Automating more cases can reduce manual work, but it also changes the human capabilities an organisation needs.
An organisation introduces artificial intelligence into its processes. The system begins resolving repetitive requests, operations governed by clear rules and situations that do not require special interpretation.
The percentage of automated cases increases. The volume of manual work decreases.
The indicators show an improvement.
But there is a question those indicators cannot answer: what kind of work still reaches people, and what capabilities does the organisation need to resolve it?
Because automating part of a process does not simply change the number of cases a team handles. It can also change the nature of those that remain.
When routine work disappears, what remains changes
Consider a claims management process.
AI can resolve common requests when sufficient information is available and the conditions for resolution are clearly defined.
But what happens when the data does not match? When two rules lead to different outcomes? When a claim involves circumstances that the procedure does not cover?
Those cases may require human intervention.
And they do not necessarily require the same amount of time, knowledge or effort as requests resolved automatically.
This concentration of exceptions is not inevitable. It depends on which cases the organisation automates, how it defines escalation criteria and what tools it provides to its professionals.
What matters is that the volume of work that disappears does not, on its own, describe the complexity of the work that remains.
The human contribution does not begin where technology ends
In the practical application of AI, people should not be reduced to validating responses or resolving incidents that the system has been unable to close.
Their contribution can be much more specific.
When a claim contains contradictory information, it may need to be checked against business knowledge. When the rules do not cover a particular circumstance, professional judgement helps interpret the context. And when a resolution has sensitive consequences, human intervention provides contextual assessment and assumes the appropriate responsibility.
There is another function that often falls outside automation metrics: feeding what has been learned back into the process.
A resolved exception may reveal missing information, a rule that needs reviewing or an escalation criterion that is not working correctly.
If that knowledge does not return to the process, the organisation will continue encountering similar problems without improving its ability to resolve them.
Human intervention is not simply about closing what AI leaves unresolved. It also helps explain why those cases remain unresolved.
This requires professionals to have access to information, the authority to act and mechanisms for communicating their conclusions. Simply having a person involved in the process does not guarantee a better outcome.
The mistake of sizing teams according to automation rates
Suppose an organisation manages to automate a significant proportion of its requests.
It would be tempting to translate that percentage directly into workforce planning.
Fewer manual cases. Fewer hours required. Less capacity to maintain.
But that equivalence can be misleading.
Automation can generate significant savings in time and costs. What should be avoided is translating the percentage of automated cases directly into an equivalent reduction in resources.
A routine case and an exception involving incomplete information do not necessarily require the same effort. The latter may involve consulting different sources, coordinating several departments, reviewing previous decisions and justifying a resolution.
Furthermore, two processes with the same automation rate may require very different operational capabilities. Resolving time-consuming exceptions is not the same as managing cases that depend on several departments, specialised knowledge or decisions requiring escalation.
That is why, before reorganising a team, it is not enough to know how many cases AI has absorbed.
Organisations also need to understand how long the remaining cases take, what knowledge they require, how many interventions they involve and what conditions allow them to be resolved correctly.
This calls for a review of the operating model: how work is distributed between AI and people, which criteria determine when exceptions are escalated, what capabilities teams need and how knowledge gained from each resolution is incorporated into the process.
At Cognodata, we approach this question by connecting data analysis with an understanding of the process and the reality of the teams operating it.
The objective is to identify which cases AI can resolve, which require human intervention and what information, criteria and capabilities must support that intervention.
This makes it possible to review team capacity without relying solely on automation rates, integrate AI with existing systems and procedures, and establish mechanisms through which resolved exceptions contribute to process improvement.
Research into automation has highlighted this difficulty for decades. In Ironies of Automation (1983), Lisanne Bainbridge described how automated systems can leave people responsible for precisely the most difficult interventions.
More recently, a Gartner survey published in February 2026, involving 321 customer service and support leaders, reported that more than 80% of the organisations surveyed expected to expand the responsibilities of their human agents as they introduced AI into customer service and support operations.
This figure reflects organisational expectations. It does not demonstrate that all remaining cases are more complex. But it does point to an important issue: automating tasks and reorganising human work are related decisions that should not be confused.
Measuring what disappears is not enough to manage what remains
The automation rate is a useful indicator. It shows what proportion of a process is completed without human intervention.
But it cannot, on its own, determine the capabilities an organisation still needs.
To do that, organisations must also examine the remaining workload: its composition, the effort required to resolve cases, dependencies between teams and the knowledge generated by each exception.
This is where the human contribution takes on a concrete operational dimension. Not as an alternative to AI, but as part of the process in which AI operates.
Well-integrated automation can improve efficiency and free people from routine tasks, allowing professional knowledge and judgement to focus on situations where they provide greater value.
This requires the remaining work to be supported by sufficient information, appropriate capabilities and clear responsibilities.
Before reorganising resources based on the volume of cases resolved by AI, one question needs to be answered:
Do we really understand what work we are leaving in people’s hands?
Frequently Asked Questions About AI Automation and Human Work
How does human work change when AI automates routine tasks?
Automation can reduce the volume of repetitive tasks and change the types of cases that reach people. Depending on automation and escalation criteria, the remaining work may require more interpretation, coordination or specialised knowledge.
Why is the automation rate insufficient for workforce planning?
Because it indicates how many cases are resolved automatically, but not how much effort the remaining cases require. Workforce planning must also consider their complexity, resolution time, required expertise and dependencies between teams.
What role should people play in AI-automated processes?
People can verify information, interpret unforeseen circumstances, intervene in sensitive decisions and contribute knowledge that helps improve the process. Their involvement should be supported by sufficient information, the authority to act and clearly defined responsibilities.
How can organisations manage exceptions left unresolved by AI?
By defining escalation criteria, assigning responsibilities, providing access to the necessary context and establishing mechanisms for incorporating lessons learned from each resolution. This allows organisations to review both process rules and the capabilities their teams need.