“The price of greatness is responsibility.”
Winston Churchill
Somewhere in a corporate chat window, an AI agent is reading a shared inbox, looking for an invoice discrepancy, drafting a supplier email and preparing to update a system of record. Someone has called it a “digital colleague.” Someone else has added a cheerful avatar. Nobody has answered the less cheerful question: who owns the decision if it does the wrong thing?
The language we use is part of the problem. Calling an agent an assistant makes it sound harmless. Calling it an intern makes it sound manageable. Neither description survives first contact with scale.
An intern cannot send three thousand personalized messages in a minute. An intern cannot access twenty systems at once. An intern cannot repeat the same bad assumption perfectly, overnight, without losing concentration or asking for help. An AI agent can.
So the management question has changed. It is no longer only “what can this tool do?” It is what authority are we giving it, under which conditions, with whose name attached to the outcome?
In How to deal with AI agents: stop prompting and start directing, I argued that better instructions are only the starting point. Direction requires a mandate, a context and a person prepared to own the result.
An assistant is not an accountability model
I understand why companies enjoy the metaphor of the AI assistant. It offers a comforting picture of the future: a diligent helper doing the boring parts while the human gets on with more interesting work.
The trouble begins when the helper can act.
A summarization tool produces text. A recommendation engine proposes a next step. An agent can read a document, select information, invoke a tool, write to a database, contact a customer or trigger a process. At that point, we are not discussing a feature. We are discussing delegation.
Delegation has always required a clear mandate. What is the objective? What is out of bounds? When must the person doing the work escalate? Who checks the result? A surprising number of AI deployments have plenty of enthusiasm around the first question and very little patience for the rest.
This is why I resist the phrase “AI employee.” Employees have contracts, training, performance reviews, legal obligations and the ability to explain why they made a choice. An agent has none of these in the human sense. It has permissions, prompts, models, data and a design that may or may not produce behavior we anticipated.
The agent is not responsible. The organization that gave it access is.
The moment automation becomes delegation
Traditional automation is usually narrow. A rule fires, a workflow moves, a record changes. We can often inspect the logic and test the expected paths.
Agentic systems change the shape of the problem. They handle ambiguous language, choose among tools, decide what information seems relevant and adapt their next action to what happened before. That flexibility can be useful. It also means a process is no longer only executed, it is interpreted.
IBM’s description of small Forward Deployed Units makes the healthy version of this model clear. The agents can generate requirements, code, test cases or defect analysis. The human team still defines the problem, provides business and technical context, and decides whether the output is fit for the client.[1]
That is not a ceremonial human in the loop. It is a division of labor.
The human role has to be designed around the moments that matter:
| Agent can do | Human must own |
|---|---|
| Collect and summarize evidence | Define the decision and the evidence standard |
| Draft a recommendation | Test whether it fits the commercial, ethical and regulatory context |
| Trigger a routine action | Set the authority threshold and escalation rule |
| Detect a pattern | Decide whether the pattern deserves action |
| Execute at speed | Carry responsibility for the impact |
A human who sees the output only after it has caused harm is not in the loop. They are in the post-mortem.
The control gap is a management gap
BCG’s 2026 Applied AI Index puts a number on the discomfort. Forty-two percent of companies expect agents to act autonomously by 2030. Only 5% have the full set of relevant controls in place today. Companies with all six controls report three times as much agentic AI value as those with only one.[2]
It is tempting to classify this as an IT issue. It is not. Security, data quality and access management are necessary, but they cannot decide what a business should delegate or who should live with the consequences.
A customer-service agent that offers an unauthorized concession is a commercial problem. A procurement agent that favors a supplier because of messy data is a governance problem. A recruiting agent that filters candidates unfairly is a people problem. A finance agent that quietly classifies a transaction incorrectly is an accountability problem.
NIST’s Generative AI Profile is useful precisely because it refuses to treat governance as a one-page policy. It recommends aligning use with legal and regulatory duties, documenting data and generated-content provenance, evaluating safety measures before deployment and on an ongoing basis, and defining an incident-reporting minimum that includes the system, source, date, impact and affected stakeholders.[3]
That sounds dry until an agent does something expensive at 2 a.m. Then dry becomes underrated.
A minimum operating model for AI agents
I would not begin with an elaborate committee. Most companies already have enough of those. I would begin with a short, enforceable agent mandate for every meaningful deployment.
1. Name the decision, not only the task
“Process customer emails” is not a mandate. “Classify standard shipping complaints and draft a response within approved templates” is closer. The sharper the decision, the easier it is to spot what does not belong to the agent.
2. Set an autonomy threshold
An agent can prepare, recommend, execute within a narrow limit or execute freely. Those are different operating modes. Write down the financial, legal, reputational and human thresholds that trigger a review.
3. Give one person the right to stop it
Not a department. A named person. If nobody has the authority and incentive to pause the system, then the company has automated momentum rather than created capability.
4. Keep an audit trail that a manager can read
Logs are not enough if they are useful only to engineers. A manager should be able to answer: what did the agent see, what did it do, what rule or prompt mattered, what changed and who approved the operating boundary?
5. Treat incidents as design input
The point of a record is not to blame the colleague nearest to the keyboard. It is to learn which conditions, data or incentives made the failure possible. NIST explicitly includes incident reporting, feedback and monitoring in its governance guidance for generative AI.[3]
This is management work. It deserves the same seriousness as a pricing decision, a new market entry or a customer-data policy.
The temptation to delegate responsibility as well as work
The business case for agents is obvious. Speed is seductive. A process that took days can happen in minutes. A small team can cover far more ground. BCG reports that the companies getting the most value from AI concentrate on a few end-to-end workflows and combine the technology with changes to people and process, not just algorithms.[2]
But there is a darker temptation inside the same story. The more complex an automated process becomes, the easier it is to say that nobody could have known. The model did it. The vendor did it. The data did it. The prompt did it.
Cui bono when responsibility dissolves into a flowchart?
Good governance does not slow down useful AI. It makes the delegation legible. It tells people which decisions remain human, which actions can happen automatically and what happens when the system encounters the messy world that no demo ever includes.
That is the difference between treating AI as another tool and becoming a digital orchestrator: the latter means designing the system around purpose and accountability, not around the most impressive demo.
An agent may be brilliant at the task we gave it. It will never be accountable for why we gave it that task, what it was allowed to do or who paid when it got the context wrong.
That part is still ours.
References
[1] IBM Think, AI is tightening consulting teams and raising the bar for who’s on them, 21 September 2026.
[2] Boston Consulting Group, The Formula for Agentic AI Value: The Applied AI Index 2026, 30 September 2026.
[3] National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, 2024.