When AI agents can work across departments, does a company still need its old structure?

1. From AI assistant to a system that takes on multi-step work

The difference between an assistant and an employee is that an assistant waits to be told what to do, one thing at a time, while an employee takes a goal and sees it through. AI is moving from the first kind to the second, and that is why the way you organize has to change with it.

An AI Assistant answers questions or produces drafts when someone asks it to. An Agent takes a goal and carries out a sequence of work: reading a request, checking data, calling an API, creating a document, requesting approval and following up. This shifts the unit of automation from “one task” to “one outcome,” and outcomes often cross systems and departments.

2. Why silo structures slow a company down

A customer asking for a quotation doesn't care that the details sit with sales, the pricing with accounting and the delivery date with production. They only care when they'll get an answer. Every time work passes from one desk to another, a queue forms.

Departments were created to bring expertise together, but customer workflows don't stop at organizational lines. The customer wants a quotation and has no interest in the fact that the data sits in sales, the pricing in Finance and the delivery date in Operations. Every handoff adds a queue, room for misinterpretation and a chance for information to slip through the cracks.

Once agents can connect data, keeping handoffs on email and spreadsheets becomes the new bottleneck. AI may prepare an answer in seconds and then wait a day for an approver, or create a document that someone then has to copy into the system by hand. So the change has to reach decision rights and the workflow itself. Bolting an agent onto the old process won't do it.

That said, you shouldn't dissolve every department. Expertise, people development, professional standards and oversight are still needed. The answer is a “Functional Home + Outcome Workflow” model: people keep a home base for their expertise but work through paths with a shared owner for the outcome.

3. Organizing around workflows

4. The role of people in an Agentic Organization

Team leads will spend less time dividing up queues and chasing work, and more time analyzing the causes of exceptions, refining the knowledge base and making capacity decisions. Frontline staff take the cases with the most ambiguity and the highest impact, so they need a complete summary of the information. They shouldn't be handed scraps of work the agent couldn't finish, with no context.

Mid-sized companies shouldn't create too many new positions. One person may wear several hats, but each hat has to be clearly defined, especially who is responsible when an agent uses the wrong data or acts outside its scope.

5. Build a control plane before the number of agents grows

A control plane is a central system that knows which agents exist, who sponsors each one, which model it uses, which data and tools it can access, how much it costs and how well it performs. Every agent needs its own identity and least-privilege permissions, instead of sharing an employee's account.

Evaluation has to be ongoing, because models, data and workflows change. Build a set of important cases, test against it before each deployment and spot-check production as well. If quality drops, the system should automatically fall back to human review instead of letting errors pile up.

6. A Lead-to-Cash example in an agentic organization

People come in at the key points: sales staff make sense of vague needs, managers approve discounts and Finance checks risky customers. If an agent finds conflicting data, it must stop and summarize what it knows and what needs deciding, instead of sending a bare error message.

7. Moving away from the old departments without causing confusion

1. Pick one value stream: Map the process and appoint an Outcome Owner.
2. Build a central state: Stop passing status around in emails and in multiple versions of files.
3. Split the tasks: Separate deterministic automation, AI judgment and human decisions.
4. Trial in shadow mode: Let the agent make recommendations without taking action, and compare it with the real team.
5. Open permissions one level at a time: Start with reading, then creating drafts, and only then writing or sending in standard cases.
6. Adjust KPIs and roles: Drop metrics that let each department optimize for itself while the workflow as a whole stays slow.

During the transition, you need a manual procedure and an incident owner. Don't run the old and new systems side by side indefinitely, because that doubles the work. Set criteria for retiring the old steps once the new system is stable and has made it through a peak period.

8. A readiness checklist for executives

If several of these are missing, start with the process and the data foundation. Don't buy an agent platform and hope the technology will make up for the organization's lack of clarity, because the system will spread ambiguity and conflict faster than before.

In short: Agents will thin the walls between departments, but the need for expertise and accountability remains. The organization of the future should combine a functional home with outcome workflows, a central state and a control plane. Companies that start designing decision rights and data today will be able to use agents across departments without losing control.