From: aidotengineer
Hbert Mishtella, who leads AI researchers in drug design at Novardis, presented at the AI Engineer World’s Fair 2025 on the topic of “Agentic Enterprise: What your CEO must know about AI” [00:00:00]. The presentation covered various angles of AI, including workflows, networks, and individual perspectives, focusing on how organizations can prepare for the future impact of AI agents [00:00:37].
Organizations may be run by AI agents in as little as three years, a statement considered realistic given the current pace of AI development [00:00:50].
Defining AI Agents
An AI agent is an application based on a large language model (LLM) that exhibits a degree of autonomy in its behavior [00:01:08]. They use an LLM to “glue together” specific functions, executing steps with planning and dynamically adjusting tool usage based on previous outcomes [00:01:22]. This capability provides a new level of functionality previously almost exclusive to humans [00:01:45].
To leverage AI agents, companies should work on evolving their digital assets to be used as tools for or even as AI agents themselves [00:01:55].
Key Capabilities of AI Agents
AI agents offer several capabilities that enable complex tasks:
- Multi-step Planning and Reasoning [00:02:26]
- Adaptability for Dynamic Work [00:02:28]
- Persistent Memory [00:02:32]
- Contextual Knowledge (e.g., Retrieval Augmented Generation or RAG for specific knowledge not in the LLM) [00:02:35]
- Interactive Workspaces (e.g., sandboxes or canvases for iterative collaboration on tasks like code or art) [00:02:46]
- Computer Vision (allowing agents to navigate graphical user interfaces and automate steps previously done by humans or robotic process automation) [00:03:02]
Workflow Perspective: The Agentic Workflow
A workflow is the execution of a defined process, often involving sequential steps that can be commoditized or specialized [00:03:32]. Technology can be used for automation (substituting a step) or augmentation (doing a step quicker or better) [00:04:15]. While machine learning could achieve some of this before, AI agents introduce an “agentic workflow” [00:04:45].
AI agents differ because they can:
- Tackle a broader set of tasks by operating on cognitive steps, often represented in natural language [00:05:04].
- Act as a “glue” between tasks, composing them together without human intervention, requiring human feedback only when necessary [00:05:20].
This enables the automation and delegation of multiple steps simultaneously [00:05:43].
Key Takeaway: To start building agents, organizations must first identify their workflows and understand the exact context around them [00:05:55]. This context, including data, systems, transformations, and steps, is often unrecorded and exists only in employees’ heads [00:06:11].
Network Perspective: Personas and Transformation
Relying solely on traditional roles and job titles is insufficient for applying AI agents because it doesn’t convey the tools, planning, or behavioral conditions needed [00:07:50]. A more helpful approach is the network perspective, where each contributor to a workflow is viewed as a “persona” [00:08:28].
Examples of personas include:
- Silent Achiever: Low communication, high performance [00:08:57].
- Individual Contributor: Good communication and performance [00:09:04].
- Connector: Links non-adjacent teams [00:09:09].
- Multiplier: Enhances others’ work [00:09:16].
- Knowledge Hub: Provides domain expertise [00:09:21].
Defining personas helps project how AI agents might impact workflows and employees [00:09:39]. For example, a silent achiever with a coding assistant might become so productive they can handle tasks previously requiring multiple people [00:11:18]. An AI agent could magnify a multiplier’s impact across an entire team, or largely substitute a knowledge hub persona by dynamically supporting all team members [00:11:56]. This leads to overall efficiency [00:13:11].
Key Takeaway: Looking at personas is a helpful first step for organizations and people to adapt to AI, making it easier to project and plan for the development of different agents [00:10:40].
Emerging Patterns and New Roles
When considering how humans and AI agents can best perform tasks, ambiguous workflows or those requiring subjectivity are better suited for humans [00:14:42]. Repetitive, tedious tasks or those requiring perfect, consistent execution are ideal for AI agents [00:15:08].
Emerging patterns include:
- Commoditization of Intelligence and Domain Knowledge: These become cheap and easily available, no longer a primary edge for employees or companies [00:15:32].
- Shift to Multi-disciplinary Know-how or Deeper Specialization: Employees and companies must pivot to these areas [00:16:06].
- Context Understanding Gains Value: In contrast to intelligence or domain expertise [00:16:46].
- AI Supercharges Doers: Individuals who perform and do can achieve significantly more (e.g., 100 times engineer) [00:17:04].
- Polarization in Positions: Leading to new specializations and less middle ground [00:17:27].
New roles might appear in AI-mature organizations, such as:
- Workflow Miner [00:17:48]
- Human-AI Orchestrator (managing human-AI interaction) [00:17:55]
Enabling Organizational Adoption and Democratization
To democratize and adopt AI agents, organizations can follow a step-by-step approach:
- Access to General Assistants: Providing basic access, which is likely already in place [00:18:32].
- Individual Assistants: Offering personalized assistants per employee with memory, context understanding, and personal databases [00:18:39].
- Employee-Built Agents: Enabling employees to build agents quickly using no-code, low-code, or coding tools [00:19:10]. This requires cognitive self-awareness to translate mental steps into explicit, buildable processes [00:19:45].
- Digital Twins: Creating “time travelers” and memory to retain organizational knowledge, allowing an agent to represent an employee’s experience and provide access to past expertise [00:20:24].
- Multi-Employee Agents: Agents serving entire teams or workflows [00:21:06].
- Swarm of Agents: Advanced systems with multiple interacting agents [00:21:12].
This progression is now possible due to the maturing software stack around LLMs and AI agents [00:21:20].
Challenges and Opportunities
Challenges
Common challenges include groundedness, hallucinations, guardrails, and security [00:21:42]. Two critical challenges are:
- Adoption: The social readiness for AI agents to drive interactions and delegate subtasks to humans, rather than just being delegated to [00:22:15].
- Ethics of Autonomy: Defining explicit values for AI agents to uphold and resolving conflicts between human and AI agent decisions [00:22:59]. As John Lennox stated, “the greater the machine’s freedom, the more it will need a moral standards” [00:24:12]. This necessitates engaging with morality, ethics, and philosophy to operate technology [00:24:23].
Opportunities
Beyond addressing challenges, AI agents present significant opportunities for executives:
- Simulate Synthetic Reality: Before real-world activities in marketing, labs, or other domains [00:24:53].
- Proactively Identify Regulatory Loopholes: Acting as “loophole lighthouses” [00:25:06].
- Scout M&A Opportunities: Finding mergers and acquisition opportunities [00:25:13].
Executive Summary
A key indicator of an outdated strategy is if it still makes sense after swapping “AI agent” with terms like “machine learning” or “data” [00:26:27]. AI agents introduce new capabilities that drive significant change [00:26:56].
For organizations, critical aspects include:
- Detailed Workflow Understanding: Crucial for maximizing benefits from agents [00:27:00].
- Democratization of Agents: Essential to fully accelerate the AI agent revolution [00:27:05].
- Human Aspect: Looking at the network of personas is vital [00:27:13].
- Rapid Transformation Readiness: Organizations must be prepared for swift changes, including drastic shifts in organizational charts [00:27:19].
- Mindset Shift: Redefining and adapting to new interactions with technology, even when AI agents lead [00:27:31].
- Ethics: Remaining a crucial part of the development and application of AI agents [00:27:47].