Here is a management question that did not exist five years ago and is now arriving in real organisations: your team of eight includes three systems. What does a one-to-one look like, how do you assess performance, and who is accountable when the software gets something wrong?
The honest answer is that nobody has fully worked this out. But the questions are becoming practical rather than theoretical, and the organisations thinking about them early are making better decisions than those improvising.
What actually changes for a manager
Start with what does not change. Managers are still accountable for output, quality and the wellbeing of their human team. Systems do not absorb accountability; they redistribute where the work happens.
What changes is the nature of the management task itself.
From directing to specifying
Managing people involves substantial ambiguity. You describe roughly what is needed, and a capable person fills the gaps with judgement, asking when genuinely stuck.
Systems do not fill gaps well. They fill them confidently and sometimes wrongly. Managing agentic capacity means specifying with a precision that most managers have never needed: what counts as done, what the edge cases are, when to escalate rather than proceed.
Managers who find this transition easiest tend to be those who were already good at writing clear briefs. Those who managed primarily through conversation and correction find it harder.
From reviewing output to designing sampling
You cannot review everything a system produces at volume, and reviewing nothing is not oversight. The task becomes designing what to check: high-consequence cases, low-confidence cases, a random sample for quality baseline.
This is a statistical skill more than a managerial one, and it is not currently taught in any management development programme.
From managing capacity to managing failure modes
Human teams degrade gracefully. Someone overloaded works slower, makes more errors, or asks for help. There are visible signals.
Systems degrade differently. They maintain apparent quality until a boundary is crossed, then fail in ways that can be sudden and correlated. The management skill is knowing where those boundaries are and watching for approach, which requires understanding the system rather than only its output.
The most consistent early mistake: treating an agentic system as a reliable junior employee. A junior asks when unsure and improves from correction, largely unprompted. A system does neither by default. The mental model that works better is a very fast process that produces confident output and requires deliberate quality control.
What happens to the human roles
The genuine version of this is neither the utopian nor the catastrophic account.
Roles composed largely of routine coordination and information transfer are most exposed. Roles requiring judgement, relationship, physical presence or accountability are least. Most roles are a mixture, which means the common outcome is composition change rather than elimination: the same job with a different balance of activities.
That sounds benign and frequently is not, for a specific reason. When the routine portion of a role is removed, what remains is the difficult portion, continuously, without the lower-intensity work that provided natural recovery. A caseworker whose straightforward cases are handled by a system spends their entire day on complex and distressing ones. That is a genuine wellbeing risk and it is routinely overlooked in workforce planning.
Roles that are emerging
Several functions are appearing that did not exist in most organisations three years ago. People who specify and maintain the behaviour of deployed systems, sitting between operations and engineering. People who design and run quality sampling. People who own the governance question of what systems are permitted to do. These are real jobs with real skill requirements, and organisations are mostly filling them by accident rather than by design.
The accountability problem
When a system makes a costly error, someone is answerable. In practice, accountability lands on the manager of the function, which is defensible, and it requires that the manager had genuine ability to control the outcome.
That condition is frequently not met. Managers inherit systems they did not specify, cannot modify, and do not fully understand, then carry responsibility for output. This is an unstable arrangement and it produces predictable behaviour: managers either over-review, destroying the efficiency gain, or disengage from oversight entirely.
The fix is not complicated but requires deliberate decision. Either give the manager real authority over system behaviour, including the ability to change guardrails and halt operation, or locate accountability with whoever does hold that authority. Splitting responsibility from control is the failure mode.
Digital departments, honestly assessed
The phrase suggests entire functions operating autonomously. The reality in most organisations is narrower: specific processes within a function running with minimal human involvement, surrounded by processes that do not.
Where fuller automation of a function does work, it shares characteristics. High transaction volume, well-defined inputs, decisions expressible as rules, contained within systems the organisation controls, and consequences of error that are recoverable.
Functions matching that profile, portions of accounts payable, first-line support triage, standard document generation, are genuinely being run this way today. Functions that do not match it, anything involving negotiation, relationship, novel judgement or irreversible external consequence, are not, and claims otherwise generally do not survive inspection.
What to do now
Three things worth doing regardless of adoption pace.
Document how your processes actually work. Not how they were designed to work. This is prerequisite to any automation and valuable independently, and it is the step most organisations skip.
Decide accountability before deployment, not during an incident. Who owns each system, who can change it, who answers for its output. Write it down.
Plan the role composition change explicitly. If you remove the routine portion of a role, decide what fills it, and consider the intensity effect. Doing this deliberately produces better outcomes than letting it happen and dealing with the attrition afterwards.