AI Engineer World's Fair 2025

Agentic Enterprise: What Your CEO Must Know About AI

Loading slides…

Read the talk

Agentic Enterprise: What Your CEO Must Know About AI

Agents can connect tasks, reshape employee networks and eventually direct human work. Preparing for that shift starts with the context that job titles and process diagrams leave out.

From a talk by Hubert Misztela

Could agents run the organization?

What would an organization need to change if AI agents began running its work? At AI Engineer World’s Fair 2025, Hubert Misztela introduces himself as leading AI researchers in drug design at Novartis, then turns to this broader organizational question. His deliberately provocative premise is that an organization might be run by agents within three years—a possibility posed in 2025, not a demonstrated outcome. Answering it requires looking at workflows, employee networks, individual capabilities and the challenges of delegating authority.

0:000:12
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:00 · section reference included

Give agents access to the enterprise

An AI agent is an application built around a large language model that exhibits some autonomy. The model connects planning with tool use: it takes a step, observes the result and adjusts what it does next. This dynamic action matters because a useful plan cannot always specify every action before execution begins. A failed operation or an unexpected result may change which tool the agent needs next.

That capability changes how an enterprise should think about its existing digital assets. Misztela proposes making each asset available as an agent tool, a retrieval-augmented generation resource or an agent in its own right. The distinction is functional: some assets support actions, some supply knowledge, and some can participate in delegated work. The enterprise’s existing resources become part of what an agent can use, rather than remaining accessible only through an employee.

Diagram connecting an LLM to action, planning, memory, retrieval and tools, with Python, Web, Files, CRM and DB below. Two purple callouts define agents and describe adapting digital assets.Open full source frame
AI agent components and the proposal to turn enterprise digital assets into tools, RAG or agents.

Several capabilities support this arrangement:

  • Planning and adaptation: reason through multiple steps and adjust control as work progresses.
  • Memory and retrieval: retain context and bring in specific knowledge that the model does not already contain.
  • Interactive workspaces: collaborate iteratively in a sandbox or canvas to refine code, artwork or another deliverable.
  • Computer use: navigate graphical interfaces to perform operations previously handled by people or, in some cases, robotic process automation.

Together, these capabilities expand both the work an agent can attempt and the ways a person can collaborate with it.

1:071:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

1:07 · section reference included

Delegate connected steps, not just isolated tasks

Consider a process represented by five sequential steps, running from left to right. Some are routine; others require specialized work. This is the workflow: the execution of the process. Misztela tentatively credits the simple diagram to Sanjit and accompanying blog posts, then uses it to distinguish two familiar applications of technology.

ApplicationWhat changes at a step
AutomationTechnology performs the step in place of a person.
AugmentationA person performs the step faster or better with technology.

Neither application began with agents. Machine learning, and systems without machine learning, could already automate or augment particular tasks. The question is what agents add to that existing pattern.

Misztela identifies two changes. First, language models can address a broader set of cognitive tasks expressed in natural language. Second, they can connect those tasks, carrying work from one to the next without requiring a person to manage every handoff. Human feedback enters where the agent needs it. In the diagram, the pink region spans this connected work: an agentic workflow delegates several steps together instead of improving only one isolated operation.

Before building that agent, however, someone must recover the workflow’s actual context. Which data does the employee use? Which systems do they enter? What transformations do they perform, and what happens on the machine between the initial request and the finished result? Much of this knowledge is never logged. An employee invents a process under time pressure, repeats it later, and gradually turns it into a routine that exists mainly in their head. Other parts may be distributed across colleagues and systems. Workflow discovery and context discovery are prerequisites for useful delegation.

3:223:40
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:22 · section reference included

Map contributions that job titles hide

Now put people into the workflow: one or two contributors at each step, plus a manager. The staffing diagram shows positions and headcount, but it does not explain how work gets done. An agent needs more specific information: which tools to use, how to plan and which conditions should change its behavior. A job description rarely supplies that level of detail.

A network perspective adds another layer. Instead of treating each contributor only as a job title, consider the kind of contribution they make. Misztela introduces five conceptual personas:

PersonaContribution to the network
Silent achieverStrong performance with relatively little communication.
Individual contributorCombines performance with communication.
ConnectorLinks teams that are not otherwise adjacent.
MultiplierImproves other team members’ work.
Knowledge hubSupplies domain expertise and necessary information.

These are ways of participating in a workflow, not replacements for formal job classifications.

Two people with the same title may contribute very differently. One may deliver independently while another makes everyone around them more effective. That difference affects which assistance would help, and what an intervention might disrupt. A team may already understand these distinctions informally; an executive needs enough visibility to compare them across the organization. If a contribution pattern repeats in several workflows, the same kind of agent may improve all of them.

7:177:32
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:17 · section reference included

How assistance changes the workflow itself

In Misztela’s conceptual example, a silent achiever gains a coding assistant and can take on work that previously occupied several people, allowing two workflow steps to merge into one. The same assistant helps other contributors, so some tasks previously assigned to two people become tasks for one. These are hypothetical changes in work allocation, not measured staffing or productivity results. Their significance is structural: assistance can remove a handoff, rather than merely speed up the work on either side of it.

Next, an agent expands the multiplier’s reach. Previously, that person mainly influenced adjacent colleagues. With assistance, their guidance and communication could reach the whole team, improving the work of people beyond their immediate connections. This intervention acts on relationships between contributors, so its effects can combine with the gains from individual coding assistance.

The knowledge hub changes differently. Misztela considers this persona especially susceptible to substitution by an agent, explicitly distinguishing the persona from the person’s position. A knowledge agent could dynamically support the entire team instead of routing expertise through one person serving a few colleagues. In the illustration, the former knowledge-hub contributor can then take on the whole workflow shown below. The proposal is a redistribution of work and expertise, not simply the removal of a named role.

Once these interventions begin, workflow structures may change quickly. Understanding the contributors remains useful even as the original diagram becomes obsolete. Across the company, recurring problems—such as the same bottleneck at the first step of many workflows—can identify which agents deserve investment first. The network analysis becomes a way to discover reusable improvements.

11:0511:16
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:05 · section reference included

Decide where human judgment belongs

The next decision is which parts of the workflow belong with humans and which should be delegated. Misztela starts with ambiguity: where the right action depends heavily on a situation that is difficult to define, humans remain the better fit in his account. Subjectivity can also be something the organization deliberately wants to preserve. A person’s judgment is sometimes part of the desired result, not noise to eliminate.

Conversely, reducing subjective variation or handling repetitive, tedious work can motivate agent adoption. Consistent execution is the objective here; perfect repetition is not an established property of autonomous agents. The useful allocation question is therefore more precise than whether a task can be automated: does the organization want situational judgment, personal interpretation or a repeatable application of an explicit process?

14:0714:20
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

14:07 · section reference included

Context becomes a source of advantage

Misztela expects broadly available intelligence and domain knowledge to become cheaper, weakening advantages that depend mainly on possessing them. Employees and companies would then need either multidisciplinary know-how or much deeper specialization in a particular area. The organizational consequence is that access to expertise matters less if everyone can obtain it; applying it in a distinctive setting matters more.

He points to reports online of senior engineers moving closer to product work and domain experts. Coding assistants help them develop faster, while domain assistants help them communicate across expertise boundaries. Agents can also act as connectors across the organization. In this setting, situated context gains value relative to general domain knowledge: knowing a field is different from understanding the particular work, constraints and relationships in front of you.

The expected benefits are uneven. People who already execute effectively may gain more capacity to act, widening differences in output. Misztela invokes 10× and potentially 100× engineers as speculation, without a defined task, metric or benchmark. He anticipates fewer middle-ground roles and more specialization, including kinds of work that are not yet familiar.

Two proposed roles follow directly from the earlier problems. A workflow miner uncovers the processes an organization actually performs. A human–AI orchestrator manages how people and agents interact. The accompanying slide extends this proposed role landscape, while the spoken examples concentrate on discovering work and coordinating its execution.

Colored icons accompany Workflow Miner, Human-AI Orchestrator, Trust Engineer, Sensemaker, Prompt Architect / Behavior Programmer, Change Amplifier and Embodied Proxy.Open full source frame
Proposed roles in AI-mature organizations.
15:2115:43
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

15:21 · section reference included

Let employees turn their own work into agents

Making these changes possible across an organization requires more than a central team building a few agents. Misztela proposes a progression in access and capability:

  1. Provide general assistants. Establish broad access to assistance.
  2. Provide individual assistants. Add memory, contextual adjustments and personal knowledge stores so employees do not have to explain the same background repeatedly.
  3. Enable employees to build agents. Use coding assistants, no-code tools or low-code tools to create agents when a workflow needs them.

The third step supports rapid changes to workflows because the people closest to the work can help implement the assistance they need.

The prerequisite is cognitive self-awareness. A finished deliverable hides the mental operations, decisions and actions across systems that produced it. An employee must be able to notice those operations, describe them as explicit steps and then translate them into an agent. Access to a builder does not by itself expose this knowledge. The employee’s ability to explain their own work connects the hidden-context problem to practical agent creation.

The next stage preserves knowledge beyond the individual’s tenure. Misztela imagines a digital twin representing an experienced employee’s knowledge and prior work. When that employee leaves, the organization would retain more than scattered documents and presentations: colleagues could ask questions about earlier decisions and reconstruct useful context. His time-travel metaphor describes queryable organizational memory, not a specified implementation.

From there, assistance expands to agents serving several employees, whole teams or complete workflows, and eventually to swarms of agents. Misztela attributes the feasibility of this progression to the maturation of the software around language models and agents. The organizational direction is clear even without a named deployment stack: begin with individual access, then support shared memory and coordinated work.

18:0818:27
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

18:08 · section reference included

When the agent delegates back to you

Groundedness, hallucinations, guardrails and security remain technical challenges, but the next barrier is cultural. Misztela turns to a DeepMind framework first published in 2023, Levels of AGI, to discuss the transition from human-led interaction to AI-led execution. The framework distinguishes capability from autonomy: a more capable system does not have to receive more authority. Its pre-talk autonomy table includes arrangements in which AI directs interaction and people contribute guidance, feedback or subtasks.

That arrangement reverses a familiar relationship. Instead of asking an assistant to perform a bounded subtask, a person gives the agent responsibility for the whole task—and may then receive subtasks from it. Are employees prepared to accept that direction? Are organizations ready to treat it as a legitimate way of working? Starting with simple agents does not remove the need to answer those questions before granting broader autonomy.

21:3521:50
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

21:35 · section reference included

Define values before granting autonomy

Autonomy also introduces governance questions that go beyond choosing the right action in an abstract ethical dilemma. If agents are to act as keepers of organizational values, those values must be explicit. People disagree about what matters, but an agent still needs a defined basis for action. The difficulty is deciding what it should preserve when those disagreements occur.

A second problem follows: how should an organization resolve conflict between a human and an agent? Misztela floats another agent—a justice agent—as one possibility, leaving adjudication unresolved. Delegating execution therefore creates a need to decide not only who can act, but who can challenge an action and how a dispute is settled.

He closes this discussion with a line attributed in the talk to John C. Lennox: “The greater the machine’s freedom, the more it will need moral standards.” In the published excerpt from Lennox’s 2084, Lennox credits Rosalind Picard for the quotation. The practical demand is to bring ethics and philosophy into the operation of autonomous technology, rather than treating them as separate from its design.

22:5723:11
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

22:57 · section reference included

Explore new work, then test the strategy

The opportunity extends beyond improving existing workflows. Misztela points to synthetic-reality simulations that let an organization explore an activity before trying it in marketing, a laboratory or another real setting. He also mentions proactive regulatory opportunities, labeled regulatory loophole lighthouses on the slide, and agents that scout for mergers and acquisitions. These are proposed directions rather than deployment walkthroughs; the broader opportunity list remains visible in the slide.

Opportunities slide lists talent-as-liquidity, synthetic reality wind tunnels, the causality engineer, always-on M&A scouts, edge-native supply chains, regulatory loophole lighthouses, mental model engineering, multi-agent strategy simulations and unbundling the CEO.Open full source frame
Executive opportunities including simulations, M&A scouting and an internal skill stock exchange.

Pursuing those possibilities also creates more work in philosophy, ethics, psychology and the social relationship between humans and AI. Misztela keeps the detailed technical implementation outside this talk’s scope. The closing question is whether the organization’s strategy actually accounts for the capabilities and relationships introduced by agents.

His diagnostic is deliberately sharp: replace AI agent in the strategy with machine learning or data. If the strategy still means the same thing, he considers it already outdated. A strategy specific to agents must account for connected execution, not just the addition of another technical capability.

That means understanding workflows deeply enough to delegate them, enabling employees to build and use agents, and examining the network of contributions behind the organization chart. The chart itself may change rapidly. The harder adjustment is accepting that technology may sometimes lead the work and ask humans to participate on its terms. Preparing for that possibility makes ethics part of organizational readiness, alongside access, context and technical capability.

24:3624:53
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

24:36 · section reference included

Resources

From the talk

  • Google DeepMind’s framework distinguishes model capabilities from deployment autonomy. Table 2 describes interaction ranging from human control to autonomous execution. This is the June 2024 revision of work first released in 2023.

  • AI applied to moralityArticle

    An excerpt from John C. Lennox’s 2084 discussing moral standards for AI, including a quotation credited to Rosalind Picard.

  • A controlled study of GitHub Copilot assistance on implementing a JavaScript HTTP server, reporting 55.8% faster task completion for the assisted group.

  • Building effective agentsArticle

    Practical guidance on predefined workflows, dynamic agents, tool interfaces, feedback and evaluation. Originally published in December 2024; the live page contains later tooling updates.

Transcript download
Full transcript

Loading transcript…