RPA follows steps which do not change. AI agents read unstructured inputs, apply business rules, and send exceptions to people.
A supplier sends an order confirmation by e-mail. The supplier changes the delivery date and puts the new date in the message text. The purchase order number is in an attached PDF. A note in the PDF gives a different quantity.
An employee must find the purchase order and compare the information. The employee then updates the order or sends the case to a different person.
The work has clear steps, but the input does not have a standard format. An RPA system frequently stops when the supplier changes the words or document layout.
Robotic process automation (RPA) works correctly when each input has a specific format. An artificial intelligence (AI) agent has a different function. The AI agent reads different sources and does tasks in given limits.
Many enterprise processes start in e-mails, documents, spreadsheets, or free-text fields. These sources do not always use the same structure.
RPA for a process with steps which do not change
RPA does the same actions again in software applications. A software robot opens an application, copies a value, and puts the value in a different system.
This method is satisfactory for a process which does not change. For example, a report always has the same columns and folder. Each value always goes into the same field.
RPA also works with software which does not have an application programming interface (API). The robot selects buttons, fields, and menus in the same way as an employee. Thus, a company does not have to replace its primary systems.
Consistency is the primary benefit of RPA. The same input and rule give the same result. A process which does not change is easy to test and monitor.
Problems occur when the process changes. A supplier uses a new template. A customer puts two requests in one e-mail. The robot knows the steps, but the robot cannot identify the business information.
Inputs without a standard format
Unstructured data does not have a specific location. A delivery date is in a subject line, paragraph, table, or attachment. Different suppliers use different words for the same change.
One supplier writes "delivery postponed." A second supplier writes "new arrival date." A third supplier writes "delivery is three days late." All three messages give a change of the delivery date.
An RPA system examines given fields or words. A small change in the input frequently causes an error. The system then stops or gives the case to an employee.
An AI agent reads the information in the source. The agent finds the applicable values and the correct business record. The agent then applies the business rules.
This process has three primary parts:
- The agent reads the source.
- The agent changes the applicable information into structured data.
- The agent prepares or does the next action.
If information is missing or incorrect, the agent sends the case to an employee. The employee receives the source, record, and information about the problem.
RPA follows steps, and an AI agent examines a case
An RPA workflow identifies the next process step.
An AI agent also examines the information. Which information changed? Which record is applicable? Do the rules let the agent do the next action?
A quality department receives inspection reports from different suppliers. The reports have equivalent information, but they use different layouts and technical terms.
RPA uploads the report when each value has a specific location. An AI agent finds the inspection values and compares them with approved limits.
The AI agent must not have authority for all cases. An approved result moves to the next step. A missing certificate causes a request. Other results go to a quality specialist.
The agent operates in clear limits and uses approved sources. A person makes the decision when the information is not clear or the risk is high.
RPA is the correct selection for many tasks
AI agents do not replace all RPA workflows. Many tasks do not contain business information which a person must examine.
RPA is a good selection when inputs have a standard structure and each case follows the same steps. Examples include data transfer, report downloads, and data entry into software without an API.
An AI model adds cost, and the result is not always the same. The model does not give a solution for an applicable problem. A rule which does not change is easier to test and audit.
The primary point is how to divide the work. Where does the agent stop its checks? Where does the action which does not change start?
An AI agent reads an e-mail and prepares structured fields. A business rule checks the necessary values. An RPA robot puts the approved data in the application.
Each component does its given type of work.
Business rules for AI agents
An AI agent reads different inputs, but process design is necessary. Approved sources, a specific result, and action limits are also necessary.
Business rules divide routine cases from exceptions. Some rules apply in all cases. For example, the process does not continue without a necessary document. Other rules let the agent make a controlled selection.
A small date change in an approved range moves to the next step. A large date change goes to a production planner. The company must give these limits before the agent starts its operation.
Each exception must contain sufficient information for a decision. The source e-mail is not sufficient. The exception shows the applicable record, the change, missing information, and the applicable rule.
This design keeps responsibility clear. The employee does not have to do all the work again before the decision.
One process frequently uses more than one method
An enterprise process frequently contains structured records, rules, documents, and human decisions. One automation method does not give a satisfactory result for all parts.
RPA does the same actions again in software interfaces. APIs move data between modern systems. Business rules check known conditions. AI agents read unstructured data. People make decisions about cases with information which is not clear, high risk, or missing authority.
The supplier confirmation example shows a system which uses all methods. The AI agent reads the e-mail and its attachment. The agent finds the purchase order, delivery date, and quantity difference.
A business rule compares the new information with accepted limits. An API or RPA robot updates a routine case. A material change goes to the production planner.
No component controls the full process. The workflow gives each task to the applicable method.
Select the automation method from the work
A company must first examine the work, not the technology name. The company must identify the inputs, decisions, actions, and exceptions of one process.
First, find where information goes into the process. RPA is sufficient when all inputs use the same fields. An AI agent has a clear function when employees must read documents first.
Second, examine each decision. Identify rules which do not change, approved selections, and cases for which human responsibility is necessary. A process without agreed rules is not prepared for automatic action.
Third, identify the necessary result. The result must be a completed operation, such as an updated order, document request, or clear exception.
"Use AI" is not an operational result.
RPA gives solutions for a specific group of tasks which occur again. AI agents extend automation to work which includes language, documents, and different inputs.
The two methods have different functions. Deterministic tools do actions which do not change. AI agents examine different cases. People keep responsibility for decisions for which business knowledge or authority is necessary.
Where Onyven fits
Onyven builds the workflow around the split this article describes. Sonata runs the defined process end to end and sends exceptions to people; Orchestra keeps the work in one view when a decision needs a wider picture.