AI changes different organizational boundaries in different ways.

AI-native is often treated as shorthand for flatter teams, fewer people, and fewer boundaries. That confuses what AI can do with what an organization should become.

I do not think companies are moving toward one final “AI-native” organizational model.

Some boundaries exist because people cannot carry enough knowledge or context. Others separate competing priorities, limit authority, assign accountability, preserve independent checks, or reflect regulatory and physical constraints.

Some of those boundaries will disappear. Others are still there for a reason, even if AI makes them easier to cross.

Before flattening teams, removing handoffs, or giving agents more authority, ask what problem the boundary was solving—and whether AI has actually removed that problem.

Boundaries Exist for Different Reasons

People cannot know everything

No one can become a great salesperson, accountant, lawyer, security engineer, and supply-chain planner at the same time. Attention, memory, training time, and available context are limited, so companies divide knowledge across roles. People usually deal with familiar problems themselves and call in specialists when they need deeper expertise.

AI changes this constraint. An agent can look across far more information than any one person could keep track of, connect context from different parts of the business, and work with tools that used to belong to separate teams.

Some teams are meant to disagree

Take a strategic customer asking for a rush order.

• Sales wants to say yes because the customer matters
• Finance worries about margin
• Operations sees the disruption to the schedule
• Risk focuses on what could go wrong

All four can see the same facts and still disagree, because they are responsible for different things.

The company still has to choose what it is willing to sacrifice: margin for growth, stability for a key customer, or risk for speed. AI can make the trade-off clearer. It cannot make the trade-off go away.

Some boundaries exist to limit power

Most companies would hesitate to let one person propose a transaction, approve it, execute it, and then review their own work. Separating those steps helps prevent fraud and hidden mistakes. This is the basic logic behind segregation of duties.

With AI agents, this separation becomes easier to blur. One agent may be technically able to propose, approve, execute, and review the same action. Capability is not authority.

That raises three questions:

• Who grants that authority?
• Who owns the outcome?
• Who can change the authority?

Every boundary has a cost

Handoffs to specialists add waiting and force people to pass context back and forth.

AI can remove a lot of that friction. But that does not mean every boundary should disappear.

• Some boundaries lose their purpose and can go away.
• Others still serve a purpose, but become much easier to work across.

One agent can cover more of the work

A customer-service employee used to ask the Order team, wait for a response, contact Finance, explain the context again, and wait again.

An agent can now check the customer, order, payment, and policy, then trigger a refund, update the record, and create a ticket in one flow. What used to require several people and systems can now happen in one flow.

But doing more does not give the agent the right to decide more. It can act across systems while still being constrained by policies, limits, and approval rules.

That is where the organizational problem starts. An agent may be able to trigger a refund even though the team running it does not own the refund policy or the payment process.

The real test is what happens when the agent cannot decide. If it can work across five teams but no one owns the exception, the process simply stops.

Cheaper coordination isn't agreement

AI is strongest when the company already knows what it wants to do: establish the facts, apply the rules, execute the policy.

If the order system says an item shipped while the billing system says it was never paid for, the job is mostly to establish what happened. If customers order 100 units but only 80 are available, software can handle much of the allocation once the company has already decided who gets priority and how low margins can go.

The hard part begins when the business has not agreed on the rule.

AI can compare the options, but it cannot decide what the business is willing to sacrifice.

Some of those choices can be made in advance and built into the workflow.

But some decisions do not fit a simple rule. A long-term customer asking for an exception may force a real judgment between margin and the relationship.

AI also forces companies to make explicit choices that used to stay buried inside meetings, manager judgment, and exception handling. It can make coordination cheaper. It does not make disagreement disappear.

Someone still has to define the goal

Agents can plan and adapt, but only inside a goal, limits, and authority someone else has set.

“Make the company better” is not enough. Better could mean revenue, margin, market share, customer satisfaction, or lower risk. AI can optimize against a set of priorities. Choosing those priorities is not an optimization problem. It is a business judgment.

Process, Decision Rights, and Organization Can Move Separately

One of the biggest mistakes in AI-native design is treating process, decision rights, and organizational structure as if they move together. They do not. In practice, at least three different things can change:

• Process: how the work flows. AI can remove handoffs, automate checks, and redesign a workflow end to end.
• Decision rights: who gets to decide. An agent may handle standard refunds while humans handle unclear cases and high-value exceptions.
• Organization: how roles, teams, ownership, KPIs, and reporting lines are structured.

When these layers move at different speeds, predictable problems appear:

• Process changes, decision rights do not. The agent finishes the analysis in 30 seconds. Approval takes two days. The bottleneck moved.
• Decision rights change, the organization does not. The agent makes the call, but when something goes wrong, nobody clearly owns the result. Decision authority moved. Accountability did not.
• The organization changes, the process does not. The company creates a new AI team and changes reporting lines, while the work still runs through email, spreadsheets, and manual approvals. The org chart changed. The work did not.

Change one layer without the others, and you often move the bottleneck instead of removing it.

Redesign the Work Before the Org Chart

Do not start from the org chart. Start from the capability the business needs and the outcome it must deliver.

Redesign the work first. Change the organization only where the work actually demands it.

If a handoff exists mainly because people cannot access the same information, carry enough context, or cover a wide enough scope, AI may remove much of the reason for it. That can justify broader roles, less coordination, more end-to-end ownership, and in some cases flatter teams.

But not every boundary should disappear. AI can make Sales and Finance easier to coordinate. It does not give them the same objective. Sales still cares about growth. Finance still cares about margin.

Sometimes redesigning the workflow is not enough. Even if people and agents can handle new-customer setup almost end to end, the work may still stall because ownership is fragmented across teams. Then the organization itself has to change.

This is the logic behind the Reverse Conway Maneuver: sometimes the team structure has to change because the work has changed—not because “AI-native” demands a flatter org chart.

AI does not change every industry in the same way

• Digital marketing: the work is digital, feedback is fast, and mistakes are often reversible. Processes and roles can change much more aggressively.
• Banking: analysis and execution can become highly automated while independent checks and legal accountability stay important.
• Manufacturing: planning and scheduling can change dramatically, while physical capacity, materials, and supplier lead times remain.
• Healthcare: patient safety, professional responsibility, and regulation keep stronger constraints in place.

Agents Can Execute, but Humans Still Set the Rules

An agent can become very good at operating inside a system. The harder question is who gets to rewrite the system.

Processing a refund under the current policy is execution. Changing the policy, expanding the agent's authority, or redefining how success is measured changes the system the agent operates inside.

Execution should inform those changes. It should not control them.

The same agent should not be able to do the work, judge its own performance, redefine success, and expand its own authority.

As agents take over more execution, people shift toward setting goals, changing rules, handling exceptions, and owning the consequences.

How to Decide What Should Change

Once the outcome is clear, ask what each boundary is still doing.

1. Why is this boundary here? Is it mainly about human limits, specialization, competing objectives, control, authority, regulation, or physical constraints?
2. What has AI actually changed? Has it removed the reason for the boundary, mainly made the boundary cheaper to cross, or barely changed the reason for it at all?
3. What might move as a result? The process, the decision rights, the organization itself, or some combination?
4. Would changing or removing it create enough value? Would the gains in speed, cost, adaptability, or customer value justify the added risk, complexity, and governance cost?

AI is not magic, and “AI-native” is not a mandate to redesign everything.

Some organizations should change dramatically. Others may adopt AI deeply while keeping much of the same structure.

A boundary can be inefficient and still be doing important work. Before removing it, understand what you might lose with it.