Real-World AI Agent Examples in Production Today

The conversation around AI agents shifted in 2025. Instead of "what could agents do?" the question became "what are agents already doing?"

The answer: quite a lot. Klarna's AI assistant handled 2.3 million customer conversations in its first month. GitHub Copilot's coding agent now contributes to 1.2 million pull requests per month. Salesforce's Agentforce platform hit $1.4 billion in ARR with 18,500 customers.

But the picture is more nuanced than the headlines suggest. Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by end of 2026, up from less than 5% in 2025. That same report warns that over 40% of agentic AI projects will be canceled by 2027. The gap between a successful agent deployment and a failed one comes down to scope, architecture, and knowing when humans still need to be in the loop.

Here are the deployments worth studying.

Klarna's support agent is the most cited production deployment for good reason. Built on OpenAI's models, it connects directly to Klarna's billing and transaction systems via secure APIs. When a customer asks about a missing payment, the agent looks up the specific transaction, checks the status, and either provides an update or initiates a trace.

The numbers from the first month were striking:

• Handled two-thirds of all customer service chats (2.3 million conversations)
• Cut average resolution time from 11 minutes to under 2 minutes
• Customer satisfaction scores matched human agents
• Projected $60 million in annual cost savings

Then came the correction. In May 2025, Klarna shifted back to a hybrid human-AI model after discovering that complex edge cases, emotionally charged disputes, and multi-step financial issues still needed human judgment. The lesson: agents handle volume well but struggle with cases that require empathy or creative problem-solving.

Klarna's NPS score sits at 73, which suggests customers are generally satisfied with the hybrid approach. The agent handles the routine work. Humans handle the exceptions.

Coding agents have moved well past autocomplete. Two deployments stand out for their scale and measurable impact.

Cognition's Devin

Devin operates as an autonomous software engineer. Give it a feature request, and it explores the codebase, writes an implementation plan, codes the solution, runs tests, and opens a pull request. Customers include Goldman Sachs, Dell, Cisco, and Nubank.

The performance trajectory tells the story. In 2024, only 34% of Devin's pull requests were merged. By mid-2025, that number hit 67%. Nubank used Devin for a large-scale monolith refactoring project and reported an 8x efficiency improvement over their previous approach. Cognition's ARR grew from $1 million in September 2024 to $73 million by June 2025.

GitHub Copilot Agent Mode

Copilot is deployed at roughly 90% of Fortune 100 companies, with 4.7 million paid subscribers as of January 2026. The agent mode contributes to approximately 1.2 million pull requests per month. Across all users, Copilot generates an average of 46% of code written, rising to 61% for Java developers.

Both tools follow the same architectural pattern: observe the codebase, plan the implementation, execute with tool use (file reads, writes, terminal commands), and reflect on test results before submitting.

Give Your Agents a Workspace That Keeps Up

Fast.io provides intelligent, persistent storage for AI agent workflows. 50GB free, MCP-native access, built-in RAG, and no credit card required. Built for agent examples real world workflows.

Sign Up Free

The enterprise platforms are building agent infrastructure, not just individual agents. Two approaches are worth comparing.

Salesforce Agentforce

Agentforce launched as Salesforce's fastest-growing product, reaching nearly $1.4 billion in ARR with 114% year-over-year growth. The platform has 18,500 customers and has processed over 3.2 trillion tokens.

The IRS Office of the Chief Counsel provides a concrete example. Their Agentforce deployment reduced manual activities in tax court case processing by 98%, cutting the time to open a case from 10 days to 30 minutes. That's the kind of result that justifies agent adoption: a well-scoped, repetitive, document-heavy process where the rules are clear but the volume is crushing.

Google Gemini Enterprise

Google took a different approach with Gemini Enterprise, launched in October 2025 at published pricing per month. Rather than selling individual agents, Google provides a no-code workbench for building custom agents alongside pre-built agent "taskforces" for research and data analysis. The platform unifies LLMs, agents, and enterprise data under a single interface.

The contrast matters. Salesforce targets CRM-adjacent workflows where agents act on structured data. Google targets knowledge work where agents need to reason across unstructured documents.

The architectural pattern that emerged in 2025 is orchestrated teams of specialized agents rather than a single all-purpose agent. Gartner reported a 1,445% surge in multi-agent system inquiries from Q1 2024 to Q2 2025.

A typical production setup looks like this:

1. Task decomposition: A coordinator agent breaks a complex request into subtasks
2. Specialized execution: Each subtask routes to a purpose-built agent (one for research, one for writing, one for code review)
3. Shared state: Agents read and write to a common workspace so they can build on each other's output
4. Human checkpoints: Critical decision points route to humans for approval before the pipeline continues

The protocol layer is standardizing around three standards. Anthropic's Model Context Protocol (MCP) defines how agents access tools and external resources. Google's Agent-to-Agent protocol (A2A) handles peer-to-peer agent communication. IBM's Agent Communication Protocol (ACP) adds governance for enterprise deployments.

Popular frameworks for building these systems include LangGraph, CrewAI, AutoGen, Google ADK, and the OpenAI Agents SDK. Each makes different tradeoffs between flexibility and ease of use.

The shared state layer is where most teams underestimate the complexity. Agents need persistent storage that supports concurrent access, versioning, and permissions. A file lock prevents two agents from overwriting each other's work. An audit trail shows which agent made which change and when. Without this infrastructure, multi-agent systems produce inconsistent, unreproducible results.

Not every deployment succeeds. Gartner predicts that over 40% of agentic AI projects will be canceled by end of 2027. Understanding why helps you avoid the same mistakes.

Scope creep kills agents first. The IRS case works because it targets one specific process with clear rules. Companies that try to build a "general-purpose business agent" almost always fail. The agent that handles refunds, writes marketing copy, and manages inventory doesn't exist in production.

Cost surprises come second. Token costs for complex multi-step reasoning add up fast. An agent that makes 50 tool calls per task at enterprise scale can generate bills that dwarf the human labor it was meant to replace. McKinsey found that 23% of organizations are scaling agentic AI successfully, but another 39% are still experimenting, often stuck on cost-benefit calculations.

Governance gaps come third. When an agent autonomously processes a refund, updates a database, or sends a message to a customer, who is responsible if something goes wrong? Companies without clear accountability frameworks stall at the pilot stage.

The pattern among successful deployments is consistent: narrow scope, measurable outcomes, human oversight on edge cases, and infrastructure that provides audit trails for every agent action.

Every production agent deployment described above has one thing in common: the agents need somewhere to store their work. Without persistent storage, an agent starts from scratch every session. It can't build on yesterday's analysis, share context with other agents, or hand off completed work to a human reviewer.

Most teams start with local file systems or S3 buckets. These work for prototypes but break down when you need versioning, permissions, concurrent access from multiple agents, or semantic search across the files agents produce.

Fast.io approaches this differently. Instead of treating storage as a dumb file system, Fast.io provides intelligent workspaces where files are automatically indexed for semantic search the moment they're uploaded. Agents and humans share the same workspaces and intelligence layer. Humans use the web UI; agents use the Fast.io API or MCP server.

Key capabilities for agent workflows:

• MCP-native access: Fast.io exposes Streamable HTTP at /mcp and legacy SSE at /sse, giving agents 19 consolidated tools for workspace, storage, AI, and workflow operations
• Built-in RAG: Enable Intelligence Mode on a workspace and agents can run semantic search with citations across all indexed files, no separate vector database needed
• Ownership transfer: An agent creates workspaces, organizes deliverables, then transfers ownership to a human client. The agent retains admin access for ongoing maintenance
• File locks: Acquire and release locks to prevent conflicts when multiple agents write to the same workspace
• Webhooks: Get notified when files change so agents can react without polling
• URL Import: Pull files from Google Drive, OneDrive, Box, and Dropbox via OAuth without local I/O

The free agent tier includes 50GB storage, 5,000 credits per month, 5 workspaces, and no credit card requirement. For teams building multi-agent systems, the shared workspace model means agent output becomes team output without manual file transfers or copy-paste workflows.