AI Will Eliminate Some Work. It Will Also Create an Entire New Layer of It.
One of the most common questions about AI is how much human work it will eliminate.
A less obvious question may prove just as important:
How much new work will AI create simply because organizations need to become explicit about how work actually gets done?
Boston Consulting Group’s 2026 Applied AI Index offers an intriguing clue: BCG found that 89% of companies expect AI to generate new work, while only 11% expect it primarily to replace existing jobs. BCG: The Formula for Agentic AI Value
Some of that new work will be easy to recognize: new AI products, new services, new analytical capabilities, new customer experiences.
But another category may be emerging underneath it: call it decision infrastructure work.
Before an AI agent can safely take action inside a business, someone has to define what that action means.
Humans have been carrying more of the operating model than we realize
Consider a customer-service process. A human employee may know:
- Which systems to check
- Which customer information can be accessed
- Which exceptions deserve escalation
- When a refund is appropriate
- When a manager should approve it
- Which workaround is acceptable when a system fails
- Which policy technically exists but is rarely applied literally
- Who owns the problem when something goes wrong
Much of that knowledge may never appear in one document. It lives across policies, software permissions, training materials, manager judgment, institutional memory, conversations, habits, and experience.
Humans are remarkably good at operating inside that ambiguity. Autonomous AI systems make the ambiguity much more consequential.
BCG argues that enterprises need to define an agent’s purpose, scope, autonomy, permitted actions, required human oversight, ownership, evaluation criteria, rollback mechanisms, and runtime controls. BCG: The Formula for Agentic AI Value
That means organizations may have to formalize operating logic that employees have been carrying implicitly for years.
Before you automate the process, you may have to discover the process
That sounds like documentation. It is bigger than documentation.
Organizations first have to determine what the real workflow actually is. The process described in a policy manual may not be the process employees use:
- Ownership may be shared.
- Exceptions may be handled informally.
- Permissions may have accumulated over years.
- Two departments may believe they own the same decision.
- Nobody may formally own another one.
- And many important decisions may still depend on some version of: “Ask Sarah. She knows how we handle those.”
That works differently once an agent is expected to act at machine speed.
BCG’s research says future-built companies expect nearly 70% of their AI value to come from reshaping and inventing business processes, rather than simply deploying better AI tools. BCG: The Formula for Agentic AI Value
That distinction matters. The opportunity is not to meticulously encode every inefficiency in today’s process. It is to make the process visible enough to challenge it.
A useful sequence may look something like:
Map → Challenge → Redesign → Define Authority → Instrument → Automate
Only after the organization understands the work should it decide which parts belong to people, which belong to AI, and which should disappear altogether.
AI creates new labor around authority
The same issue appears wherever an agent moves beyond generating information and begins taking action. Someone has to decide:
- What may the agent see?
- What may it change?
- How much money may it commit?
- Which decisions are reversible?
- Which actions require approval?
- What confidence is sufficient?
- What happens when the evidence conflicts?
- Who can stop the system?
- Who is accountable afterward?
BCG’s agent-control framework includes scope and oversight, ownership, evaluation and rollback, security, auditability, memory and context, and tool integration. BCG: The Formula for Agentic AI Value
Those controls do not emerge automatically from the model. They have to be designed—and then maintained.
AI systems will create maintenance work too
Operating logic changes. Prices change. Regulation changes. Customers change. Risk tolerance changes. Products change. Employees discover exceptions. Models improve.
A decision boundary that made sense six months ago may no longer make sense today.
So this new infrastructure cannot be documented once and forgotten. It will need to be observed, tested, revised, versioned, audited, and sometimes challenged completely.
That creates another class of work: maintaining the organization’s decision logic.
In traditional software, we maintain code. In agentic organizations, we may increasingly maintain:
- Rules
- Authority boundaries
- Escalation paths
- Workflow definitions
- Business context
- Evaluation criteria
- Exception policies
- Human review points
- Evidence requirements
That is a substantial operating responsibility.
This may become one of AI’s hidden implementation costs
There is a tendency to imagine AI deployment as primarily a technology purchase:
Buy the Model → Connect the Data → Deploy the Agent → Measure Productivity
But if organizations want agents to act reliably across customer service, onboarding, finance, sales, operations, procurement, compliance, or other consequential workflows, there is a layer of organizational work between having intelligence and granting authority.
That layer may become one of the largest sources of both implementation effort and competitive advantage. The companies that do it well will not merely have better AI. They will have a clearer understanding of how their own businesses work.
The emerging management question
This leads to a different way of thinking about AI labor.
AI will certainly automate some existing tasks. But it may also force companies to create a new discipline around making work, authority, and decision logic explicit.
Before an organization can safely automate more of the business, it may first have to answer a surprisingly difficult question:
Can we actually explain how the business makes decisions today?
For many companies, discovering the answer may be the beginning of the real AI transformation.