Latest survey data from more than 1,250 development teams reveal an amazing reality: 55.2% plan to build more complex proxy workflows this year, but only 25.1% succeed in deploying AI applications to production. The gap between ambition and implementation highlights the key challenges of the industry: How do we effectively build, evaluate and scale an increasing number of autonomous AI systems?
Let’s focus on the actual implementation challenges and the scope of capabilities the development team is navigating, rather than debate the abstract definition of “agent”.
Understand the autonomous framework
Similar to self-driving cars making progress through defined levels of capability, AI systems follow a trajectory of development, each level based on previous functions. This six-level framework (L0-L5) provides developers with a practical lens to evaluate and plan their AI implementation.
- L0: Rules-based workflow (follower) – Traditional automation with predefined rules without real intelligence
- L1: Basic Responder (Executor) – Reactive Systems that process input but lack memory or iterative reasoning
- L2: Using Tools (Actors) – A system that actively decides when to call external tools and integrates results
- L3: Observation, Planning, ACT (Operator) – Multi-step workflow with self-assessment function
- L4: Fully autonomous (explorer) – a persistent system that independently maintains state and triggers actions
- L5: Total Creativity (Inventor) – Systems that create novel tools and methods to solve unpredictable problems
Current Implementation Reality: Where are most teams today
Implementation reality reveals a sharp contrast between the theoretical framework and the production system. Our survey data shows that most teams are still in the early stages of implementation maturity:
- 25% of strategy development is still in place
- 21% are building proof of concept
- 1% testing in a test environment
- 1% of people have deployed production
This distribution emphasizes the practical challenges from concept to implementation, even at lower levels of autonomy.
Technical challenges at the autonomous level
L0-L1: Foundation building
Today, most production AI systems run at these levels, with 51.4% of the team developing customer service chatbots and 59.7% focusing on document parsing. At this stage, the main implementation challenge is integration complexity and reliability rather than theoretical limitations.
L2: The current frontier
This is where cutting-edge development is currently available, with 59.7% of teams using vector databases to root their AI systems in factual information. Development methods vary greatly:
- Build 2% using internal tools
- 9% leverage third-party AI development platforms
- 9% purely rely on rapid engineering
The experimental nature of L2 development reflects evolving best practices and technical considerations. The team faced significant implementation barriers, with 57.4% of barrier management being their primary concern, followed by use cases (42.5%) and technical expertise gap (38%).
L3-L5: Obstacles in Implementation
Even with significant progress in model capabilities, basic constraints hinder progress towards higher levels of autonomy. Current models show a key limitation: they train data too much rather than show real reasoning. This explains why 53.5% of teams rely on timely engineering rather than fine-tuning (32.5%) to guide model output.
Things to note about technology stack
The technology implementation stack reflects current functionality and limitations:
- Multimodal integration: text (93.8%), file (62.1%), images (49.8%) and audio (27.7%)
- Model providers: OpenAI (63.3%), Microsoft/Azure (33.8%) and Human (32.3%)
- Monitoring method: On-premises solutions (55.3%), third-party tools (19.4%), cloud provider services (13.6%)
As the system becomes more complex and monitoring capabilities become more important, 52.7% of teams are now actively monitoring AI implementations.
Technical restrictions hinder higher autonomy
Even today’s most complex models show a fundamental limitation: they train data too much rather than show real reasoning. This explains why most teams (53.5%) rely on timely engineering (32.5%) to guide model output. No matter how complex your engineering is, the current model is still struggling with real automatic reasoning.
The technology stack reflects these limitations. While multimodal features are growing (in 93.8% content, 49.8% images and 27.7% audio), base models from OpenAI (63.3%) (63.3%), Microsoft/Azure (33.8%) (33.8%), basic constraints of anthropomorphism.
Development methods and future directions
For today’s development teams that build AI systems, data generates some practical insights from data. First, collaboration is crucial – effective AI development involves engineering (82.3%), subject matter experts (57.5%), product teams (55.4%) and leadership (60.8%). This cross-functional requirement makes AI development fundamentally different from traditional software engineering.
Looking to 2025, the team is setting an ambitious goal: 58.8% plan to build more customer-facing AI applications, while 55.2% are preparing for more complex agency workflows. To support these goals, 41.9% focus on improving team skills, while 37.9% are building organization-specific AI for internal use cases.
Monitoring infrastructure is also evolving, with 52.7% of teams currently monitoring AI systems in their production. Most (55.3%) use in-house solutions, while others leverage third-party tools (19.4%), cloud provider services (13.6%), or open source monitoring (9%). As the system becomes more complex, these monitoring functions will become more and more important.
Technology roadmap
As we look forward, progress towards L3 and beyond will require basic breakthroughs rather than gradual improvements. However, the development team laid the foundation for a more autonomous system.
For building teams at higher levels of autonomy, the focus areas should include:
- A powerful evaluation framework This goes beyond manual testing and can programmatically verify the output
- Enhanced monitoring system Can detect and deal with unexpected behaviors in production
- Tool Integration Mode Allows AI systems to interact securely with other software components
- Inference verification method Distinguish between true reasoning and pattern matching
Data show that competitive advantage (31.6%) and efficiency improvement (27.1%) have been achieved, but 24.2% of the teams have no measurable impact yet. This highlights the importance of choosing the right level of autonomy for your specific technical challenges.
As we enter 2025, the development team must maintain a pragmatic attitude while experimenting with the model to achieve a model that can achieve more autonomous systems in the future. Understanding the technical capabilities and limitations of each autonomous level will help developers make informed building decisions and build AI systems with real value, not just technical novelty.
