The year 2026 marked a shift from the "initial year of AI agents" to the "year of AI agent execution." Salesforce's Agentforce exceeded $800M in ARR (up 169% year-over-year) and 29,000 adopting companies, while Gartner predicts that "by the end of 2026, 40% of enterprise applications will embed task-specific AI agents."
On the other hand, only 11% of companies run AI agents in production environments, making the polarization between companies that "experimented without seeing results" and those "building competitive advantage through company-wide rollouts" strikingly clear. Based on the latest survey data, this article unpacks the structure of this polarization and outlines actionable steps for small and medium-sized enterprises.
What Are AI Agents? The Decisive Difference from Chatbots
First, let us establish the definition of an "AI agent." An AI agent is an AI system that autonomously gathers information, makes decisions, and executes actions toward a given goal.
While conventional chatbots operate as passive mechanisms that "wait for user input and respond based on preconfigured rules or training data," AI agents feature the following three characteristics:
- Autonomy: Once a goal is set, it independently gathers necessary information, formulates a plan, and executes multiple steps
- Adaptability: Able to adjust actions in response to environmental changes or errors
- Tool integration: Connects with external APIs, databases, and SaaS platforms to complete operations on behalf of humans
For example, whereas a traditional chatbot merely returns a figure when asked "What are this month's sales?", an AI agent autonomously handles complex composite tasks such as "analyzing why this month's sales fell below target, summarizing remedial measures, and posting them to Slack."
Adoption Status Shown by Recent Surveys: 2026 in Numbers
Here is an overview of data as of 2026 published by major research organizations.
Gartner's Projections
- By end of 2026: 40% of enterprise applications will feature task-specific AI agents (up from less than 5% in 2025)
- By 2028: At least 15% of day-to-day business work decisions will be made autonomously through agentic AI
- Through end of 2027: More than 40% of agentic AI projects will be canceled due to escalating costs, unclear business value, or inadequate risk controls
Salesforce's Performance
- Agentforce ARR (annual recurring revenue) reached $800M (a 169% year-over-year increase)
- Number of adopting companies reached 29,000 (up 50% quarter-over-quarter)
- Average ROI for enterprise adopters reached 171%
Anthropic's Insights
- Advocated "simplicity," "transparency," and "thoughtful interface design" as the three principles of building AI agents
- While one agent per task was the mainstream in 2025, 2026 marks the full-scale arrival of "multi-agent" systems where multiple agents divide roles
Current State of Japanese Enterprises (PwC and MIC Surveys)
- Enterprises that have implemented generative AI account for about 25% of the total (1 in 4 companies)
- Adoption stands at 50% among companies with 10,000 or more employees, but remains at only 15.7% for those with under 1,000 employees
- Companies establishing generative AI usage policies reached 49.7% (up from 42.7% the previous year)
What these numbers demonstrate is the reality that while the technology itself is maturing rapidly, the gap continues to widen between organizations capable of mastering it and those that are not.
3 Reasons Enterprises Are Polarizing
Why is a massive performance gap opening between companies despite having access to the same technology? Let us examine the three structural factors that emerge from survey data.
Can You Overcome the Pain of Workflow Redesign?
AI agents do not deliver sufficient results simply by being "added" onto existing workflows. Because the business processes themselves must be redesigned, the workload on front-line teams temporarily increases during early adoption.
For example, introducing an AI agent into accounting operations requires decomposing the end-to-end flow—from receiving invoices to journal entries, payments, and reconciliation—and redesigning which stages to assign to AI and where humans should provide approvals.
Successful companies overcome this "redesign pain" through executive leadership commitment, whereas many enterprises stop at merely appending tools to existing flows, falling into a state of "we introduced it, but see no results." This is also the primary reason Gartner projects a project cancellation rate exceeding 40%.
Maturity of the Data Foundation
For AI agents to make decisions autonomously, accurate and structured data is essential. However, many Japanese companies struggle with the following challenges:
- Siloed data: Different departments use separate systems, fragmenting data
- Low data quality: Inconsistent terminology, missing values, and duplicate records remain unaddressed
- Lack of real-time capability: Batch processing dominates, preventing agents from referencing data immediately
Behind the contrast between large enterprises of 10,000+ employees reaching 50% AI adoption and SMEs lingering at 15.7% lies a substantial disparity in investment capacity for data infrastructure. However, with cloud-based data integration services becoming more affordable in recent years, phased implementation is becoming increasingly feasible for SMEs as well.
Presence or Absence of a Governance Structure
Because AI agents make decisions and take actions on behalf of humans, a governance framework to track and control "what the AI did" is indispensable.
Leading companies establish governance structures such as the following:
- Guardrails: Rules restricting the scope of execution to prevent the AI from performing unintended operations
- Draft & approve: A two-stage process where the AI creates drafts, and humans review and approve them before execution
- Audit logs: Mechanisms that record all AI operations so they can be verified later
Money Forward's "AI Cowork" has designed this aspect in a progressive manner, drawing attention as a governance model for AI agents. On the other hand, among companies that postpone governance design, adoption itself frequently halts due to concerns over security incidents and operational error risks.
Pioneering Case Studies Among Japanese Companies
Having understood the structure of this polarization, let us examine how frontrunner companies are putting AI agents into practice.
Money Forward "AI Cowork"
Scheduled for release starting in July 2026, "AI Cowork" is an AI agent service that autonomously executes back-office tasks. In response to natural language requests like "Process this month's accounting tasks together," multiple AI agents collaborate in parallel to complete the work, handling invoice generation, payment requests, receipt reconciliation, cash flow forecasting, and more. Money Forward has announced a target of ¥15 billion in AI-related ARR by 2030.
SoftBank × Seino Information Service "Logistics AI Agent"
Collaborating with SoftBank, Seino Information Service built an MVP embedding AI agent capabilities into its warehouse management system "SLIMS." AI assists from situational assessment through decision-making and action on the logistics floor, achieving major improvements in delivery efficiency. Against the backdrop of the 2024 problem (driver shortages), AI agent adoption in logistics is expected to accelerate further.
SoftBank "AGENTIC STAR"
SoftBank launched "AGENTIC STAR," an enterprise AI agent platform, in December 2025. It allows companies to use AI agents in a SaaS model to comprehend business goals and autonomously advance tasks in collaboration with team members; starting in March 2026, external connectivity and development platform provisioning models have also rolled out.
3 Steps SMEs Should Take Right Now
Some may feel that "it is still too early for us." However, as technology matures and services become more affordable, now is the prime opportunity for SMEs to establish the foundation for AI agent utilization.
Step 1: Inventory Operations and Identify Tasks to Delegate to AI
Begin by visualizing current business workflows and listing repetitive, standardized tasks with clear evaluation criteria. Candidates include tasks such as:
- Creating and sending invoices
- First-line response and triage of inquiry emails
- Drafting routine reports and status updates
- Data entry and transcription work
Rather than aiming for company-wide implementation all at once, starting with a single department and a single workflow is the key to success.
Step 2: Start with Small Improvements to Data Foundations
Large-scale data infrastructure projects are unnecessary. Start with these baseline improvements:
- Migrate to cloud tools: Consolidate paper records and local Excel files into the cloud
- Standardize naming conventions: Resolve naming discrepancies across filenames, client names, and product titles
- Select SaaS with open APIs: Prioritize services that publish APIs with future AI agent integrations in mind
These can be achieved using tools costing from a few thousand to tens of thousands of yen per month, serving as essential preparatory work before introducing AI agents.
Step 3: Experience Tangible Results with a Small PoC (Proof of Concept)
Using the identified tasks and refined data, execute a small 2- to 4-week PoC. The key is establishing the following metrics in advance:
- Time saved: How much time required for the target workflow decreased before and after introducing the AI agent
- Error rate: How error rates changed compared to when tasks were performed manually
- Cost: The balance between tool fees and reduced labor costs
Reporting PoC results with concrete numbers to leadership facilitates smooth investment decisions for the next phase.
Outlook for Late 2026 Through 2027
Finally, let us review upcoming trends.
Multi-Agent Systems Going Mainstream
While assigning a single task to a single AI agent was standard in 2025, "multi-agent systems" where multiple agents share roles and operate in parallel will become the norm from late 2026 through 2027. For example, specialized agents might handle legal checks, code generation, and email drafting respectively, while an orchestrator agent coordinates the entire process.
Rapid Rise of Industry-Specific Agents
Alongside general-purpose AI agents, agents deeply specialized in specific industries and functional domains will proliferate. Agents embedding domain knowledge for logistics, manufacturing, healthcare, legal, accounting, and other areas will emerge, lowering the barrier to adoption even further.
Shift to Performance-Driven AI Investment
Gartner notes that "from 2026 onward, AI investments will be evaluated under the same standards as other capital expenditures." The era of "adopting AI just to see what happens" is over, and we are entering an era where projects failing to show concrete ROI will see budgets cut. That is precisely why starting small and systematically stacking up tangible wins is the right approach.
Conclusion
In 2026, enterprise AI agent adoption shifted from "experimentation" to "execution." As illustrated by Gartner's projection of "over 40% of projects canceled," whether an enterprise can build an operational mechanism that delivers results—not merely adopting the technology—separates the winners from the losers.
The three barriers driving polarization—workflow redesign, data foundations, and governance—are not technology problems; they are organizational challenges. In other words, with executive commitment and a phased approach, SMEs can fully overcome them.
Start with a single workflow in a single department. Experience real results through small PoCs, accumulate structured data, and expand coverage. Building on these steps forms the foundation of competitiveness in the multi-agent era of 2027 and beyond.
For consultation on AI agent adoption and PoC support, feel free to reach out to GleamHub.
