NEWS June 27, 2026
✍️ Carlos Montiel
⏱ 10 min read
In 2022, chatbots answered questions. In 2024, they could hold conversations. In 2026, autonomous AI agents complete end-to-end tasks that used to require human teams. Klarna replaced the work of 700 support agents. JP Morgan saves 360,000 hours of legal work a year. How do you bring this to your company?
The Fundamental Difference: Chatbot vs. Agent
| Dimension | Chatbot | Autonomous Agent |
|---|
| Interaction mode | Reactive — only responds | Proactive — initiates actions |
| Duration | A single conversation | Hours or days if needed |
| Tools | None or very limited | APIs, databases, code, emails, web |
| State | Stateless — forgets between sessions | Stateful — persists context and progress |
| Decision-making | Makes no decisions | Decides, self-corrects, escalates when needed |
| Output | Text on screen | Real changes in external systems |
The 4 Levels of Autonomy in 2026
1
Assistant
Answers questions, generates text, classifies information. Doesn't execute actions. Example: basic ChatGPT answering support questions.
2
Copilot
Suggests concrete actions, the human approves and executes each step. Example: GitHub Copilot suggesting code, the developer applies it manually.
3
Supervised Agent
Executes tasks with periodic supervision. The agent works, the human reviews milestones. Example: Claude Code refactoring an entire module, the developer reviews at the end.
4
Autonomous Agent
Completes end-to-end workflows, only escalates edge cases to a human. Example: Salesforce Agentforce resolving 84% of tickets with no human intervention.
Most companies in 2026 operate at Level 2–3. Level 4 is being tested in specific use cases with human escalation mechanisms as a safety net.
Real Cases of Agents in Production
Salesforce Agentforce
84%
Autonomous resolution rate across 380,000+ customer support interactions
Klarna (Fintech)
2.3M
Conversations in the first month. Equivalent to 700 human support agents
JP Morgan COiN
360K hrs
Hours of legal work saved per year by automatically reviewing contracts
Duolingo
40
Languages served simultaneously by agents that generate personalized exercises
A 5-Step Agentic Strategy for LatAm Companies
- Identify candidates: Processes with high repetition, clear rules, and existing digital data. Ideal examples: replying to support emails, processing HR requests, generating periodic reports
- Start supervised (Level 3): The agent proposes, the human approves. Collect quality data for 2–4 weeks to understand where the agent gets it right and where it fails
- Define autonomy thresholds: If the agent has over 95% approval in a task category, consider giving it full autonomy for that specific category
- Implement human-in-the-loop: For high-impact decisions (above a certain amount, affects multiple customers, requires human expertise) always keep an escalation step
- Measure real ROI: Resolution rate, average time per task, cost per interaction, end-user satisfaction. Without metrics you can't justify expansion or spot problems
The Most Underestimated Risk: Agent Identity
Only 22% of organizations treat AI agents as entities with formal identity and access controls (Gartner 2026). This means that at 78% of companies, AI agents operate with a human user's permissions or with excessive access to critical systems.
Best practice: every agent should have its own identity (service account), the minimum permissions necessary for its task, its own audit logs, and a human owner responsible for its behavior. Treat the agent like a new employee who needs onboarding and supervision.
Ready to take an AI agent to production?
Carlos Montiel designs agentic strategies, implements agents with LangGraph and LangChain, and guides you from pilot to production for companies in Guatemala and Latin America.
Design my agentic strategy
Carlos Montiel
Enterprise AI Solutions Architect · guatemalia.com
Implements autonomous agents with LangGraph, LangChain, and Claude for companies in Guatemala and Latin America. Contact: guatemalia.com/en/#contact