Truly Useful AI: Reviewing Pascal Bornet's Five-Level Guide to Getting Stuff Done with Agents
- Nancy Wang
- Sep 12
- 7 min read
Updated: 4 days ago
In the fast-evolving AI landscape, too many organizations remain stuck in a "brilliant advisor trap" – leveraging artificial intelligence to think and analyze, but failing to harness its power for getting stuff done in the real-world.
So argues Agentic Artificial Intelligence: Harnessing AI Agents to Reinvent Business, Work and Life by Pascal Bornet and colleagues, a hefty volume (Irreplaceable Publishing, 2025) which addresses this critical gap with a five-tiered framework grounded in practical guidance for using the right type of agentic capabilities for different business challenges.
Bornet's team of 27 co-authors spans business, academic and research roles, bringing decades of collective experience implementing AI solutions across industries. But what sets this work apart is its real-world grounding: the authors discuss their successes and failures with refreshing honesty that shuns hype and hyperbole.
From this sprawling review emerges a powerful lens that brings into focus the evolution of AI from basic process automation to the sophisticated agentic systems we are starting to see today – along with a compelling reminder to appropriately scope solutions.
This core insight validates what we've observed at Agentic Foundry: successful AI transformation requires matching intelligence levels to specific challenges, and shaping better ways to collaborate.
Agent Capabilities: The SPAR Framework
Before diving into the five-level progression, the authors introduce a crucial conceptual foundation: the SPAR framework (Sense, Plan, Act, Reflect).
This framework provides a systematic way to understand how AI agents operate in their environments, much like how we might evaluate a new human colleague's capabilities.
Sensing refers to an agent's ability to perceive and gather information from its environment – from reading documents to monitoring alerts and user inputs.
Planning encompasses the agent's reasoning capabilities: how it processes information, evaluates options and develops strategies to achieve its goals.
Acting involves the agent's capacity to execute tasks in the real world – sending messages, updating databases, triggering workflows or controlling physical systems.
Reflecting captures the agent's ability to learn from experience, evaluate its performance and adapt its future behavior.
The SPAR framework serves as a diagnostic throughout the book, helping readers understand how these capabilities evolve across the five intelligence levels.
More importantly, it provides a practical lens for evaluating any AI agent system: rather than getting lost in technical specifications, decision-makers can ask focused questions about each component and how well they align with their specific needs.
Five Levels: A Taxonomy for Autonomy
The book's most significant contribution lies in its Agentic AI Progression Framework – a five-level taxonomy that maps the evolution from basic automation to full autonomy.
Unlike traditional maturity models where higher levels are always better, this framework recognizes that different organizational needs require different levels of AI capability. A Level 2 solution that reduces processing costs by 40% may deliver far more value than a Level 4 system that never successfully deploys.
The authors shrewdly use the analogy of automotive assistance technologies: while fully autonomous driving might be technically possible, many drivers prefer the predictability and control of basic cruise control.
Level 0: Manual Operations
At Level 0, humans perform all tasks without technological assistance. This represents the traditional workplace where employees rely on basic digital tools like spreadsheets and email but handle all processing manually.
The authors use this as the baseline to demonstrate the transformative potential of each subsequent level. While seemingly primitive, Level 0 operations remain critical in many contexts requiring human judgment, creativity and emotional intelligence.
For many organizations even today, this also marks the reality.
Level 1: Rule-Based Automation
Level 1 introduces basic automation following fixed rules, comparable to simple cruise control that maintains speed but requires human oversight for all other driving functions. Consider advanced macros or sophisticated if-then logic.
In business contexts, this includes basic robotic process automation (RPA), simple scripts and rule engines that can handle data entry and straightforward workflows.
The authors emphasize that Level 1 systems excel at repetitive, high-volume tasks with clear business rules. They provide the crucial foundation for higher levels of automation by establishing process standardization and data quality.
However, Level 1 systems are brittle – they break when encountering exceptions or scenarios outside their programmed parameters.
The SPAR capabilities at this level are rudimentary: basic sensing through predefined inputs, simple planning following predetermined steps, deterministic acting and minimal reflection limited to basic error reporting.
Level 2: Intelligent Process Automation
Level 2 represents a quantum leap, combining traditional automation with cognitive abilities like natural language processing, machine learning and computer vision.
The authors compare this to advanced driver assistance systems that can handle both speed and steering but still require human supervision. Level 2 systems can make decisions based on data patterns, handle semi-structured information and adapt to variations in input.
Such systems might excel in scenarios like invoice processing, where they extract information from documents, validate against business rules and route for human approval. This level marks the first wave of "intelligent automation" the authors pioneered in their consulting work.
The SPAR capabilities become more sophisticated: semi-structured sensing that can interpret various data formats, basic decision-making planning that weighs multiple factors, more nuanced acting with conditional logic and improved reflection through performance monitoring and basic learning loops.
The authors note that many organizations have found their "sweet spot" at Level 2, achieving significant productivity gains while maintaining human oversight for complex decisions.
Level 3: Agentic Workflows
Level 3 represents the emergence of true AI agents – systems that can generate content, plan multi-step processes, reason through complexity and adapt based on context.
The automotive analogy here is autonomous highway driving: the system can navigate complex situations independently but may need human intervention in truly unusual circumstances.
Level 3 agents might understand natural language instructions like "improve customer satisfaction through better service delivery" and translate this into executable workflows.
In these situations, agents demonstrate sophisticated reasoning capabilities, can chain multiple tools together and maintain context across extended interactions. This is where the authors' expertise in large language models and modern AI becomes evident.
The SPAR framework reaches maturity at Level 3: sophisticated sensing with contextual awareness, complex planning involving multi-step reasoning and resource orchestration, dynamic acting through tool chaining and adaptive execution and meaningful reflection through limited learning and long-term memory storage.
The authors provide compelling examples of Level 3 agents handling end-to-end customer service scenarios, conducting market research and managing complex approval workflows with minimal human oversight.
At the same time, the authors are honest about Level 3 limitations. These systems can still make errors in edge cases, may struggle with tasks requiring deep domain expertise and need careful monitoring to ensure alignment with organizational values and objectives.
Level 4: Semi-Autonomous Systems
Level 4 systems approach true autonomy, capable of setting their own sub-goals, learning from experience and adapting their strategies over time.
The authors compare this to fully autonomous driving in most conditions, requiring human intervention only in extraordinary circumstances.
While acknowledging that Level 4 systems are still largely experimental, the authors provide fascinating glimpses of their potential. These systems can conduct their own research, develop strategies, execute complex multi-phase projects and continuously improve their performance through experience.
A fascinating insurance company case study hints at Level 4 capabilities where AI agents not only process claims but proactively identify fraud patterns, optimize workflow efficiency and suggest policy improvements.
SPAR capabilities at Level 4 represent a qualitative leap: comprehensive sensing across multiple data streams and environments, strategic planning with long-term goal optimization, autonomous acting with self-correction capabilities and deep reflection through continuous learning and strategy refinement.
Level 5: Fully Autonomous Systems
Level 5 represents the theoretical endpoint: fully autonomous systems that can operate independently across all scenarios within their domain.
Like fully autonomous vehicles that never require human intervention, Level 5 AI agents would handle complete business functions from strategy through execution.
The authors approach Level 5 with appropriate caution, noting that such systems remain largely theoretical. The challenges are not just technical but involve fundamental questions about accountability, control and the role of human judgment in complex decisions.
In addition the authors agree that Level 5 may be more relevant for specific, well-defined domains rather than general business operations.
The Value of A Portfolio Approach
What makes this five-level framework useful is its recognition that organizations need different levels for different functions. As we have also seen on more than a few occasions, when organizations make the leap into AI, they often assume more is always better.
The authors reveal that successful AI transformations typically involve what we might call a portfolio approach: Level 1 and 2 systems handling routine operations, Level 3 agents managing complex workflows and careful experimentation with Level 4 capabilities in appropriate contexts.
This thinking provides a springboard for capability building. Companies can't jump directly from manual operations to Level 3 agents – they need to build the process discipline, data quality and change management capabilities that come with implementing Levels 1 and 2 first.
Building on their five-level framework, the authors explore the critical implementation factors that determine success, drawing on their SPAR framework to detail a complementary lens for understanding how AI agents operate at each level.
The exploration of the "Three Keystones" – Action, Reasoning and Memory – explains why many AI implementations fail despite impressive benchmark scores, and underscores the importance of a portfolio approach that spreads bets on a number of ideas.
To wit, the authors make a compelling case that traditional AI benchmarks like HumanEval (measuring task execution) and MMLU (assessing knowledge breadth) are like academic credentials – they tell you something important about capability, but they don't predict real-world performance.
An AI system might score 95% on reasoning tests yet struggle to remember customer preferences across interactions, execute actions reliably in complex business systems or adapt its reasoning to unusual but important edge cases.
Without all three keystones working in harmony, even the most sophisticated AI becomes like a brilliant graduate who can ace exams but can't handle the messy realities of actual work.
This insight helps explain why so many organizations see impressive demos and pilot results, only to struggle with production deployments that feel brittle and unreliable..
Implications and Horizons
"Agentic Artificial Intelligence" provides a pragmatic framework for understanding agentic capabilities as well as practical guidance for implementation. The five-level taxonomy represents a notable contribution that offers a portfolio approach or AI transformation.
The authors' recognition that different levels serve different purposes – and that higher levels aren't always better – reflects the mature perspective that comes from hands-on implementation experience.
Their detailed exploration of each level's capabilities, limitations and appropriate use cases provides invaluable, hands-on guidance for organizations navigating the complex landscape of AI agent technologies.
As AI agents transition from prototypes to business-critical systems, Bornet and colleagues have provided a conceptual foundation for leaders to evaluate investments and build organizational capabilities systematically.
At Agentic Foundry, these ideas aren’t just theoretical — we’re pressure-testing them daily. This work is about connecting data and workflows to get stuff done, it’s about building a suite of agents that solve problems today.
Agentic Foundry: AI For Real-World Results
Learn how agentic AI boosts productivity, speeds decisions and drives growth
— while always keeping you in the loop.
