# Process automation with AI: Where the leverage is greatest

Author: Marin Niehues, Niehues Consulting · Published: 2025-06-16 · Updated: 2026-11-16 · AI Adoption & Operations
Source: https://www.niehuesconsulting.com/blog/process-automation-with-ai/

Process automation with AI has the greatest leverage in high-volume processes where employees read and classify unstructured information every day. It means automating workflows with language and vision models that understand free text and documents. This applies to inboxes and stacks of documents, as well as to handovers between departments. There AI now does what rule-based automation fails at: it understands content that follows no fixed structure, with an error rate that can be measured and controlled through sampling. The return depends on whether the organization anchors the automated workflow in day-to-day operations.

This article provides an overview and does not constitute legal advice.

**The greatest leverage lies in processes with high volume and variable inputs, provided an error causes manageable costs. There every improvement immediately affects many transactions.**

## What is the difference between RPA and AI automation?

RPA replicates fixed click sequences on structured data, while AI automation understands changing content and makes classifications a person used to make. Rule-based automation such as robotic process automation (RPA) or workflow rules follows fixed rules and requires structured inputs: one field, one value, one target. A software robot transfers data from one screen to another, just as a person would. As soon as inputs vary, because customers phrase things freely or suppliers send documents in different formats, rules reach their limit.

Language models handle exactly this variation. They read a text and assign it to a category, extract data and prepare a decision. In practice the two approaches complement each other. The combination is called intelligent automation: AI understands the input, and RPA or an interface executes the step in the system. Hyperautomation describes the next step, the coordinated combination of several tools with the aim of automating entire process chains end to end. AI agents go a stage further, planning and executing several steps on their own.

## Which processes are suited to AI automation

The selection question therefore shifts from technical feasibility to business value. Almost any process involving text can now be automated as a prototype. Which one delivers the greatest return is decided by a grid of four criteria.

**Four criteria for process selection**
- Volume: how many transactions run per week, and how much handling time does each one take?
- Variability: how much do the inputs differ, and how many special cases occur?
- Share of judgment: which part of the work is routine, and which requires judgment?
- Cost of errors: what does a wrong result cost, and how quickly is it noticed?

A simple ranking follows from the criteria. Processes that score well on all four are the starting candidates. Processes with high error costs follow at a later stage, in which AI makes suggestions and people decide.

## Four application areas with examples

In practice, successful initiatives cluster around four patterns. What they share is that most of the handling time goes on reading and searching as well as transferring data.

- Intake processing: requests or applications are classified automatically and routed to the right place. Example customer inbox: AI recognizes concerns such as a complaint and extracts customer number and case. The case lands in the right team with a draft reply.
- Data transfer between systems: information is extracted from documents and transferred into core systems, eliminating media breaks. Example incoming invoices: AI reads paper and PDF invoices and matches them against purchase order and goods receipt before posting starts.
- Decision preparation: the system summarizes documents, checks them against defined criteria and presents a proposal to the case handler. Example claims notification: AI links photos and description to the policy and proposes the handling route.
- Knowledge retrieval: employees get answers from internal policies and manuals instead of asking colleagues.

For handling documents, including the effect of e-invoicing on incoming invoices, the article on [AI document processing](/blog/ai-document-processing/) describes the technology in detail.

The return comes from volume. Every minute saved per transaction is multiplied by the volume, which is exactly why the intake processes with the largest volumes head every ranking. A second effect adds to this: when employees no longer have to read and transfer, time is freed for queries and exceptions. A field study of more than 5,000 customer service agents shows that an AI assistant raises issues resolved per hour by 14 percent on average, and by 34 percent for new and less experienced agents.

## Understand the process before automating it

A process with detours still contains those detours after automation, only faster. Every process automation therefore begins by mapping the current workflow. Business analysis records how many transactions arise, which variants exist and where handling time is actually lost. The bottleneck often lies elsewhere than assumed, in the query back to the customer or in an approval.

Process mining helps here. The method, developed largely by the computer scientist Wil van der Aalst, reconstructs the actual workflow from the timestamps that ERP and workflow systems log anyway, and shows variants and waiting times. It pays off when the process runs entirely in systems and has several thousand transactions a month. If much of it runs through email and paper, interviews and a sample of cases deliver a robust picture faster.

The business case emerges from the mapping. It sets benefit against effort per process and provides the basis for the sequence. The sequence is a management decision because each process ties up business unit time for testing and acceptance. How to calculate benefits and costs robustly is described in the article on the [AI business case](/blog/ai-business-case-benefits-and-costs/).

**Calculation path: at what rate does automation pay off?**
1. Current cost per transaction: handling time times hourly rate.
2. Cost per transaction with AI: operating cost per transaction plus share of exceptions times handling time per exception times hourly rate.
3. Annual saving: (1 minus 2) times annual volume.
4. Payback: one-off costs for prototype and integration divided by the annual saving.
5. Minimum rate: insert the desired payback period and solve for the share of transactions that must run through without intervention.

If the minimum rate is above what the prototype measures, the process is the wrong candidate for now. The calculation also shows that one exception with a long handling time depresses the return more than many short ones.

## From prototype to operations

A prototype proves its benefit only when it runs in the real workflow with real transactions. A test with prepared sample data says little about the effect under the fluctuations of daily business. The prototype therefore measures three values in the live operation of a team:

- Handling time per transaction compared with the manual workflow
- Error rate compared with previous handling
- Share of transactions that run through without human intervention

The third value determines the return and also shows how well the rules for exceptions work. A process is ready for operations when AI completes the majority of transactions on its own, the error rate is at most at the level of manual handling, and the remainder goes to people with a proposal.

The transition to operations decides whether the benefit lasts. Every automated process needs a responsible person in the business unit and a defined route for exceptions. Ongoing quality control through sampling completes the set-up. Whoever handles the exceptions in the future takes on a more demanding task and needs suitable onboarding. Why initiatives often stall at this point is explained in the article [Why AI pilots do not make it into operations](/blog/why-ai-pilots-fail-to-reach-operations/).

## How long does it take to automate a process?

A prototype in the real workflow is ready in six to eight weeks, and an integrated process with its own data needs three to six months to reach regular operations. The range is set by three factors: the number of interfaces to core systems, the availability of data and the time the business unit provides for testing and acceptance. Each additional system extends the duration. Transferring the same approach to a second, similar process is considerably faster because components and operating rules are already in place. The overall frame is described in the article [From pilot to scale](/blog/from-pilot-to-scale-ai-operations/).

**Automation approaches compared**
Criterion | Rule-based automation | Individual AI tools at the workplace | With Niehues Consulting
Unstructured inputs | Only with fixed formats | Possible, varies by person | Processed uniformly in the process, with rules for exceptions
Measuring benefit | Measurable per bot | Barely measurable | Metrics per process from the prototype onwards
Responsibility | With IT | With each individual | Process owners with defined escalation
Transition to operations | Through an IT project | Does not happen, hence no scaling | Planned from the start, with enablement of the business units

## Niehues Consulting as a partner for process automation

[Niehues Consulting](/) in Düsseldorf supports process automation from selection through to adoption in the business unit. The approach follows the Niehues Delivery Model with its phases understand, build and introduce. AI management produces business analysis and requirements, the business case and a sequence of applications. Prototypes prove the benefit in the real workflow, and operationalisation transfers them into operations with internal or external teams.

For process automation, three of the model's review questions matter most. What is the baseline, and how do we measure success? Without measured handling time per transaction, every saving remains an estimate. Does the prototype prove the benefit in the real workflow with real data? Only then is the rate known that decides the business case. Who owns operation from day one? Exceptions and sampling need a named person in the business unit.

Marin Niehues leads initiatives from the delivery perspective: he builds organizations that deliver and brings AI into operations. The free initial consultation lasts 30 minutes and clarifies which process offers the greatest leverage: [arrange an initial consultation](/#kontakt).

**The order of decisions**
Process automation with AI delivers the greatest return where volume and variation meet. The organization first selects the process using the four criteria and then proves the benefit in the real workflow. After that it sets responsibilities for operations. Whoever steers these steps from one hand reaches regular operations with the least loss of time.

**Sources**
- Brynjolfsson, Li, Raymond: Generative AI at Work, NBER Working Paper, 2023
- van der Aalst: Process Mining: Data Science in Action, book, 2nd edition 2016
- IEEE: IEEE 2755 Guide for Terms and Concepts in Intelligent Process Automation, standard, 2017

## Frequently asked questions

### Which processes can be automated with AI?

Suitable processes are those in which people read and classify texts or documents before passing requests on. The higher the volume and the more variable the inputs, the greater the leverage. Processes with high error costs start with AI suggestions and a decision by people.

### How does AI automation differ from traditional automation?

Rule-based automation such as RPA follows fixed rules and requires structured inputs. AI automation processes free text and documents in changing formats. The two approaches complement each other, and many processes combine them as intelligent automation.

### What is hyperautomation?

Hyperautomation is the coordinated combination of several automation tools with the aim of automating entire process chains end to end. It includes RPA and AI, supplemented by process mining and workflow systems. For executive management, the term matters less than which process becomes measurably faster or cheaper.

### Do you need process mining before automation?

Process mining pays off when the process runs entirely in systems and has high volumes. It then shows variants and waiting times from real data. If the process runs mainly through email or paper, interviews and a sample of cases are the faster route.

### How can you tell whether automating a process pays off?

Automation pays off when the measured share of transactions without intervention exceeds the minimum rate from the business case. The business case sets the benefit from handling time and volume against the effort for set-up and operation and takes error costs into account. A prototype in the real workflow supplies the measurements that make this calculation robust.

### How much does AI process automation cost?

Costs arise in three blocks: prototype and integration as one-off costs, ongoing model and operating costs per transaction, and the handling of exceptions. The largest item is usually integration into core systems. Whether the investment pays off is shown by the calculation path above using your own values.

### Who is responsible for an automated process in operation?

Responsibility lies with the business unit that uses the process. A named person steers exceptions and quality, while IT or an external team runs the technology. This allocation is settled before regular operations start.
