Source: MIT Technology Review — https://www.technologyreview.com/2026/05/26/1137584/rethinking-organizational-design-in-the-age-of-agentic-ai/
Agentic AI will disappoint companies that treat it as another software layer. That is the sharpest point in MIT Technology Review’s piece on enterprise AI agents: the technology may be advancing faster than the organizations trying to absorb it.
The article cites a large ambition gap. According to the source, 85% of organizations say they want to become “agentic” within the next three years, but their current operations and infrastructure are not ready for that shift. The weak points are not only technical. Companies cite readiness gaps across people, processes, and workflows.
That matters because the promise being sold is not a better chatbot. Enterprise agents are being pitched as systems that can execute full workflows with limited human input: coordinating tasks, making decisions, adjusting to changing conditions, and improving performance over time. If that is the target, then plugging agents into a human-first operating model is not enough.
The mistake: adding agents to a model built for humans#
Prasun Shah, global CTO for workforce consulting and chief AI officer at PwC UK Consulting, frames the current problem bluntly. Many organizations are “embedding AI employees into what is a human operating model,” he told MIT Technology Review. He compared it to adding “sticky tapes” to parts of an operating model that is already breaking.
That is a useful warning. A conventional enterprise workflow assumes human actors, human pacing, human approvals, and human boundaries between departments and applications. AI agents challenge each of those assumptions. They can operate across systems, act at machine speed, and coordinate steps that used to sit in separate roles or teams.
The result is not automatically efficiency. It can also be confusion: unclear ownership, brittle integrations, poor escalation paths, and managers who remain accountable for decisions they cannot fully observe.
The source cites early proving grounds in customer service, HR, and sales, where AI agents could accelerate business processes by 30% to 50% and reduce low-value work time by 25% to 40% when deployed at scale. Those figures are meaningful, but they should be read as conditional. The gains depend on deployment quality, workflow choice, data access, and governance. They are not a guarantee that any enterprise agent project will produce that range of improvement.
“Agentic business transformation” is a vocabulary play, but the gap is real#
The article introduces “agentic business transformation,” or ABT, a term coined by enterprise agentic AI platform Ema with HFS Research. The phrase is vendor-led, so it deserves some skepticism. New technology waves often arrive with new vocabulary designed to make older categories feel obsolete.
Still, the distinction Ema’s CEO Surojit Chatterjee draws is useful. In his framing, digital transformation moved companies from paper to software. AI transformation added AI to existing processes. Copilots assist humans with tasks. ABT, by contrast, means integrating AI agents into the organization itself.
That last claim is the real issue. If agents are expected to become active participants in work, the organization has to define what they are allowed to do, how their work is measured, who supervises them, when they escalate, and how errors are handled. Calling that ABT may or may not stick. The underlying design problem will not go away.
Shah argues that the term helps force a broader redesign: operating model, workflows, decision rights, and performance management systems. That is the right level of analysis. A company cannot safely automate judgment-heavy work while leaving decision authority, risk ownership, and success metrics untouched.
The tech stack has to become connective tissue#
The first pillar in the ABT framing is the technology stack. The source’s core point is that most enterprise stacks were built for human-operated, application-centric workflows. Agents change the actor.
A human employee can open one system, interpret context, message another team, check a spreadsheet, wait for approval, and update a record. It is slow, but the social and procedural context is often implicit. An agent operating across multiple systems needs that context exposed in a form it can use. That means access, permissions, data quality, integration design, auditability, and guardrails become central architecture questions.
Chatterjee argues that the value of agents is not as another layer on top of the stack, but as connective tissue across it. Shah makes a similar point: agents can retrieve and interpret data from multiple applications and coordinate higher-level tasks. If they can contextualize well enough, they may create competitive differentiation.
That is plausible, but it also defines the risk. The more systems an agent can touch, the larger its blast radius. An agent that can read across tools, write into workflows, and trigger actions can be powerful. It can also propagate bad assumptions faster than a human team would.
The architectural question is therefore not just “Can the agent connect?” It is “What should the agent be able to infer, decide, and change?” Enterprises that skip that distinction will mistake integration for readiness.
The workforce redesign is not just headcount math#
The second pillar is the workforce. The source argues that current workforce structures still reflect industrial-era hierarchy: standardized processes, clear task divisions, strategic business units, and managers who advance by optimizing output from teams below them.
Agentic AI blurs that model. If agents can execute, coordinate, and optimize tasks without constant managerial intervention, managers lose some execution work and gain a different kind of responsibility. They must manage hybrid teams where human employees and AI agents both affect outcomes.
Shah points to issues around trust, explainability, psychological safety, and status dynamics. Those are not soft side-notes. They shape whether workers accept agent output, challenge it appropriately, or defer to it in situations where human judgment should still dominate.
There is also a practical management problem: if a manager supervises a process partly executed by agents, what does performance management measure? Human productivity? Agent throughput? Quality of escalation? Error reduction? Customer outcomes? The old dashboard may reward the wrong behavior.
The article also cites McKinsey’s prediction that by 2030, three-quarters of current jobs will require redesign, upskilling, or redeployment. Predictions at that scale should not be treated as precision forecasts. But the direction is credible: if agents become embedded in core workflows, job design will change faster than job titles.
Metrics decide whether agentic AI becomes useful or theatrical#
The third pillar named in the source is success metrics, though the provided material cuts off before developing it in detail. Even from the earlier sections, the metric problem is clear.
If companies measure agentic AI only by time saved, they may automate shallow work and miss deeper process redesign. If they measure only cost reduction, they may create brittle systems that look efficient until an exception hits. If they measure only adoption, they may reward tool usage instead of business value.
Better metrics should separate several layers:
- workflow cycle time, not just individual task speed
- quality and error rates after agent intervention
- human escalation load and resolution quality
- decision traceability and audit completeness
- customer or employee outcomes tied to the workflow
- operational resilience when systems or assumptions change
This is where the enterprise AI conversation often becomes too vague. “Productivity” is not a metric by itself. A useful agent should improve a defined workflow under defined constraints, with observable effects and known failure modes.
What enterprises should check before scaling agents#
The practical lesson is not to avoid agentic AI. It is to stop treating it like a feature deployment.
Before scaling agents across sensitive workflows, organizations should answer a few hard questions:
- Which workflows are being redesigned, not merely accelerated?
- What systems and datasets can the agent access, and why?
- What decisions can it make without human approval?
- Who owns failures caused by agent action or recommendation?
- How are agent decisions logged, reviewed, and challenged?
- What changes for managers supervising hybrid human-agent teams?
- Which metrics prove value beyond demo-stage speed?
The MIT Technology Review piece is partly built around a vendor term, and that should be visible to readers. But the stronger point does not depend on the term. Agentic AI forces companies to confront a design problem they could ignore with copilots: once software starts acting inside workflows, the organization itself becomes part of the system architecture.
That is where most deployments will either compound value or expose the cracks.