Process automation: preparing decisions from data, leaving exceptions to people
Process automation starts with a measurement, because a poor process is still poor once it is automated and merely runs faster.
- What it delivers: In our model calculation for a fictitious company, the new workflow saves 54,600 euros a year in the middle scenario, out of today's 135,000 euros in checking costs. This figure already deducts monitoring and the cost of running the model, but not yet the one-off investment. The flattering calculation based on the automation rate would have promised 108,000 euros. In the cautious scenario the new workflow ends up costing more than the current one.
- What it requires: a large number of similar cases, stored decisions, and an outcome that can be verified later.
- When it does not pay off: when a fixed rule is enough, when volumes are small or when the workflow itself is disputed.
Process automation: definition and distinctions
Process automation means handing recurring steps of a business process to a system that either follows fixed rules or uses a model learned from data. This page deals with the second case: steps in which a person currently checks, approves, or plans, and in which the right decision depends on data.
Four terms are often confused. Rule-based automation carries out whatever can be written as an if-then rule. Robotic process automation, or RPA, is software that performs rule-based routine tasks on structured data [1]. Intelligent automation combines RPA with artificial intelligence, such as machine learning [1]. On this page, AI automation means the part in which a model prepares a decision from data, using a forecast, a classification, or anomaly detection.
Where rules are enough, rules are the better choice, because they are cheaper and easier to trace. A model only pays off where a fixed rule would be wrong in too many cases. We sell neither RPA licenses nor workflow orchestration software. Our contribution is the model behind the decision and its integration with your business application. Before that, we test whether the model decides better than the current workflow, and we tell you the result even when it is negative.
The call to redesign the workflow first is not new. In 1990, Michael Hammer argued in the Harvard Business Review that companies should redesign their processes instead of simply automating the existing ones [2]. As early as 1983, Lisanne Bainbridge had described, using industrial plants as her example, how automation can expand the problems with the human operator [3].
How process automation with AI works
If you want to automate business processes, you first need figures on the current workflow. Without a baseline, you can neither demonstrate a benefit later nor notice when something goes wrong.
| Step | What happens | What it produces |
|---|---|---|
| 1. Measure the process | volumes, handling times, error rate, and variants are derived from system data | the baseline for every comparison |
| 2. Separate the decisions | for each step it is determined whether a fixed rule is enough or whether the decision depends on data | a list that assigns each step to rules, a model, or a person |
| 3. Build the model | a model learns how to classify a case from past decisions and their outcomes | a proposal for each case, with an indication of how certain it is |
| 4. Test against the current workflow | the model is compared on held-out cases with the staff who handle such cases today and with a simple rule | the comparison and the threshold above which a case counts as an exception |
| 5. Integrate into the process | the result appears in the business application as a proposal that a person confirms | a prepared decision at the workstation |
| 6. Monitor and retrain | spot checks and the comparison with later actual outcomes continue | a model whose quality is measured continuously |
You can recognize a reputable provider by the first two steps. They measure your current workflow before making an offer and tell you which steps can manage without a model.
Our production example, set in a fictitious sample company, shows what step 5 can look like. There the model's output appears as a maintenance alert in the maintenance management system: “Machine 14 has a failure risk in nine days, recommended maintenance window: Sunday night shift.”
Worked example: AI automation in replenishment, 60,000 order proposals
The following example is a model calculation for a fictitious company. A mid-sized wholesaler plans its purchasing in an ERP system that generates order proposals according to fixed rules. A case here is a single order proposal, that is, an item with a quantity, a date, and a supplier. A planner checks each proposal, adjusts it if necessary, and approves it.
The model classifies every proposal either as a standard case, if it flags nothing unusual, or as an exception with a reason, such as unusual demand. Approval stays with the planner, who approves the standard cases in bulk and handles the exceptions one by one. Here, the automation rate means the share of proposals that are not checked individually.
| Assumption | cautious | middle | favorable |
|---|---|---|---|
| Order proposals per year (cases) | 60,000 | 60,000 | 60,000 |
| Checking time per proposal today, in minutes | 3 | 3 | 3 |
| Cost per working hour, in euros | 45 | 45 | 45 |
| Automation rate (share of standard cases), in percent | 70 | 80 | 90 |
| Exceptions checked individually, in percent | 30 | 20 | 10 |
| Time per exception, in minutes | 8 | 6 | 4 |
| Monitoring per year, in hours (bulk approval, spot checks, comparison with actual figures) | 480 | 320 | 160 |
| Running the model per year, in euros (computing power, retraining, interface) | 18,000 | 12,000 | 9,000 |
Checking currently takes 3,000 hours a year, which at 45 euros an hour comes to 135,000 euros. A flattering calculation multiplies this amount by the automation rate and arrives at 108,000 euros in the middle scenario. It leaves out three items.
Exceptions are the first item. The 12,000 proposals checked individually are the difficult ones and take 6 minutes instead of 3, because the planner has to follow the model's reasoning. That comes to 1,200 hours or 54,000 euros. The second item is monitoring, at 320 hours or 14,400 euros. The third is the cost of running the model, at 12,000 euros. This amount is an assumption we have set, and no price can be derived from it. That leaves a net effect of 54,600 euros a year out of the 135,000 euros, or roughly half of what the flattering calculation promised.
| Scenario (amounts in euros) | Flattering calculation | Exceptions | Monitoring | Operation | Net effect per year |
|---|---|---|---|---|---|
| cautious | 94,500 | 108,000 | 21,600 | 18,000 | −12,600 |
| middle | 108,000 | 54,000 | 14,400 | 12,000 | 54,600 |
| favorable | 121,500 | 18,000 | 7,200 | 9,000 | 100,800 |
In every scenario the net effect stays below the flattering calculation, because monitoring and operation are added and an exception takes longer than today's check. In the cautious scenario the new workflow costs more than the current one. In that scenario, exceptions and monitoring together take 2,880 hours, almost as many as today's checking, and it is only the cost of running the model that turns the net effect negative.
For your own rough calculation, use this formula:
Net effect per year = today's checking costs − cost of exceptions − monitoring − running the model
Today's checking costs are the number of cases multiplied by the checking time and the hourly rate. The cost of exceptions is the number of cases multiplied by the exception rate, the time per exception, and the hourly rate.
Three caveats apply to this formula. The one-off cost of building and integrating the model sits outside it and is set against the net effect. In addition, the calculation initially yields only freed-up hours. In the middle scenario 1,480 of 3,000 hours are freed up, and whether they turn into money depends on how they are used. Finally, the calculation only counts what is contained in today's checking costs, so it gives no euro amount for stockouts and overstock. Both can still be measured, because every order later shows whether the item ran out or sat on the shelf. The pilot compares both figures with the situation before the model, and how that comparison turns out is open.
With 6,000 proposals a year, today's checking costs come to 13,500 euros. If running the model costs as much as in the middle scenario, that item alone accounts for 12,000 euros. In that case individual checking remains the more economical solution.
Module 8 of our chocolate series shows, using a verification case, how a model pre-sorts and a person decides.
When process automation pays off and when it does not
Process automation with a model pays off when three conditions come together: a high volume of the same kind of case, past decisions stored with their outcomes, and a way of establishing afterwards whether a decision was right. Without the third condition, nobody can measure whether the model is doing a good job. In the worked example it is met, because stockouts and overstock show for every order how good the decision was.
Four wrong conclusions are easy to draw:
- The automation rate is not the saving. Without individual checks the easy cases go through, while the expensive ones stay with the planner.
- A disputed process stays disputed with a model. If two planners decide the same case differently, the model learns the contradiction as well.
- Full automation can only come at the end. First the model makes proposals that a person confirms. The share of cases without individual checks grows with the measured results.
- Responsibility stays with a person. It must be settled before the start who answers for a wrong decision and who handles the exceptions.
Where a decision concerns natural persons, such as private customers or employees, legal limits apply as well. Article 22 of the General Data Protection Regulation (GDPR) restricts decisions that are based solely on automated processing and that produce legal effects concerning the person or similarly significantly affect him or her [4]. Whether a project falls within its scope has to be assessed legally case by case.
We advise against automation when a fixed rule decides the case just as well, when volumes are small or when the process itself is still disputed.
Process automation in the data science process
A process automation project runs through the six phases of CRISP-DM, the widely used process model for data science projects [5]. Three of them are decisive for a project of this kind. Phase 1 defines which decision the model prepares and what an error costs. Phase 5 tests the model against the current workflow and sets the threshold for exceptions. Phase 6 brings the result into the business application and monitors it. For running the models, most projects at mid-sized companies need no more than a partly automated pipeline that a person triggers. That holds as long as only a few models are involved, the data changes over weeks rather than minutes, and a person approves every model change.
All six CRISP-DM phases at a glance Phase 5: model validation Phase 6: deployment and monitoringFrequently asked questions about process automation
What is process automation?
Process automation hands recurring steps of a business process to a system that either follows fixed rules or has learned from data how a case should be classified. People remain in charge of the exceptions and of monitoring, and they continue to bear the responsibility for the decision.
What is the difference between process automation with AI and robotic process automation?
Robotic process automation follows fixed rules and suits steps that always run the same way. An AI model comes in where the decision depends on the individual case, for example whether an order proposal can be accepted unchanged. The two approaches can be combined.
What does process automation cost and when does it pay off?
The calculation has four components: today's checking costs, the cost of exceptions, monitoring, and running the model. The balance of these four components is the net effect per year, which is set against the one-off investment. With small volumes the numbers are hard to make work, because a base level of monitoring and operating cost remains.
What data does AI automation need?
The model needs past cases with all the information that was available to the person handling them, together with that person's decision and the later outcome. If this information only exists in emails or free text, the project starts with clean data capture, and the model follows later.
How long does it take for process automation with AI to take effect?
The time depends on what comes before and after the model. First the current workflow has to be measured, and it must be clear whether past decisions are stored in a form that can be analyzed. Afterwards, the model initially only makes proposals that a person confirms. The measured results determine how long that phase lasts.
How does a process automation project begin?
It begins with a measurement that shows how many cases there are, how long each takes, how many errors occur, and which variants exist. Then each step is assessed to see whether a rule is enough. A model and a pilot with human confirmation follow only for the steps in which the decision depends on data.
Sources
- Wil M. P. van der Aalst, Martin Bichler, Armin Heinzl: Robotic Process Automation. Business & Information Systems Engineering 60 (4), 2018, pp. 269-272, DOI 10.1007/s12599-018-0542-4; and Kam K. H. Ng, Chun-Hsien Chen, C. K. M. Lee, Jianxin Roger Jiao, Zhi-Xin Yang: A systematic literature review on intelligent automation: Aligning concepts from theory, practice, and future perspectives. Advanced Engineering Informatics 47, 2021, article 101246. https://doi.org/10.1016/j.aei.2021.101246
- Michael Hammer: Reengineering Work: Don't Automate, Obliterate. Harvard Business Review, July-August 1990, reprint 90406. https://hbr.org/1990/07/reengineering-work-dont-automate-obliterate
- Lisanne Bainbridge: Ironies of Automation. Automatica 19 (6), 1983, pp. 775-779. https://doi.org/10.1016/0005-1098(83)90046-8
- Regulation (EU) 2016/679 of the European Parliament and of the Council of 27 April 2016 (General Data Protection Regulation), Article 22 (“Automated individual decision-making, including profiling”). https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32016R0679
- Christoph Schröer, Felix Kruse, Jorge Marx Gómez: A Systematic Literature Review on Applying CRISP-DM Process Model. Procedia Computer Science 181, 2021, pp. 526-534. https://doi.org/10.1016/j.procs.2021.01.199
Run the numbers on your process with us
The number of cases per year, the time per case and an hourly rate are enough for a first assessment of whether process automation pays off for you.
Book a free initial consultation