BUSINESS INSIGHTS Introduction
A quiet shift is under way inside large organisations. For decades, getting anything done at scale meant moving a request through layers of people: a line manager, a department head, an approvals committee, perhaps a shared-services team. AI agents are beginning to compress that chain. Unlike earlier generations of software, these systems can plan a sequence of steps, pull data from multiple applications, take limited actions and hand off exceptions to a human only when judgement is genuinely required. That capability is prompting a serious question in boardrooms: are AI agents flattening the corporate hierarchy, or simply moving the work of coordination somewhere else? The answer matters for how CEOs, CIOs and CHROs design their organisations over the next five years, and it is more nuanced than either the optimists or the sceptics currently suggest.
Key Takeaways
- AI agents can now execute multi-step workflows across enterprise systems, reducing the need for some coordination layers built purely to move information between people.
- Middle management is not disappearing wholesale; its focus is shifting from routine supervision towards exception handling, governance and people development.
- Functions such as finance, HR, customer service, IT and procurement are already piloting agentic workflows, with measurable but uneven results.
- Genuine organisational flattening depends on trust, data quality and governance maturity, not just on the technology being available.
- Accountability, oversight and security risks remain significant, and human judgement is still essential for ambiguous or high-stakes decisions.
- The most likely outcome is not a hierarchy-free enterprise but a differently shaped one, organised around outcomes and agent-human teams rather than pure reporting lines.
What Makes AI Agents Different From Earlier Automation
Robotic process automation and rules-based scripts have existed for years, quietly handling narrow, repetitive tasks such as data entry or invoice matching. AI agents are a step further. Built on large language models and augmented with planning, memory and tool-use capabilities, they can interpret a goal expressed in natural language, break it into steps, call on multiple software systems, and adjust their approach based on what they find along the way.
The distinction matters for organisational design. A single task-automation bot removes friction from one step in a process. An agentic system, by contrast, can coordinate several steps that previously required a person to relay information between departments — checking a budget, verifying a policy exception, updating a record and notifying stakeholders, all without a human physically routing the request from desk to desk. McKinsey’s 2025 State of AI research found that a meaningful share of enterprises are already scaling agents in at least one function, with IT, knowledge management and engineering furthest ahead, and that organisations pursuing more fundamental workflow redesign are seeing considerably stronger returns than those simply bolting agents onto existing processes.<sup></sup> Read McKinsey’s 2025 State of AI report
This is the crux of the flattening debate. Historically, much of what middle managers did was not strategic judgement but coordination: relaying status updates, chasing approvals, translating instructions between systems and teams, and checking that policies were followed. If an agent can perform that coordination reliably, the argument goes, fewer human links are needed in the chain.
Where the Hierarchy Is Genuinely Compressing
Direct Access Replacing Layered Approval
In several enterprise functions, employees are beginning to interact directly with intelligent systems rather than routing requests through several human approvers. In procurement, an agent can check contract terms, compare supplier pricing and initiate a purchase order within pre-set thresholds, cutting out steps that once required a buyer, a manager and a finance reviewer to each touch the same request. In IT service management, agents can triage tickets, resolve common issues and escalate only genuine anomalies, reducing the queue that once required a tier of human dispatchers.
The World Economic Forum’s Future of Jobs Report 2025 reflects this shift at the level of the labour market itself: employers surveyed expect a substantial share of core skills to change by 2030, and clerical and administrative roles — many of which existed to coordinate information flow — are among those projected to decline fastest, even as new technical and analytical roles grow.<sup></sup> See the WEF Future of Jobs Report 2025
Faster Decision Loops
Where agents are trusted with well-defined decisions, the time between a request and an outcome shrinks. A finance team using an agent to reconcile transactions and flag anomalies does not need a multi-day cycle of human review before escalation; the agent narrows the pool of issues requiring judgement to a fraction of the original volume. That said, speed alone does not equal flattening — it equals faster throughput within whatever structure remains. Businesses exploring how automation reshapes decision cycles have covered similar ground in From Hype to Infrastructure: How AI, Crypto, and Autonomous Agents Are Becoming the Backbone of the Global Economy, which examines how autonomous systems are moving from experimental pilots into core operational infrastructure.
Where the Hierarchy Is Simply Being Redrawn
Redistribution, Not Removal
A more sober reading of the evidence suggests that responsibility is being redistributed rather than eliminated. McKinsey’s research into what it calls the “agentic organisation” argues that traditional organisation charts, built around delegation through fixed reporting lines, are giving way to more fluid work structures organised around outcomes and small, cross-functional human-and-agent teams — but this still requires people to design workflows, set guardrails and own results.<sup></sup> Explore McKinsey’s analysis of the agentic organisation New roles are emerging as a direct consequence: agent orchestrators who design and supervise multi-agent workflows, hybrid team leads who manage blended human-and-agent groups, and specialists responsible for monitoring agent performance and correcting drift. These are still management roles. They simply manage a different mix of resources.
The Governance Burden Grows, Not Shrinks
Autonomous systems that can take action inside enterprise applications introduce a governance requirement that did not exist when a human was the one clicking “approve.” Someone still has to define what an agent is permitted to do, audit its decisions, and intervene when it behaves unexpectedly. GEBM’s own coverage of When an AI Model Goes Rogue: Inside the AI Kill Switch Act illustrates how quickly governance questions move from theoretical to urgent once autonomous systems are given real authority inside a business. Far from removing a layer of oversight, agentic adoption tends to create a new, more technical one.
Functional Impact Across the Enterprise
Finance. Agents increasingly handle transaction matching, variance analysis and first-pass commentary on management accounts, freeing finance business partners for forecasting judgement rather than data assembly.
Human resources. Agentic tools can screen applications, schedule interviews and handle routine employee queries, though final hiring and disciplinary decisions remain with people, for legal and fairness reasons alike.
Customer service. Tiered support, where a query once moved from frontline agent to supervisor to specialist, is compressing as AI agents resolve more enquiries end to end, escalating only complex or sensitive cases.
IT. Ticket triage, patching and routine configuration changes are increasingly agent-led, freeing engineers for architecture, security exceptions and novel incidents.
Procurement. Contract review, supplier comparison and routine approvals within defined thresholds are strong early use cases, though strategic negotiations stay human-led.
Operations and project management. Agents can track milestones and generate status updates automatically, reducing the reporting burden that once justified additional coordination roles.
Comparison Table: Traditional Hierarchy vs Agent-Assisted Structure
| Dimension | Traditional Hierarchy | Agent-Assisted Structure |
|---|---|---|
| Information flow | Passes sequentially through several approval layers | Flows directly between employees and systems, with agents pre-processing routine steps |
| Decision speed | Constrained by meeting cycles and sign-off chains | Near real-time for well-defined, low-risk decisions |
| Manager’s core task | Supervising routine execution and relaying updates | Setting strategy, managing exceptions, coaching people |
| Coordination mechanism | Human relay between departments | Agent orchestration across systems, human oversight of outcomes |
| Accountability | Clear line-management ownership | Requires explicit governance frameworks to remain clear |
| Failure mode | Slow, bureaucratic bottlenecks | Fast-moving errors if oversight is weak |
Common Mistakes Organisations Make
Treating agent deployment as a technology project rather than an organisational redesign. Installing agentic tools without rethinking who is accountable for the outcomes they produce tends to create confusion rather than efficiency.
Removing human checkpoints too quickly. Enthusiasm for faster throughput sometimes leads leaders to strip out review steps before the underlying agent has proven reliable across edge cases, creating exposure to compounding errors.
Underinvesting in data quality. Agents are only as good as the systems they draw from; fragmented or poorly governed data undermines even well-designed agentic workflows.
Ignoring middle managers in the transition. Organisations that treat existing managers purely as a cost to be automated away, rather than as the people best placed to define exception-handling rules and coach hybrid teams, tend to lose valuable institutional knowledge.
Failing to define escalation paths. Without clear rules for when an agent must hand a decision to a person, accountability gaps appear precisely in the moments that matter most.
Assuming flattening is automatic. Simply adopting agentic software does not restructure an organisation; deliberate redesign of roles, incentives and governance is required to realise any structural benefit.
Future Trends: The Next Three to Five Years
Expect the language of “org charts” to increasingly compete with what McKinsey has termed “work charts” — structures organised around outcomes and task exchange rather than fixed reporting lines. Enterprise software vendors will continue building orchestration layers that let specialised agents coordinate with each other and with human teams, extending today’s point-solution agents into more integrated systems. Governance will professionalise: expect dedicated AI oversight roles, internal audit functions for agent decisions, and clearer regulatory expectations, particularly in financial services and healthcare. Management itself will likely bifurcate into two tracks — people leadership, focused on development and judgement calls, and system governance, focused on defining what autonomous agents are permitted to do. Organisations that invested early in enterprise AI infrastructure are likely to move fastest, since agentic systems depend heavily on existing data and compute foundations. Leadership itself is evolving in parallel, a theme explored in Global Leadership | CEOs Shaping Industry in 2025.
Practical Considerations for CEOs, CIOs and CTOs
Leaders considering agentic adoption should distinguish between task-level automation, which handles a single defined activity, and true agentic coordination, which spans multiple systems with real autonomy — the latter carries materially higher governance requirements. Before scaling, it is worth piloting in a contained function with measurable outcomes, defining explicit boundaries for agent authority, and building an audit trail for consequential decisions. Equally important is engaging existing managers early, since their process knowledge often determines whether an agentic workflow proves reliable or brittle. This tracks with McKinsey’s finding that a substantial share of firms adopting agents have already experienced at least one AI-related incident, with better-performing organisations distinguished by human-in-the-loop controls and centralised oversight rather than by more advanced models.
Frequently Asked Questions
Are AI agents actually eliminating management jobs? Not in a wholesale sense. Evidence to date shows management roles evolving rather than disappearing outright, with routine coordination tasks decreasing and responsibilities shifting towards exception handling, governance and people development. Broad, unsupported claims of mass management elimination are not currently supported by employer survey data.
What is the difference between an AI agent and a chatbot? A chatbot typically responds to individual queries within a conversation. An AI agent can plan a sequence of steps, call on multiple tools or systems, and complete a multi-step task with limited supervision, adapting its approach as new information emerges.
Which business functions are adopting AI agents fastest? IT, knowledge management, customer service and finance are among the functions with the most mature agentic deployments, largely because they involve well-defined, high-volume, rules-based tasks that are easier to automate safely than judgement-heavy work.
Do AI agents remove the need for human oversight? No. Autonomous action inside enterprise systems increases the need for defined governance, audit trails and clear escalation rules, even as it reduces the volume of routine human intervention required for everyday tasks.
Will AI agents make organisations flatter overall? Many organisations will become leaner in coordination layers built purely around information relay, but new governance and orchestration roles are likely to emerge, meaning the total structure changes shape rather than simply shrinking.
What risks should businesses watch for when deploying AI agents? Key risks include accountability gaps when escalation paths are unclear, security exposure from agents with broad system access, poor decisions from agents operating on incomplete data, and employee disengagement if the transition is handled without transparency.
How should middle managers prepare for agentic AI? Managers who focus on developing skills in exception handling, cross-functional coordination, AI governance and people leadership are best positioned, since these are the areas least likely to be automated and most valuable in an agent-assisted structure.
Final Thoughts
The honest answer to whether AI agents are flattening corporate hierarchies is that they are doing something more specific: removing layers built solely to move information and approvals between people, while adding new layers built around governing autonomous systems. For some organisations, particularly those with heavily bureaucratic approval chains, that will feel like genuine flattening. For others, it will feel like a lateral shift in where responsibility sits rather than a reduction in structure. What is consistent across both outcomes is that human judgement, accountability and leadership do not become less important — they become more concentrated, applied to fewer but higher-stakes decisions. The organisations that get this right will not be the ones that adopt the most agents, but the ones that redesign accountability and governance deliberately enough that speed does not come at the cost of control.
