Teams built around AI already work in ways that can feel like a preview of the future. Agents help write code, prepare releases, conduct research and coordinate tasks. From that perspective, it is easy to imagine companies becoming networks of people and software agents, with every employee directing a much larger volume of work.

I believe this direction is plausible. But moving from a small AI-first team to an established enterprise changes the problem. The challenge is no longer just what the technology can do. It is how to introduce it into an organisation that must keep operating, honour its commitments and account for its decisions.

That is why I expect company-wide redesign to take longer than the most optimistic demonstrations suggest. Here are the constraints that matter.

1. Integration begins where the demonstration ends

An agent creates lasting value when it can work with the company's actual systems: product data, customer records, finance, procurement, logistics and approval processes. Connecting those systems also means dealing with access rights and dependencies.

A seemingly small action can affect an invoice, a shipment or a customer commitment. A working demonstration proves that a task is possible. It does not yet prove that the task can be performed reliably inside the business.

2. Data carries the history of the company

Conflicting records, manual corrections and spreadsheets often exist for a reason. They preserve exceptions and workarounds that have kept the business moving for years.

AI does not automatically reconcile those differences. Without clear ownership, agreed definitions and quality checks, it can turn inconsistent data into convincing but unreliable answers. Preparing data is therefore part of the transformation, not a preliminary inconvenience.

3. Responsibility does not disappear when execution is automated

The moment an agent moves from recommending an action to taking it, several questions become unavoidable. Who approves a price change? Who is responsible for a contract? Where is the record of the decision? How can it be explained and, if necessary, reversed?

These controls take time to design. Their purpose is to limit the consequences of error. Faster execution is useful only when responsibility remains clear.

4. Access is an organisational question too

Information security is not simply a technical switch. Access reflects confidentiality, intellectual property, personal data obligations and the responsibilities of different teams.

A capable agent may need information held across several departments. Giving it that access requires a decision about acceptable risk, not just a better prompt. Security review can therefore become a practical limit on deployment even when the model performs well.

5. Processes are agreements between departments

A process diagram rarely shows the whole arrangement behind it: budgets, performance measures, ownership, escalation routes and reporting duties.

Changing the process means renegotiating those arrangements. AI may reveal unnecessary steps, but it cannot by itself settle who will carry the cost or take responsibility for the result.

6. Resistance can be rational

Automation changes more than the amount of work. It can change who makes decisions, whose judgement is visible and who bears the risk when something goes wrong.

Some resistance reflects those concerns rather than a general fear of technology. A credible change programme must address roles, incentives and accountability alongside the tools.

7. A successful pilot is only the beginning

Pilots usually operate within a narrow scope, with close attention from a small team. Wider use requires training, support, quality control, incident handling and ongoing checks on cost and model performance.

That less visible work determines whether the initial improvement can last. It also explains why scaling often takes longer than building the first useful prototype.

8. Speed needs boundaries

AI can shorten many execution cycles. It can also repeat a pricing, logistics or reporting error across many transactions before anyone notices.

The appropriate level of autonomy depends on the consequences of the action. Approval gates, monitoring and rollback procedures should reflect that risk. They are part of a workable system, not evidence that the transformation has failed.

9. New responsibilities come before a new organisation

Established roles are unlikely to disappear all at once. I expect an intermediate stage in which people take responsibility for AI operations, training, context, evaluation and oversight.

This work makes broader redesign possible. Once a company understands where agents are reliable and where human judgement is essential, it can change its structure with more confidence.

A realistic sequence of change

My expectation is a gradual sequence: experiments, shared standards, clear oversight, controlled autonomy and then wider organisational redesign. Companies will move through it at different speeds.

AI can accelerate individual tasks much faster than a business can revise all the agreements around them. That difference is not a reason to abandon the ambition. It is a reason to plan the transition honestly.

I believe in a different model of management. I do not expect it to arrive in a single quarter. The pace will depend on how quickly organisations can make new ways of working dependable—and take responsibility for them.