AI does not primarily remove job titles; it changes the composition of work. Whether that produces capability or risk depends on role design, decision rights and accountability — the organisational work, not the tooling work.
Almost every executive conversation about AI and work eventually arrives at the same anxiety: which jobs disappear? It is a reasonable question, and it is mostly the wrong one.
The question that determines your outcome is not which roles disappear. It is how the work gets divided between people, agents and human-AI teams — and whether somebody has designed that division, with named accountability at each step.
What changes when AI arrives
Work is not a list of job titles. It is a set of activities, each with a decision, a judgement, a system interaction and a record. AI changes the composition of those activities, often without changing the title attached to them.
Consider a claims handler. The title survives. The activity changes substantially: the reading, the summarising, the first-pass assessment and the correspondence drafting move toward automated or agent-supported execution; the adjudication of the genuinely contested case moves more heavily onto the human. Which is now the higher-value part of the role? Almost certainly the adjudication.
Notice what did not happen: nobody decided the job was eliminated. The job was recomposed, and if the recomposition was not designed deliberately, it happened by accretion — the human absorbed everything the system could not confidently do, and the exception queue grew until the role's value was mostly triage.
That is the failure mode. Not mass unemployment — quiet value destruction across a hundred roles, where each person spends their day on the residue that automation has not yet cleared.
Allocation is the real design decision
Cyberwisdom works with three allocation models, and choosing between them per activity is the core organisational design task of an AI programme:
Human-owned. The activity stays with people. AI supplies context, drafting and preparation; the human decides and remains accountable. This is right where consequences are severe and judgement is genuinely contested — legal judgment, clinical decisions, complex client relationships, and anything where the organisation's own values are in play.
Agent-owned. A bounded process is completed end to end by an agent within explicit authority limits, with defined triggers for escalation. This is right for high-volume, well-specified, verifiable work: first-line triage, document review, routing, standard drafting, evidence assembly.
Human-AI team. A person and an agent share the process with an explicit division of labour. The agent carries the volume; the human handles the exceptions and holds the accountability. This is often the most productive model, and the least frequently designed deliberately.
Note what all three require: an explicit decision about where the boundaries sit, who holds accountability at each boundary, and what triggers a hand-off. None of that is a tooling question.
Coordination, not headcount
There is a well-worn argument that AI will let smaller teams outperform larger ones. We would put it more carefully than that, because the unqualified version is not true and the qualified version is more useful.
A smaller team with well-designed AI-enabled roles can outperform a larger team using AI inconsistently. The mechanism is coordination, not headcount. A coordinated team moves work through the organisation with less friction, fewer handoffs, fewer waiting states and fewer duplicated efforts. An uncoordinated team with the same tools produces more activity and the same output — plus the overhead of coordinating a larger population through a changed process.
The practical implication is uncomfortable: the constraint on AI productivity in most enterprises is not the number of people or the number of licences. It is the number of places where work stops and waits for a human who has to reconstruct what they are being asked to do.
Decision rights get harder
Automation changes where judgement is exercised, and this is where most of the unresolved design work sits.
When an agent recommends and a human approves, the approval is often a formality — the human is unlikely to read what they are approving. This produces accountability theatre: a control that appears to keep a human in the loop while removing one from the actual decision. It is worse than no human checkpoint, because it generates false assurance.
The design discipline is to place human checkpoints where consequence lives, and to be willing to have none where they would be theatre. A checkpoint that no competent reviewer would ever override has consumed time and produced no assurance.
Capability is not the same as familiarity
There is a persistent confusion in executive conversations between the capability to use AI and familiarity with it. An organisation can be full of people who use AI daily and have no idea what they are doing — and, separately, contain people who have not adopted AI at all and could transform a process in a week if given the right context and permission.
This matters because the two require completely different interventions. Familiarity requires exposure. Capability requires context, authority and permission. Training is only one of them, and it is usually the easiest one to arrange, which is why it is overdone.
The interventions that actually move capability are unglamorous: giving people the authority to act on an agent's output, making the escalation path work, and making it easy to see what the system knew and why it recommended what it recommended.
What good looks like, practically
If an organisation has redesigned its work well for AI, several things are observable — and none of them are about tools:
- There is a documented map of which activities are human-owned, agent-owned or shared, and it is reviewed when the process changes.
- Escalation is routine rather than exceptional. Agents hand over work with context attached, and receiving one is an ordinary part of someone's day.
- Decision rights have been rewritten explicitly. There is a written statement of who decides what, including where an agent decides and what happens when it is wrong.
- Capability is visible. You can tell who in the organisation can redesign a workflow using AI, and it is not simply everyone who has a licence.
- The exception queue is shrinking. Volume is moving to the agent estate, and human time is concentrating on the work that genuinely requires a person.
- Measurement is behavioural, not attitudinal. You know what got faster and what got better, not just whether people say they like the new tools.
The argument for treating this as the main work
There is a reason to spend an executive agenda on role design rather than on tool selection. Tool selection is a decision with a bounded set of options, a vendor comparison, and a defensible answer. Role design has no such comfort — it exposes genuine disagreements about what the business is for, who is accountable for what, and where the organisation is willing to be wrong in public.
Those disagreements need resolving anyway. The AI programme simply forces them to the surface earlier. Organisations that treat that as an inconvenience tend to get a shadow operating model. Organisations that treat it as the actual work tend to get a capability that compounds.
The tooling will keep changing, and it is changing fast. The work decomposition, the decision rights and the accountability model are what remain — and they are, in the end, the whole of what determines whether the organisation is more capable than it was last year.