Business process automation: where to start

Business process automation involves using technology to automatically execute and coordinate tasks, data exchanges, and steps within an operational workflow, reducing manual work and making the process more efficient and easier to manage.

Depending on the problem being addressed, automation can deliver different benefits, which can be grouped into four main areas.

• Operational efficiency: reduces the time spent on repetitive tasks and manual interventions, freeing up resources for higher-value activities.
• Scalability: makes it possible to handle larger volumes of cases, requests, or transactions without proportionally increasing workload and required resources.
• Quality and reliability: reduces rework, data transfer errors, and redundant steps, making the exchange of information between systems and process stages smoother.
• Visibility and control: makes it easier to monitor timing, volumes, and exceptions, identify bottlenecks, and assess results over time.

These areas also provide a useful starting point for understanding where automation can have a tangible impact.

The first step in implementing process automation is identifying the activities with genuine potential for improvement. The clearest signals often emerge from day-to-day operations: repetitive tasks, manual handoffs between systems, recurring checks, or growing volumes that require an increasing amount of effort.

Once a potential area for improvement has been identified, it is important to examine it within the context of the entire workflow. Take document classification, for example: which system does the document come from? What information does it contain? Where should the extracted data be recorded? What activity is triggered next? Which cases require verification?

The same logic applies to an email, an administrative case, a CRM update, or order processing. Analysing the workflow end to end helps determine where to intervene, which dependencies need to be considered, and whether automation actually solves the inefficiency or simply shifts it to another stage of the workflow.
To reconstruct the process accurately, it is useful to involve the people who carry it out every day, as they are familiar with exceptions and undocumented steps. When volumes justify it, system logs can also be analysed through process mining to observe the actual workflow, which often differs from the one described in formal procedures.
If redundant steps emerge, it is better to simplify them before automating them: automating an inefficient process only makes it inefficient faster.

After identifying the activities with the greatest potential for improvement, the next step is to determine which ones to prioritise. For a first project, it is generally best to choose an activity that is relevant, well-defined, and sufficiently limited in scope to make it possible to assess both feasibility and results. Four questions can help narrow the field.

Does the volume justify the intervention?
A frequent activity can consume a significant amount of time even if each individual execution takes only a few minutes. It is therefore important to consider frequency, time spent, people involved, and rework. An activity performed only a few times a month may remain a low priority even if each execution is demanding, while a short task repeated hundreds of times can consume more resources than it initially appears.

Is the workflow logic clear?
The clearer the steps, decisions, and exceptions are, the easier it is to identify what can be automated and which approach to use. When the workflow is still poorly defined, mapping it clearly is already a necessary part of the project, even before selecting a technology.

Is the data available?
It is important to verify where the required information is stored, in what format, and how accessible it is. Data quality and availability directly affect feasibility: information may exist but be spread across different systems or require manual steps to retrieve it, making the project more complex than it first appears.

Can results and exceptions be monitored?
For a first initiative, it is useful to choose an area where both outcomes and exceptions can be monitored clearly, while maintaining human oversight where necessary, especially in more sensitive or high-impact stages.

A good starting point therefore combines operational impact, feasibility, and control.

Once the process has been selected, the next step is to define the technological approach that best fits the workflow structure, the systems involved, and the level of variability in the activities. The approaches described below are not mutually exclusive: in most projects, they are combined.

• Application integrations, APIs, and orchestration components: connect systems, data, and applications so they can exchange information and trigger actions in a coordinated way. They are useful when a process spans multiple tools and manual handoffs between systems need to be eliminated.
• Workflow engines: manage the sequence of activities within a process by defining rules, approvals, responsibilities, and exceptions. They are particularly useful in workflows involving multiple people or several operational stages.
• Robotic Process Automation (RPA): automates repetitive, rule-based activities by replicating actions that a person would perform on existing applications and interfaces. It can be especially useful when systems cannot be easily integrated via APIs. Because it operates through user interfaces, however, it is more sensitive to changes and requires maintenance: when API integration is possible, it is generally the more robust option.
• Artificial Intelligence: extends automation to steps that require interpreting, classifying, or analysing less structured information. In email routing, for example, a rule can manage messages identified by specific codes or predefined fields; when routing depends on the meaning of the request, AI can interpret the content and support classification. Since a model’s output is probabilistic, it is good practice to define confidence thresholds: above the threshold, the case proceeds automatically; below it, it is routed for human review.
• AI agents: systems based on language models that, given an objective, can plan and execute multiple steps autonomously by interacting with applications and data through APIs and dedicated tools. They are suited to more complex and variable activities, such as managing a request end to end, but require clear rules on what they are allowed to do, traceability of their actions, and human control points.

When AI becomes a stable part of day-to-day operations, it is important to design AI systems integrated into business processes, connected to the data and applications that support the workflow.

Once the technology has been selected, automation needs to be embedded into the existing operational flow by designing the right integration between systems and applications. The automation should interact with the applications, systems, and process steps already in use, avoiding the creation of isolated solutions.

For example, if a system classifies a request or extracts information from a document, the output must correctly feed the next activity: updating a management system, triggering a workflow, generating a notification, or initiating a control.

Integration is therefore part of the use case design itself: input, processing, and output must work as components of the same process.

Governance aspects are also part of the design: who is responsible for the automation once it is in production, how the actions performed are tracked, how personal data is handled in compliance with the GDPR, and, for AI components, which requirements arise from the AI Act. An automation without a clear owner and ongoing maintenance tends to lose effectiveness over time as systems, rules, and data evolve.

Before launching the project, it is necessary to define the desired outcome and the KPIs that will be used to measure it, ideally referring back to the benefit areas identified at the outset: time, operational capacity, quality/reliability, or workflow control.

The starting point is a baseline, meaning a reference situation that makes it possible to compare performance before and after the intervention. Depending on the use case, relevant metrics may include processing times, volumes handled, manual effort required, rework, SLA compliance, or the number of cases managed with the same level of resources. In projects involving AI, additional indicators may be useful, such as classification accuracy, output quality, or the percentage of cases requiring human review.

Finally, it is advisable to proceed in stages: a pilot project with a limited scope and predefined success criteria makes it possible to assess results against the baseline and gradually involve the people concerned before extending automation to larger volumes or additional processes.

What we have described in this article is the path that turns an operational need into a first project with clear objectives and a well-defined scope. However, when the project involves the use of artificial intelligence, it also becomes important to understand whether the company has the right conditions to introduce AI into its processes and move the use case beyond experimentation.

This is the topic we explored in the webinar “AI in Italian Companies: Who Is Succeeding and What Did They Do First?”, where we looked at how to identify the initiatives with the greatest potential, which factors influence the transition from experimentation to production, and which conditions should be assessed before investing in an AI project. The goal is to focus resources and attention on the most sustainable use cases and those with the greatest potential to generate operational results.

If you are considering how to introduce AI into your business processes and want to understand where to start, watch the webinar, available in Italian.