the revenue, profit, and cash flow of listed companies.