The Bad Review or the Standing Ovation

Opening thesis

Every company deploying agentic AI is preparing a performance, whether it knows it or not. Some will receive the bad review: failed pilots, uncontrolled agents, rising costs, unclear accountability and lost trust. Others may earn the standing ovation: clearer workflows, safer AI action, better decisions, stronger customer experience and measurable operating value.

The next AI challenge is not creating more agents, but governing the actors they become.

This is the closing question in the operating logic series. Fragmentation already shapes how companies work [1]. Operating logic already determines who owns the way work moves [2]. Vendor dependency may become agentic when companies let external platforms define how AI enters their workflows [3]. And intelligence is not only inside the model; it also lives in company context, human judgment, workflows, memory and permissions [4].

Agentic AI brings these questions together. An AI agent may classify, draft, reason, retrieve, summarize, plan or recommend. But once that capability is placed inside the operation, connected to systems, guided by a workflow and allowed to affect outcomes, it becomes something more serious. It becomes an actor. And when an actor performs, the company owns the consequence.

1. Fragmentation is already on the stage

The future operation will not start from a clean page. Most companies are not broken. Their operations are functional, commercially viable and often impressively adaptive. People solve problems locally, teams create workarounds, departments develop habits, managers know exceptions by memory, and knowledge travels through meetings, messages, spreadsheets and informal judgment.

That is not failure. It is how organizations survive commercial pressure, growth and change. But it does mean that fragmentation is already present inside the company [1]. It lives in handovers, disconnected systems, local practices, unclear ownership, duplicated data, informal decision rules, unstructured documents and knowledge that only certain people carry.

Agentic AI will inherit this reality. If the company does not make its operating logic explicit, AI will not automatically repair fragmentation. It may reproduce it. It may accelerate it. It may turn local workarounds into automated workflows and isolated pilots into a new layer of machine-speed complexity.

Key point. Fragmentation is not only a technical condition. It is an operating condition. If AI enters the company without a shared operating layer, future operations may become more fragmented, not less.

2. If the company does not define the logic, the vendor will

The temptation will be to wait. Many organizations will assume that the largest vendors will eventually package agentic AI into familiar enterprise software. That may feel safe. It may even be the right choice for specific capabilities. But it also repeats a familiar pattern: convenience in exchange for control.

The next vendor lock-in may not begin with software licenses. It may begin when vendor systems define how AI actors move through the company: how work is routed, which data becomes visible, which workflow is optimized, which approval points matter, which actions are automated, and which exceptions are ignored [3].

Vendor capability is not the same as company logic. A vendor can provide a powerful platform, but the company still needs to own the logic of its own operation: what matters, who is responsible, what may be automated, where human authority remains and which outcomes define success [2].

If the company does not define its own operating logic, the vendor increasingly defines it on the company's behalf.

3. Agents are capability. Actors are situated roles.

The market is becoming fluent in the language of agents. Sales agents, support agents, research agents, coding agents, finance agents, HR agents and service agents are appearing everywhere. But the word agent can hide an important distinction.

An agent is capability. It can perform a skill: analyze, search, draft, classify, summarize, reason or recommend. An actor is what that capability becomes when it is placed into a situated operational role. The actor is no longer floating outside the business. It enters a scene.

A scene contains other people, prior context, workflow dependencies, permissions, consequences, customers, timing, risk, commercial goals and human authority. A good actor does not merely execute a line. It understands where it is, who else is on stage, what happened before and what the next move may trigger.

The distinction. Agent = capability. Actor = capability placed into a situated operational role. The actor needs scene awareness, expert logic, permissions, human authority, auditability, cost awareness and consequence ownership.

This is why the next challenge is not simply to create more agents. It is to decide which actors deserve to exist, what performance they are responsible for and how their work can be constrained, reviewed and improved.

4. The actor does not yet see the entire scene

Many agents today can perform impressive skills, but they do not necessarily oversee the full scene. They may understand the task but not the operation. They may see a prompt but not the ownership behind it. They may retrieve a document but not understand which human decision made it authoritative. They may update a system but not understand the downstream consequence.

This is the problem that connects the entire series. Fragmented operations make the scene difficult to read [1]. Unclear ownership makes the scene difficult to direct [2]. Vendor-defined workflows may make the scene easier to use but harder for the company to own [3]. Model capability alone cannot solve this, because intelligence also lives in the company's context, rules, relationships, memory and judgment [4].

Before AI actors can perform safely, the company must make the scene legible. It must define the relevant context, ownership, dependencies, permission boundaries, human authority points and audit requirements around the action.

An actor cannot perform well if it does not understand the scene it has entered.

5. Expert logic is the missing layer

The missing layer between generic agent capability and operational performance is expert logic. Expert logic is not a prompt. It is the structured expression of what the company knows about its own work.

It includes domain rules, escalation criteria, commercial judgment, risk thresholds, exception patterns, approval requirements, workflow sequences, customer context, local knowledge and the experience that usually sits inside human teams. It tells the actor what matters, what to check, what not to decide, when to stop, when to escalate and how to make a recommendation useful.

Without expert logic, an agent may produce output. With expert logic, an actor can contribute to operating performance. This is the difference between a clever pilot and a real capability.

Expert logic defines. What matters. What context is relevant. What action is allowed. What action is prohibited. What requires escalation. What requires human approval. What outcome should be measured.

6. Rehearsal before performance

A good company should not let AI actors walk straight onto the live stage. Before AI performs inside the operation, its behavior should be rehearsed. It should be tested against scenarios, edge cases, pre-mortems, risk reviews and expert critique.

This is where the theatre metaphor becomes operational rather than decorative. The point is not to create a visual spectacle. The point is to create a way to rehearse, govern and evaluate performance before consequences reach customers, employees, partners or regulators.

Rehearsal asks: what would the actor do? Live performance asks: what is the actor doing? Review asks: what happened, who owned the consequence and what should improve?

Improvisation belongs first in rehearsal. Only what is tested, governed and approved should move to the live stage.

7. Good actors need controlled improvisation

Rigid automation is not the highest form of AI value. In many business situations, the best outcome comes from controlled improvisation: the ability to explore alternative routes, compare possible responses, identify hidden risks, test a better recommendation or recognize a pattern the fixed workflow did not anticipate.

This is familiar to research and development. Progress rarely comes from following a fixed script alone. It comes from making connections, testing assumptions and seeing relationships others do not yet see. But good R&D is not random. It is disciplined exploration inside boundaries, with feedback, evidence and review.

AI actors should be allowed to test ideas before perfect execution. They should be able to run pre-mortems, draft alternative actions, expose assumptions and show what could go wrong. But those improvisations must not overstep the boundaries that protect the company. Freedom without direction becomes risk. Direction without freedom becomes rigid automation.

Controlled improvisation. The goal is not to reduce AI actors to rigid automation. The goal is to give them enough expert logic to perform well, enough room to improve the scene, and enough boundaries to prevent improvisation from becoming operational risk.

8. The critics in the room are not blockers

Before a performance reaches the audience, critics can improve it. In companies, those critics are not only legal or compliance teams. They are domain experts, process owners, commercial leaders, operational managers, risk owners and people who understand the consequences of a decision.

Human-in-the-loop is too often reduced to a final approval button. That is not enough. Human authority must be designed into the system as direction, expertise, exception handling, review, teaching and accountability.

A human may define the role. Another may provide the expert logic. Another may approve a sensitive action. Another may handle exceptions. Another may review the performance afterwards. Human authority is the mechanism by which the company remains accountable for AI participation.

This also gives AI projects a better chance of success. RAND reported in 2024 that, by some estimates, more than 80 percent of AI projects fail, about twice the failure rate of non-AI IT projects. Gartner predicted in 2025 that over 40 percent of agentic AI projects would be cancelled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. These are not only technology problems. They are operating problems. [A][B]

9. Auditability becomes operating logic

When AI only suggests, accountability may remain easier to contain. When AI starts to act, accountability becomes operational. The company must be able to reconstruct what happened: what the actor saw, what it inferred, what it recommended, what it executed, which rule or expert logic applied, who approved or overrode it, what changed and what should improve.

Auditability is often treated as compliance administration. In agentic operations, it becomes a management capability. A company that cannot audit AI performance cannot safely scale it. It cannot learn which actors deserve more freedom. It cannot identify which actors should be constrained, retrained or retired. It cannot prove value, risk reduction or human control.

This is also where legal and regulatory implications become practical. The EU AI Act is being implemented in phases, with major rules applying from 2 August 2026 and additional obligations for certain high-risk systems later. Recent legal research on AI agents under EU law emphasizes the need to map external actions, data flows, connected systems and affected persons. The practical direction is clear: organizations will need better visibility into what AI systems do, which systems they touch and which consequences they create. [C][D]

The actor may perform the action, but the company owns the consequence.

10. More agents can mean worse performance

As agent creation becomes easier, companies may be tempted to multiply agents across every department, workflow and exception. But more agents do not automatically create a better company. They can create duplicated work, conflicting outputs, unclear ownership, rising compute costs, security risk, audit gaps and management confusion.

This is already being described as agent sprawl. Gartner has warned that by 2028 an average global Fortune 500 enterprise may have over 150,000 agents in use, up from fewer than 15 in 2025, creating significant complexity and management challenges. [E]

AI performance also has a cost profile. It consumes compute, energy, budget, governance capacity and trust. The International Energy Agency projects that global data-centre electricity consumption could roughly double by 2030 to around 945 TWh in its base case. Agentic AI may make this question more visible: who pays for the action, who benefits from it, and who decides whether the actor still deserves to perform? [F]

The question is not how many agents a company can create, but which actors deserve to perform.

11. Hotels show the problem clearly

Distributed hospitality makes this tangible. A hotel group with 10, 20 or 40 locations does not need 40 isolated AI experiments. It needs a shared operating logic where local signals, guest experience, maintenance, events, reviews, sales, housekeeping and management learning can improve together.

A guest complaint may begin at the front desk, relate to housekeeping, reveal a maintenance pattern, affect the review score, require management attention and create a commercial recovery opportunity. A generic support agent may see only the message. A governed actor needs to understand the scene.

The same applies to event leads, recurring maintenance issues, late room readiness, supplier delays, upsell patterns, negative review signals and operational handovers. The intelligence is not only in the model. It is in the history of the location, the group-level pattern, the team's experience, the commercial context and the human decision about what should happen next [4].

If each location builds its own AI logic, the organization may automate fragmentation. If each vendor imposes its own workflow, the organization may lose control over how work is interpreted and improved. But if the company makes its operating logic explicit, AI actors can help local knowledge travel, repeated problems become visible and performance improve across the group.

12. The company is judged

A guest will not review the model. A client will not complain to the workflow engine. A regulator will not judge the prompt in isolation. The organization will be judged by the performance it allowed into the operation.

That is why the bad review or the standing ovation is ultimately about accountability. The actors may perform, but the company owns the consequence. If AI acts without scene awareness, without expert logic, without human authority, without auditability and without cost discipline, the performance will eventually be judged. If it goes wrong, the bad review belongs to the organization that allowed the actor onto the stage before it was ready.

The standing ovation will not go to the company with the most agents. It will go to the company that knows which actors deserve to perform, how they are rehearsed, where they may improvise, when human authority must intervene, how consequences are audited and how every performance improves the next one.

The actors may perform. The company is judged.

This is the closing point of the series. Fragmentation explains why change is hard [1]. Operating logic explains where control really sits [2]. Agentic vendor lock-in explains why ownership may shift if companies wait too long [3]. Intelligence beyond the model explains why context, workflow and human judgment matter [4]. And this final article explains what happens when AI enters the stage: the company must govern the actors it allows to perform [5].

That is the difference between the bad review and the standing ovation.

Series references

• [1] Why Fragmentation Makes Change Hard
• [2] Who Owns the Operating Logic?
• [3] The Next Vendor Lock-in May Be Agentic
• [4] The Intelligence Is Not Only in the Model
• [5] The Bad Review or the Standing Ovation

External notes

• [A] RAND Corporation, The Root Causes of Failure for Artificial Intelligence Projects, 13 August 2024.
• [B] Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025.
• [C] European Commission AI Act Service Desk, Timeline for the Implementation of the EU AI Act.
• [D] Nannini et al., AI Agents Under EU Law, arXiv, April 2026.
• [E] Gartner, Gartner Identifies Six Steps to Manage AI Agent Sprawl, 28 April 2026.
• [F] International Energy Agency, Energy and AI: Energy demand from AI.

Share