According to McKinsey’s global survey (2025), 23% of companies are scaling AI agents in at least one function within the organization, and 39% are in the experimentation stage, together bringing the total engaging with agents in some form to 62%. However, the share of companies reporting that they are scaling agents in any specific business function remains at 10% or below across every function.

According to G2’s August 2025 survey, 57% of companies have AI agents in production, 22% are in the pilot stage, and 21% are pre-pilot.

Meanwhile, a Lucidworks survey of more than 1,600 AI leaders reveals a harsher reality: while more than 70% of companies have adopted generative AI in their operations, only 6% have fully implemented agentic AI, and just 2% are running multiple agents in production.

Market Size and Growth Rate

The AI agent market is growing at a CAGR of 46.3%, and is projected to expand from $7.8 billion in 2025 to $52.6 billion by 2030. Salesforce’s research also points to a 282% year-over-year increase in AI adoption.

According to Gartner, as of Q1 2026, some form of AI agent is embedded in 80% of new or updated enterprise applications, a sharp rise from 33% in 2024. 31% of companies have agents in production, and the gap between industries is significant: banking and insurance leads at 47%, followed by healthcare at 18% and government agencies at 14%.

Payback Period

According to data from BCG and Forrester, the median payback period for agent deployments is 5.1 months. By function, SDR (sales development representative) agents pay back fastest at 3.4 months, while finance and operations agents take 8.9 months.

Named Overseas Company Case Studies (Public Information)

JPMorgan (US, Finance)

More than 200,000 JPMorgan employees use the company’s in-house LLM Suite platform, delivering an estimated 15 million hours of labor savings and generating more than $2 billion in business value annually. The company has more than 450 AI use cases in production, applied to tasks such as automatically drafting M&A memos, automating trade settlement, and fraud detection. CEO Jamie Dimon has said, “AI is a living part of our business.”

Klarna (Sweden, Fintech) — A Case with Both Success and Failure

Klarna’s AI customer service agent handles work equivalent to 853 full-time employees and has achieved $60 million in cost savings. The company claims customer satisfaction is “on par” with human agents.

However, this case carries an important lesson. Klarna initially carried out large-scale workforce reductions in favor of AI agents, but an internal review in early 2025 found that customer satisfaction had declined for complex or emotionally charged inquiries. CEO Sebastian Siemiatkowski publicly admitted the company had “gone too far,” and subsequently began rehiring human staff. It is a case that left behind a lesson: simple, routine inquiries can be automated, but interactions involving emotion and judgment still require humans.

Morgan Stanley (US, Finance)

Morgan Stanley’s AI agent for wealth management generates meeting minutes and action items after advisor meetings and automatically syncs them with Salesforce CRM, achieving a 98% adoption rate among financial advisor teams. This is an exceptionally high figure, considering that voluntary adoption rates for typical enterprise software are generally below 60%. (AI Monk)

In addition, Morgan Stanley introduced its in-house legacy code modernization agent, “DevGen.AI,” in January 2025. The agent has reviewed more than 9 million lines of legacy code, saving developers an estimated 280,000 hours. As a result, 15,000 developers have been freed from code translation work and can now focus on strategic tasks.

Salesforce (US, SaaS)

Salesforce uses AI agents internally, achieving $100 million in annual cost savings and generating more than 3,200 deals. AI agents autonomously reach out to low-priority prospect leads internally nicknamed “sawdust,” uncovering deals that human staff had previously been unable to keep up with.

Siemens (Germany, Manufacturing)

Siemens has deployed AI on its manufacturing lines, reducing machine downtime by 20% and avoiding millions of dollars in production losses. The company applies AI agents to predictive maintenance, quality control, and energy optimization, positioning it as a leading example in the manufacturing industry.

ServiceNow (US, IT Service Management)

By integrating AI agents, ServiceNow has reduced the processing time for complex customer service cases by 52%. Its primary applications are automated incident response and change management automation within IT departments, making it a reference case for enterprise IT operations automation.

Walmart (US, Retail)

Walmart has built “Wallaby,” a proprietary retail-focused LLM trained on decades of POS, customer, and inventory data. It processes real-time sales data from 4,700 stores and autonomously decides on replenishment orders without human approval. Rather than a general-purpose model, its defining strategy is a “purpose-built model trained on the company’s own data.”

Introduction: What Is an AI Agent?

Much of the AI we have encountered until now has been the “chatbot (conversational AI)” type, which returns answers to questions. However, the paradigm shift now underway is an evolution one step further, into the “AI agent.”

The Evolution from “Chatbot” to “Agent”

If the conventional chatbot is a “knowledgeable conversation partner,” the AI agent is an “autonomous operator that gets the work done.”

From AI That Waits for Instructions to AI That Thinks and Acts on Its Own Toward a Goal

A chatbot generates an answer from its training data in response to an input such as “tell me about X.” An AI agent, by contrast, is given an ambiguous goal such as “resolve the stock shortage,” and then attempts to autonomously carry out the entire process: checking current inventory, calculating the quantity that needs to be ordered, selecting an appropriate supplier, and finally completing the purchase order.

The Three Core Elements of AI Agents: Reasoning, Tools, and Action

For an agent to maintain autonomy, the following three elements must form a loop.

Reasoning: Using an LLM (such as Claude) as the “brain,” logically working out what to do next.
Tool Use: Leveraging tools that operate external systems, such as SAP’s APIs, Excel, email, and web search.
Acting: Based on the plan it has assembled, actually executing tool commands and evaluating the results.

Why Are Agents Needed for Core System Automation Now?

Until now, RPA (Robotic Process Automation) has played the leading role in automating core systems such as SAP. However, RPA has a weakness: it “cannot deviate even a millimeter from a predetermined procedure (scenario).”

The reason AI agents are now in demand is to handle the “complexity” and “exceptions” specific to core business operations.

Bridging Gaps Between Processes

Work that is completed entirely within a core system alone is rare. In practice, tasks such as reading a delivery-date reply that arrived as a PDF by email, updating a purchasing document in SAP, and reporting to the person in charge via chat if there is a delay, span multiple applications. Agents connect these seamlessly.

The Surge in “Unstructured Data”

Core systems handle well-organized data (structured data), but the entry point of real-world operations is unstructured data such as “ambiguous emails from customers” or “handwritten instructions.” Agents with advanced reasoning capabilities have become essential for understanding this data and converting it into a format the system can accept.

Automating Judgment That Used to Depend on Individuals

Judgment logic that only veteran staff on the front line used to possess — for example, “when this error occurs, enter this value based on past experience” — is now starting to be replaced by AI agents that read through past logs and manuals.

Latest Examples: The Emergence of Autonomous Agents

From 2025 into 2026, the concept of “Agentic Workflow” has become mainstream.

For example, rather than having a single AI do everything, leading enterprises have begun testing collaborative, multi-agent automation in which an “inventory analysis agent” detects an anomaly, hands a task off to a “procurement agent,” and an “approval agent” seeks a decision from a human.

Leading Examples: How AI Agents Are Transforming the Workplace

The true value of AI agents lies not merely in automating data entry, but in “completing processes that involve situational judgment.” The following introduces the latest use cases from the areas of “supply chain” and “finance and accounting,” where progress has been especially remarkable.

Latest Use Cases from Japan and Overseas

① Supply Chain: Completing the Process from Anomaly Detection to Proposing Alternatives and Placing Orders

With conventional systems, the best that could be done when a stock shortage or delivery delay occurred was to issue an alert. The latest AI agents autonomously carry out what happens “beyond” that.

Flow of Action:

Detection: Continuously monitor logistics data and production progress in real time to identify the risk of a stock-out occurring in the coming days.
Analysis: Search past transaction data and a list of alternative suppliers, and formulate three recovery options that balance shipping cost, delivery time, and quality.
Execution: Automatically send an inventory-confirmation email to the optimal alternative supplier, read the reply (in natural language) confirming availability, and automatically update the purchase requisition in SAP.

Effect: The lag between grasping the situation and making a decision has shrunk from days to minutes, dramatically reducing the lost business opportunities caused by stock shortages.

② Finance and Accounting: Automatically Analyzing Anomalies in Financial Reports and Drafting Confirmation Emails

In monthly closing and budget management, the work of pinpointing “why did this cost spike” among a mountain of figures is extremely labor-intensive.

Flow of Action:

Identification: Automatically extract SG&A and cost items that exceeded a threshold compared with the previous month or the budget.
Deep-Dive: The agent traces back through related journal entry details, receipts, and even the travel expense request system to investigate, and puts factors such as “a spike in airfare costs tied to a specific large project” into words.
Communication: Compile the findings into a report, and if there are unclear points, draft (or send) an email to the person in charge in the relevant department that reflects an understanding of the context, such as “could you please confirm the costs for project XX?”

Effect: Accounting staff are freed from “chasing numbers” and shift toward “management work” — confirming and approving the analysis results the AI presents.

The Depth of Automation: Success Stories with Human-in-the-Loop

Currently, entrusting every process to full automation by AI carries risk from a security and accountability standpoint. What has proven successful instead is a semi-automated model that incorporates “human-in-the-loop.”

What Is Human-in-the-Loop?

A design in which humans handle only the important decisions and final approvals within a process that AI executes.

Example Patterns:

Integrated Approval Flow: The agent notifies a human through Slack or Microsoft Teams with a button, saying, for example, “I propose replenishing inventory for the following reason. Do you approve?” The moment the human clicks “Approve,” data is written to the core system (SAP).
Branching by Confidence Score: When the AI is confident in its judgment (a confidence score of 90% or higher), it executes automatically; when it is not confident, it flags the case and escalates it to a human.

Rather than “AI doing everything on its own,” the model is “AI creates a perfect draft, and a human gives the final stamp of approval.” This division of roles is, at present, the realistic and fastest route to success for AI agent adoption in the enterprise space.

Other AI Agent Case Studies

Sales Domain

Inquiries and Quotes (AI + RPA): AI agents reference past deals, costs, and profit margins to estimate prices, while RPA automates the creation and sending of quotations. Quote preparation time is reduced by 50-70%. Transcription errors are nearly eliminated.
Order Receipt (AI + RPA): AI-OCR and AI agents analyze diverse order forms received by fax, email, and EDI and convert them into data, which RPA then automatically enters into the core system. When stock is insufficient, AI automatically alerts the purchasing staff. At Matsuya (a food trading company), RPA alone reduced annual working hours by 3,276 hours, and the company is now conducting a PoC for full automation with AI agents (public case study from Uchida Yoko).
Shipping (RPA-led): RPA automatically executes the entire sequence of processes from inventory allocation to creating shipping instructions and linking instructions to the WMS. In areas where the API between SAP and the WMS is not fully built out, RPA in particular functions as the “connector” (public case study from QUNIE).
Billing (AI + RPA): RPA retrieves sales data from the ERP and automatically issues and sends invoices. AI agents judge irregularities in account classifications to speed up monthly closing. Monthly closing has been shortened from 8.5 days to 3-4 days (BlackLine Finance Benchmark 2025).

Production Domain

Production Planning (AI-agent-led): SAP’s production planning agent autonomously checks the availability of materials, capacity, and schedules, and automatically releases production orders once conditions are met (SAP Q1 2026 GA). Demand-forecasting AI links with external data to dynamically update MRP. 61% of manufacturers report cost reductions from AI (Vellum 2025 survey).
Manufacturing Execution (AI + RPA): AI analyzes equipment sensor data in real time to detect failure predictions and quality anomalies. RPA handles the automatic creation of production instructions and posting production results to the core system. At Fujita Manufacturing, RPA automated everything from order processing to production instructions, cutting 1,850 hours annually (public case study from WinActor). At an aerospace manufacturer, maintenance planning workload was reduced by 30% (ERP Today 2025, company name undisclosed).
Inventory Transfer (RPA-led): RPA automatically synchronizes inventory data updates between sites and channels multiple times a day. Automatic notifications are triggered when stock levels fall below a threshold. At Koyo Toki, automatic updating of inventory status reduced out-of-stock orders (public case study from RoboTANGO).

Procurement Domain

RFQ (AI-agent-led): AI agents automatically analyze suppliers’ email replies, comparing price, delivery time, and terms across multiple suppliers and parts at once to recommend the optimal option. In September 2025, Epicor began offering an RFQ automation agent on a success-fee basis (billing only for deals that close). Procurement cycles are shortened by up to 60% (Zycus case study), and sourcing cycles are cut by 40% (McKinsey survey).
Purchase Ordering (AI + RPA): Demand-forecasting AI calculates order quantities from inventory and sales data, and RPA automatically creates and sends POs to suppliers. At ORO AI, automating PR review (purchase requisition checks) shortened the cycle by 70%, and touchless ordering increased 2.9-fold (public data from ORO AI).
Goods Receipt (RPA-led): RPA automatically performs three-way matching of the delivery note, purchase order, and receipt, then automatically registers the goods-receipt data in the ERP. Combining this with AI-OCR also supports paper delivery notes.
Invoice Matching (AI + RPA): AI-OCR reads supplier invoices in a variety of formats, and AI agents perform three-way matching against the PO and goods-receipt data. RPA automatically approves matching items and places them into the payment queue, while mismatches are escalated. The three-way matching automation rate is 85-92% (Aberdeen Group 2025). The cost per invoice has dropped 71%, from $10.18 to $2.94, and the approval cycle has shortened from 14 days to under 3 days (Ardent Partners 2025).

Finance and Management Accounting Domain

Cash Receipts (AI + RPA): RPA automatically matches incoming payment data against accounts receivable data (cash application). For irregular cases such as partial payments, an AI agent analyzes the situation and proposes a resolution to the person in charge, allowing staff to focus solely on confirming exceptions.
General Ledger, Journal Entries, and Monthly Closing (AI-agent-led): AI agents analyze transaction content and automatically propose the optimal account classification and journal entry. Monthly ledger reconciliation and variance-explanation reports are also generated automatically. SAP’s cash management agent autonomously processes bank statements, cutting the workload for cash position management by 70% (SAP Connect 2025). First-year ROI for finance teams is 200-400% (Deloitte CFO Signals Q4 2025).
Payments (RPA-led): RPA fully automates payment-due-date management, generating remittance data, and submitting it to the banking system. Early payment discount opportunities are also flagged automatically.
Actual Costing (AI-agent-led): AI automatically collects and aggregates data on production results, material costs, labor costs, and overhead, and automatically generates factor analysis and improvement proposals for cost variances. Deviations from standard cost are monitored in real time.
Standard Costing (AI-agent-led): AI autonomously monitors market fluctuations in material and labor costs by linking with external data. It proposes the timing and revised values for reviewing standard costs, and automatically runs budget-comparison simulations.

The Affinity Between SAP and AI Agents

The biggest barrier when implementing AI agents in a business is “the state of data organization” and “process complexity.” SAP holds an “AI-friendly” foundation on both counts that overwhelms other systems.

SAP as a “Treasure Trove of Structured Data”

For an AI agent to operate autonomously, it needs to be able to accurately read and write data within a system. SAP has long been designed around the philosophy of centrally managing all of a company’s data, and its data structures are defined with extreme rigor.

Seamless Integration Through APIs (OData/BAPI)

SAP has a very rich set of interfaces (APIs) for manipulating data from the outside.

OData: A web-standard protocol that provides data in a format that is easy for AI agents to read.
BAPI (Business Application Programming Interface): A mechanism for calling SAP’s internal business logic from the outside.

AI agents (such as Claude) can “understand” these API specifications and autonomously determine which endpoint to call to, for example, “check inventory” or “register an order.” Precisely because SAP’s data model is standardized, AI can use its “tools” without hesitation.

The Benefits of Standardized Business Processes

What is most difficult for an AI agent is “work whose rules change with the mood of the moment.” In environments where SAP has been implemented, on the other hand, it is common for business processes to be standardized based on “best practices.”

Fields with Clear Rules Are Exactly Where AI Excels

Operations within SAP are governed by rigid business rules such as “document type,” “status management,” and “approval workflow.”

“If the status is A, then perform process B next.”

“If the value of this field is 1,000,000 yen or more, department manager approval is required.”

An environment where rules are documented this explicitly is, for an LLM that excels at logical reasoning, the most “conquerable field” in which it can deliver its best performance.

Synergy with the “Clean Core” Strategy

SAP’s recently advocated “Clean Core” strategy — a policy of leveraging standard functionality and minimizing add-ons — further eases the adoption of AI agents. The more standard a process is, the more accurately AI can perform the work on its behalf by drawing on its vast accumulated knowledge.

Column: The Importance of the Semantic Layer

Recently, attention has also turned to using a “semantic layer” that gives meaning (context) to SAP data. By having AI recognize data not simply as “Table: MARA” but as “material master data,” the reasoning accuracy of the agent improves dramatically.

The Potential and Limits of SAP Joule

SAP Joule is an AI assistant built into the entire SAP ecosystem, and it is currently rapidly strengthening its functionality as an “AI agent.” However, not every business task can be completed by Joule alone.

What SAP Joule Can Do: Joule’s greatest strength lies in the fact that it is natively integrated into each SAP module (S/4HANA, SuccessFactors, Ariba, and others).

The Convenience of a “Co-Pilot” Built into Standard SAP Functions

Understanding Context: AI is aware of which screen the user currently has open and what task they are performing. For example, on an inventory inquiry screen, simply asking “is this enough?” prompts it to answer by cross-referencing the current order status.
Inherited Security and Permissions: Because Joule directly inherits SAP’s role-based permissions (RBP), the risk of AI arbitrarily disclosing data a user should not be able to see is extremely low.

Simplifying Data Queries and UI Operations Through Natural Language

Transaction Shortcuts: There is no need to memorize complex T-codes (transaction codes) or menu hierarchies. Simply asking “turn last month’s Tokyo branch sales into a graph” instantly generates an analysis report.

Joule Studio & Agent Builder: A 2026 update has established a no-code environment for creating custom “AI skills” and “agents,” making it possible to train Joule on specific routine tasks.

Current Limitations and Challenges

On the other hand, for companies with large, complex system environments, there are “walls” that Joule alone cannot resolve.

Ability to Handle Customized Proprietary Add-Ons: SAP’s strength is flexible customization, but Joule still struggles to fully understand and operate proprietary add-ons (Z-functions) built up over many years, or business logic that deviates significantly from the standard.

In environments where “Clean Core” is not maintained, the “standard pattern” that Joule should reference has been eroded, which can result in responses and executions that fall short of the expected accuracy.

Constraints in Complex Integration with External Tools Outside SAP (SaaS, Proprietary Databases)

For many companies, business operations are not completed within SAP alone. Salesforce, Google Drive, in-house core systems, or old on-premises legacy systems — in an environment where these coexist, performing “extremely complex orchestration” that spans across systems can sometimes exceed the flexibility offered by Joule’s standard connectors alone.

In particular, when it comes to advanced reasoning processes such as deeply interpreting unstructured data outside SAP (such as PDF contracts) using an advanced reasoning engine like Claude, and then reflecting the results back into SAP, external AI agent platforms currently have the advantage.

Summary: Joule or a Custom Agent?

Joule is best suited as “a tool that maximizes efficiency for work within SAP.” However, for companies aiming for “end-to-end” automation that goes beyond company boundaries — optimizing the supply chain or incorporating integration with complex external data — “differentiating between” and “coexisting with” a custom AI agent built on an external LLM such as Claude is the practical answer as of 2026.

The Complementary Relationship Between Claude (LLM) and External Agent Platforms

Why Are Custom AI Agents Necessary, Not Just SAP Joule?
- Complex exception handling and judgment using Claude’s advanced reasoning capabilities.
Architectural Thinking
- Operating core systems through “multi-agent” setups using Claude combined with LangGraph, CrewAI, and similar frameworks.

Future Potential: Toward the Autonomous Enterprise

From “AI That Operates” to “AI That Optimizes Management”
- A world where real-time management decisions and execution connect seamlessly.
The Future of Agent Orchestration
- An organizational vision in which multiple AI agents collaborate across departmental boundaries.

First Steps Toward Implementation

The importance of starting small.
Considering governance and security (how much authority to grant AI).

About the author — Zhang (Data & Analytics)

Leads data utilization and AI adoption, from analytics platform design to implementing and validating AI agents in business operations.

Have a question about this article?

Ask the author directly — no sales pitch, just an answer.

Ask about this article →

AI Agents: The Potential to Replace Human Work

What Is the AI Agent Adoption Rate?

AI Agent Adoption Overseas