By Dr Kevin Shepherdson, CEO, Straits Interactive
Workflow Execution Is Not Workflow Ownership
In Part 1, we examined why the idea of the “AI employee” is so appealing—and where the metaphor begins to break down.
Agentic AI is clearly moving beyond simple assistance. Modern agents can pursue objectives, use tools, follow organisational procedures, connect to enterprise systems and execute increasingly complex sequences of work. Concepts such as Anthropic’s Agent Skills strengthen this further by giving agents reusable procedural knowledge, while technologies such as MCP improve their ability to reach the data and systems needed to act.
That combination can make an agent look remarkably employee-like.
But Part 1 drew an important distinction. An agent may increasingly possess capability, procedural knowledge, connectivity and execution ability without possessing the full set of characteristics associated with employment. Real employees operate within organisational context. They exercise judgement, handle ambiguity, receive delegated authority, build relationships, carry professional or ethical obligations and remain accountable for the consequences of their actions.
This led to the central argument:
An AI agent can increasingly be given the capabilities of a worker without acquiring the responsibilities of an employee.
The comparison also exposed another important dividing line. The “AI employee” analogy becomes strongest when we look at what the system can do. It becomes much weaker when we ask what the system is authorised to decide, how reliably it performs outside the normal path, and who remains accountable when something goes wrong.
Part 2 takes that argument one step further.
The issue is no longer simply whether an agent can execute individual tasks. Agentic systems are increasingly being presented as capable of running complete workflows—and potentially entire functions. That raises a more difficult question:
When does executing a workflow become equivalent to owning it?
The answer matters because workflow automation is where technical capability begins to collide with organisational reality. A system may perform most of the operational steps and still depend on humans for judgement, approval, exceptions, relationships and accountability.
That is where we now continue.
Workflow Execution Is Not Workflow Ownership
Imagine an AI procurement agent that can receive a request, retrieve previous purchases, draft specifications, identify approved suppliers, issue an RFQ, extract incoming quotations, compare suppliers and prepare a recommendation.
That is substantial automation.
It may even mean that the AI performs most of the information-processing work within the process.
But a procurement manager may still negotiate important terms, resolve supplier problems, approve expenditure and take responsibility for the final decision.
The agent has participated across the workflow. It may even have orchestrated most of the routine pathway.
It has not therefore become Procurement.
This helps clarify the AI Change Impact Ladder:
Task → Workflow → Function → Product/Service → Business Model → Organisation
A highly autonomous agent completing a bounded activity may still represent only Task-level change.
When AI begins coordinating multiple connected activities and hand-offs, we move into Workflow-level change.
Function-level transformation goes further. Multiple workflows, roles, staffing assumptions, control points and performance measures begin changing together.
The ladder therefore measures the scope of organisational change, not the percentage of work performed autonomously.
An agent may execute a workflow without owning the workflow.
The organisation still owns the objective, authority, controls, exceptions and consequences.
Why Coding Can Give Us the Wrong Impression
Software development has produced some of the most impressive Agentic AI demonstrations, and that is not accidental.
Code exists inside a highly digital environment. An agent can inspect files, modify software, execute code, observe an error, run tests and try again. Outcomes can often be evaluated automatically.
Anthropic's research on Claude Code has shown considerably more automation-style interaction than ordinary chatbot use. That is meaningful evidence that people are delegating more execution to agents.
It is not evidence that the corresponding proportion of a software engineer's job has disappeared.
Software engineering also involves architecture, ambiguous requirements, security, maintenance, business trade-offs and production accountability.
And most organisational work is less machine-friendly still.
A customer becomes angry. Two databases disagree. A supplier stops responding. A manager changes the objective midway through the process. A policy is ambiguous. The correct action depends on something everyone in the department knows but nobody has written down.
These are not peripheral interruptions.
They are part of real work.
The Real-World Workflow Gap
This is why research from organisations such as METR requires careful interpretation.
METR's task-completion time-horizon research shows that AI agents are becoming capable of completing increasingly substantial multi-step tasks. But METR itself cautions against translating those results directly into claims about whole-job automation.
Its benchmark tasks are heavily concentrated in areas such as software engineering, machine learning and cybersecurity and tend to be self-contained, clearly specified and objectively assessable.
An agent completing a task equivalent to several hours of skilled human work therefore does not mean it can autonomously perform several hours of a manager's, lawyer's, engineer's or consultant's working day.
The difference is what I would call the Real-World Workflow Gap:
An agent may be capable of executing a technically coherent sequence of activities while the organisation remains unable—or unwilling—to delegate the complete real-world workflow.
Inside that gap sit ambiguity, authority, incomplete information, relationships, exceptions, physical action, risk and accountability.
That is why task capability should not be automatically extrapolated into job replacement.
Human-in-the-Loop Is Too Simple
The answer is not to put a human behind every agent checking every action.
If someone has to verify every retrieval, approve every tool call and reread every output, much of the economic advantage disappears.
A more useful concept is human re-entry.
The human establishes the objective and the boundaries of delegated authority.
The agent then executes the normal pathway.
Where uncertainty rises, a threshold is exceeded, an exception appears or a consequential decision is reached, the human re-enters.
After that intervention, the agent may continue executing the workflow.
The future operating model therefore looks less like:
Human → AI → Human → AI → Human
and more like:
Human sets objective → AI executes within boundaries → Exception or threshold → Human re-entry → AI continues
The design challenge is deciding where those re-entry points belong.
Too many, and automation becomes expensive theatre.
Too few, and the organisation may surrender control faster than the system earns trust.
A Better Term: The Delegated AI Executor
For these reasons, I prefer Delegated AI Executor to “AI employee”.
The progression is better understood as:
AI Assistant → AI Executor → AI Workflow Agent → Agentic System → Delegated AI Executor
What increases across this progression is not employment status.
It is execution capability and delegated authority.
That distinction matters because authority can be constrained.
A procurement agent might autonomously gather quotes but require approval before awarding a contract.
A customer-service agent might provide refunds below a specified threshold but escalate anything larger.
A recruitment agent might screen and organise applications but be prohibited from making a hiring decision.
This gives organisations a much more practical design question:
How much execution authority should we delegate?
rather than:
Which employee can we replace?
The Six Agentic AI Tests
Before an agent is allowed to perform substantial organisational work, I would suggest that the CAIO apply six tests.
There is a useful pattern here.
Better models strengthen Capability.
Agent Skills strengthen Procedure.
MCP and integrations strengthen Connectivity.
Those developments explain why AI increasingly looks like an employee.
But they do not automatically solve:
Authority → Reliability → Accountability
Those remain management and governance problems.
And that is where the CAIO earns his or her keep.
Ask a Better Question
The “AI employee” metaphor will probably remain popular because it captures something genuinely new. AI is moving from answering questions towards executing real work.
That change should not be underestimated.
Agents may eventually perform a remarkable proportion of the routine activities currently carried out by employees. They may orchestrate substantial workflows and materially reshape entire functions.
But performing work is not the same as holding a job.
The better management question is therefore not:
Can this AI replace an employee?
It is:
What work, execution authority and decision rights can we safely and economically delegate to AI, what must remain human-led, and under what circumstances must humans re-enter the workflow?
That question accommodates rapid improvements in AI capability without pretending that capability automatically brings judgement, authority or accountability with it.
An AI agent may one day do most of the routine work inside a particular role.
The organisation will still need to decide what it is allowed to do, where it must stop, what happens when reality departs from the procedure—and who ultimately answers for the result.
That is why the distinction matters.
An AI agent can increasingly be given the capabilities of a worker without acquiring the responsibilities of an employee.
