One of the most common questions when an organisation begins adopting artificial intelligence is which processes it could automate. It is a logical question.
AI can classify information, interpret documents, generate content, analyse large volumes of data, answer queries, recommend actions and, through agents, even perform certain tasks across other systems. The possibilities continue to expand.
But that is precisely why another question should come first: Which parts of the process should we not automate?
Maturity in AI does not mean removing human involvement whenever it is technically possible. It means understanding where automation creates efficiency, where artificial intelligence should extend people’s capabilities and where human judgement remains essential.
Automating a task is not the same as automating a decision
This distinction is fundamental. Many business processes combine mechanical tasks with others that require interpretation, context or decision-making.
Consider the management of an application or request. A system can receive it, identify the document type, extract information, check for missing data, query internal systems and prepare a proposed response. All of this can significantly reduce manual work. But it does not necessarily follow that the system should also determine the final outcome.
We can automate much of the work required to make a decision without automating the decision itself. This separation allows organisations to capture much of AI’s efficiency while preserving professional judgement where it is needed.
Analyse the task, not the job
There is another common simplification when discussing automation: analysing entire professions or jobs.
‘Can this job be automated?’ Usually, that question is too broad.
The same professional performs very different activities throughout the day. Some are repetitive. Others require expert knowledge. Some involve finding and organising information. Others require negotiation, interpreting exceptional circumstances or taking responsibility for a decision.
It is therefore more useful to break a process down into specific tasks and decisions.
- What information does the person receive?
- What do they do with it?
- Which parts follow relatively stable rules?
- Where do exceptions occur?
- Which decisions require context?
- What are the consequences of an error?
Only then can we determine what role artificial intelligence should play.
Four ways to introduce AI into a process
The choice is not simply between automating and not automating. We can think in terms of at least four levels.
- The first is automate.
AI performs a task within defined conditions without requiring human intervention in every case. This may be appropriate for repetitive activities with limited risk and easily verifiable outcomes. - The second is assist.
AI prepares information, identifies patterns, summarises documentation or proposes alternatives, while a person continues to perform the main task. - The third is recommend.
The system analyses information and proposes a decision, but responsibility for accepting, modifying or rejecting it remains with a person. - The fourth is reserve the decision.
AI may participate in supporting activities, but the central decision deliberately remains under human judgement.
Not every process needs to reach the first level. In fact, in many cases the greatest value may lie in the intermediate ones.
The risk of error matters
There is one particularly important variable when determining the appropriate level of automation: what happens when the system gets it wrong. Not all errors are equal.
If an AI system incorrectly classifies an internal document and the error can easily be detected, the consequences may be limited. If an automated decision affects a person, changes critical infrastructure, commits a significant amount of money or produces an external communication that is difficult to reverse, the analysis changes.
The greater the potential impact, the more attention should be paid to controls, validation and oversight. But impact is not the only variable.
The ability to detect the error matters too. A system may make frequent but obvious mistakes that are easy to correct. Another may fail very rarely but in ways that are difficult to detect.
Automation should consider both dimensions: the consequence of an error and our ability to identify it before it causes harm.
Whether we can go back matters too
Reversibility is another useful criterion. Some actions can easily be undone. A classification can be corrected. A draft can be discarded. A recommendation can be ignored.
Other actions are much harder to reverse. A communication that has already been sent, a financial transaction that has been executed, a change deployed to production or a decision affecting a third party may have consequences that persist even after the error is discovered.
The less reversible an action is, the stronger the case for introducing confirmation mechanisms, limits or oversight. This is particularly relevant with AI agents.
When a system can move directly from reasoning to action, the architecture needs to determine which actions it can perform independently and which require authorisation.
Context can matter more than accuracy
There is another limitation that cannot always be solved by improving the model. Some decisions depend on information that is not contained in the available data.
There may be tacit knowledge, exceptional circumstances, relationships between people, strategic priorities or information that was never formally recorded. A professional may incorporate that context naturally. A system can only work with what it can access.
This means that an AI system may achieve a high level of accuracy using the available information and still not be the right tool to make a particular decision autonomously.
The questions are if it is correct and if it genuinely has all the context required to decide.
Human oversight must add value
Keeping a person in the process does not automatically solve the problem either. There is a risk of creating purely formal oversight.
If a system produces hundreds of recommendations and a person has to validate them quickly, they may end up accepting them almost automatically. Human involvement exists, but human judgement has effectively disappeared.
When designing oversight, we therefore need to ask exactly what we expect the person to contribute.
- Do they need to verify information?
- Interpret an exception?
- Provide context the system does not have?
- Take responsibility for a decision?
- Intervene only when a particular risk threshold is exceeded?
Oversight works when the person has sufficient information, time and authority to challenge the system. Otherwise, adding an approval button provides little genuine control.
Automating to increase capacity
There is another way to think about automation with artificial intelligence. Instead of asking how many people we can remove from a process, we can ask how much more those people could achieve with the technology.
A professional who previously spent much of their time searching for information can receive it already organised. A technician can focus on exceptions while the system handles routine cases. A specialist can assess more situations because AI performs an initial analysis. A team can spend less effort producing information and more time interpreting it.
This changes the objective. AI ceases to be solely a tool for replacing tasks and becomes a way of extending professional capability.
In many knowledge-intensive processes, this may create considerably greater value.
Start with the process, not the AI
There is an understandable temptation whenever a new technological capability appears: immediately look for somewhere to use it. It happens constantly with AI.
We have a model capable of analysing documents, so we look for documents to give it. We have agents, so we look for processes they can execute. The order should be reversed.
First, understand the process. What outcome does it pursue? Which tasks make it up? Where is time consumed? Where do errors occur? Which decisions require expertise? What exceptions exist? What are the consequences of getting something wrong?
Then determine which technology can improve each part. At some points, that will be generative AI. At others, traditional automation. Elsewhere, analytics, business rules or system integration.
And in some cases, the right decision will be to keep the process under human responsibility.
Technology should adapt to the problem, not the problem to the technology.
Intelligent automation also knows when to stop
For years, we have associated digital transformation with automation. AI allows us to take that automation much further.
That is precisely why we need more judgement, not less. The question is no longer simply what a machine can do.
We need to decide what we want it to do, under which conditions, with what information, within what boundaries and under what level of oversight. For some tasks, the answer will be full automation. For others, AI will act as a copilot. Elsewhere, it will prepare a recommendation.
And there will be decisions where its most valuable role is to provide a person with better information so that they can exercise their own judgement. The most advanced organisation will not necessarily be the one that removes the most human involvement from its processes.
It will be the one that places automation and human judgement precisely where each creates the greatest value. Because knowing what we can automate is a technological capability.
Knowing what we should not automate is a design capability.