Lunatec · A Guide for Decision-Makers
Which automation for which process
and when?
A practical guide for decision-makers. We show how to systematically assess which business processes in your organisation are suited to RPA, Intelligent Document Processing and Agentic AI, with thresholds, ROI logic and real examples from projects with 120+ clients.
E d i t o r i a l
Why this dossier
If you are reading this guide, you are probably facing one of these situations: accounting can no longer keep up with the invoice volume. HR manually processes thousands of sick notes per month. Procurement keys order confirmations from 800 different supplier formats into SAP. Or customer service is running a 39-day turnaround, with customer complaints growing louder by the day.
The answer is automation. But in 2026 “automation” is an umbrella term for at least three technology tiers with very different strengths, costs and prerequisites. Which of them is right for which process; that is the real decision.
This guide gives you the tools to judge that for yourself. Since 2017, we have worked as a UiPath Diamond Partner and Microsoft Partner with over 120 clients, from the mid-market to DAX-listed groups. What follows is the practice from those projects, not the theory from a whitepaper.
What you will be able to do after this guide: look at a process in your organisation and judge, in 15 minutes, whether it is suited to automation, which technology fits, roughly what it costs, and how long it takes to pay off. If it does, you will also know what to look for in a partner.
Table of Contents
In this guide
The Decision Toolkit for Automation
A self-assessment suitability scorecard, a quick ROI calculator and a 12-question checklist for partner selection. As a PDF, 30+ pages, in the same design as our case booklet.
Thank you! Your toolkit is ready.
The download starts via the button below. If you have any questions about using it, do get in touch.
Open the toolkitPart A · 01
Fundamentals
Status Quo
Automation 2026: where we stand
In the mid-market, three maturity tiers of automation exist in parallel today; often within the same organisation, sometimes even within the same department. Each tier has a clear field of application and clear limits.
Tier 1 · since 2015
Classic RPA
Robotic Process Automation (RPA) was introduced in companies around 2015 and is today the most widespread form of process automation. An RPA bot operates existing software like a human: it clicks buttons, copies data from one system to the next, fills in forms and sends emails. What it does not do is make decisions; it executes what the rules prescribe.
60–80 %
of all business processes fully automatable
Tier 2 · since 2020
Intelligent Document Processing
Around 2020, the ability to “read” unstructured documents emerged: incoming invoices, contracts, emails, and scanned forms. Document Understanding combines classic image processing (OCR) with machine-learning models that recognise not only characters but also meaning.
85–90 %
Tier 3 · since 2024
Agentic AI
Since 2024, a new class of systems has been available: Agentic AI, also called AI agents. An agent is not a bot that works through a predefined sequence, but software that understands a task, makes a plan, deploys bots and tools, evaluates the result and self-corrects when errors occur.
90–95 %
Why most organisations are still stuck at Tier 1
In our conversations with decision-makers, we often hear the same three statements: “We tried RPA, but the bots keep breaking.” “We ran a pilot, but it never got beyond five processes.” “We have an internal team, but it can’t keep up with demand.”
In all three cases the cause is rarely the technology. It lies in three errors in the set-up:
They started with the wrong process, the business case was not calculated systematically, or maintenance and scaling were not thought through.
With Agentic AI you don’t get a better bot, but a technology that copes better with the errors above. That said: anyone who makes the same set-up errors with Agentic AI as with RPA simply fails more expensively.
Example
A worked example: the evolution of automation
The same process at four maturity levels. Each tier builds on the previous one: Agentic Automation does not replace RPA; it extends it. The purple fields show what is automated; the marked steps show where a human decides within the process.
RPA = Robotic Process Automation · IDP = Intelligent Document Processing · HITL = Human in the Loop
The fundamental split: agent decides, bot executes
One of the most common confusions in marketing material is the claim that Agentic AI is “the new RPA” or will “replace” RPA. That is factually wrong. The two technologies solve different problems and work together.
In concrete terms: in a real end-to-end process, say, processing an incoming invoice from a supplier whose PDF layout has just changed, the division of labour works like this:
- The agent (decides): it recognises that the layout does not match the standard pattern. It decides whether the new fields can be mapped unambiguously or whether a query is needed. It prioritises the invoice with an eye on early-payment-discount deadlines.
- The bot (executes): it opens the SAP screen, enters the data validated by the agent, clicks through the approval workflows and archives the invoice in document management.
This split is not just technical. It is also economically decisive: agents are considerably more expensive to run than bots (they use LLM calls, each of which costs money). If an agent makes an LLM call for every routine action, the business case quickly turns unattractive. The art is to deploy agents only where decisions and judgement are needed, and to leave execution to the cheaper bots.
What you already have today without knowing it
Before you think about an automation initiative, it is worth looking at the status quo. In almost every organisation above 500 employees, hidden automation already exists, usually grown from employees’ own initiative:
- Excel macros maintained by a single co-worker that collapse when that person leaves the company.
- Outlook rules that sort, forward or auto-reply to emails.
- Power Automate flows or Zapier connections that individual departments have built themselves.
- Scripts in PowerShell or VBA that run overnight and have no documentation.
These islands matter in two ways. First, they are a strong indicator that there is automation pressure in the organisation; employees have already helped themselves. Second, they are a risk, because they are disorganised, undocumented and not compliant.
Part B · 02
Assessment Methodology
Methodology
The 5 suitability criteria for a process
Whether a process is suited to automation can be answered against five criteria. We use these questions in our discovery workshops with new clients because, in practice, they prevent most poor decisions.
1. Volume: Is the effort worth it?
Automation costs development and operating effort and only pays off above a certain transaction frequency. Rules of thumb for the most important process types:
These figures are not absolute. A low-volume process with a high cost of error (e.g. compliance-relevant access reviews) pays off even at a much lower frequency, as our case at an audit firm shows.
2. Repetitiveness: how much is rule, how much is judgement?
A process can be repetitive in terms of frequency yet require an individual judgement each time. A tax adviser signs 200 annual financial statements a year. That is frequent, but each one is different. Here an RPA bot can drastically reduce the manual effort before and after signing (data extraction, document generation, dispatch, filing), but the professional judgement stays with the human.
The question: what share of the process follows clear rules, and what share is genuine judgement? If 70 per cent is rule-based, automating that 70 per cent pays off. The 30 per cent of judgement stays with the human, who finally has enough time for it.
3. Data quality: what comes in?
This is the most common reason RPA projects fail. Classic RPA only works with structured data: Excel tables, database rows, clearly defined form fields. As soon as incoming data is unstructured; a PDF with free text, a scanned letter, an email enquiry; you need Document Understanding, i.e. Tier 2.
4. Exception rate: how variable is the process?
If a process produces more than 20 per cent exceptions – special cases that don’t fit the standard pattern – classic RPA becomes unattractive. The bot keeps breaking, staff have to step in, and the benefit drops rapidly.
This is exactly where Agentic AI is strong: an agent can assess exceptions in context and make decisions instead of breaking. In our actuator case, 800 different supplier formats are the norm, not the exception. Agentic AI processes them all with a 0 per cent error rate.
5. System stability: will this still exist in 12 months?
A simple but often-forgotten criterion: will the system you are automating be replaced or fundamentally changed in the next 12 to 24 months? If so: wait. An RPA bot is tightly coupled to the user interface of the target system. If the screen changes, the bot breaks. Building an RPA bot for SAP a few months before an SAP migration is usually a losing move.
Exception: if the automation supports the migration itself (e.g. data migration between old and new), even a short-lived bot can pay off. But that is a special case, not the standard case.
Yes → continue to the next check · No → process optimisation instead of automation (or a strategic decision).
Economics
The ROI calculation in 4 variables
Once a process meets the five suitability criteria, it comes down to economics. We calculate the business case in four variables and recommend that our clients do the same before every project starts. If you can’t write the calculation down in 15 minutes, you shouldn’t start the process.
Variable 1: manual effort today
Handling time per item × frequency × internal hourly rate. Example: 5 minutes per incoming invoice × 10,000 invoices per year × €50 internal hourly rate = €41,667 of manual effort per year.
Common mistake: the handling time per item is set too low, because co-workers won’t admit their own pace. In workshops, we often measure it for real and find, in 80 per cent of cases, that the actual effort is 30 to 50 per cent higher than the self-assessment.
Variable 2: expected automation rate
What percentage of items will the bot or agent handle completely without human intervention? Realistic figures, depending on the technology mix:
Variable 3: implementation costs
Here are the orders of magnitude we see in practice, heavily simplified and with no guarantee for your specific case:
Not included in this calculation are the upfront costs for process documentation, IT connectivity and compliance reviews. Those usually come on top.
Variable 4: running costs
Often overlooked but highly relevant: an automation has to be operated. Licences, infrastructure, maintenance when systems change, monitoring. Typical orders of magnitude per process per year:
Calculating break-even: a real case
An example from our order-confirmation case at an actuator manufacturer.
Honestly calculated
The hidden costs and the hidden benefits
If you want to calculate a business case honestly, two effects belong in it that appear in no whitepaper template.
Hidden costs that many underestimate
- Process documentation upfront: before automating, the as-is process must be cleanly documented. That costs 1 to 4 weeks of effort on the client side and often takes longer than expected, because the process is more varied in people’s heads than assumed.
- Change management: employees whose work is being automated need clarity about their future. With automotive clients, we have learned that a works-council agreement at the start is quicker than an escalated conflict at the end.
- IT connectivity: API access, service accounts, permissions: This sounds trivial, but in large organisations it often takes 4 to 8 weeks.
- Compliance reviews: data protection, audit requirements, permissions. Particularly relevant for banks, insurers, healthcare and the public sector.
Hidden benefits that often don’t make it into the calculation
- Error reduction: the direct costs of manual errors, corrections, rework, lost early-payment discounts, and customer complaints are rarely measured. Yet they are often higher than the time effort itself.
- 24/7 availability: bots work at night, at weekends and on public holidays. In one Bavarian city, 35 per cent of resident parking permits are now processed outside office hours. Citizens get their permit faster without the authority adding staff.
- Scalability without headcount growth: as the business grows, process volumes grow with it. At one fibre-optic provider, the automation team today handles 190,000 transactions per year with the same headcount as before scaling.
- Employee satisfaction: people freed from repetitive tasks can focus on skilled work. We measure employee satisfaction as a KPI in almost every project; the scores typically rise by 15 to 25 per cent.
Strategy
Top-down or bottom-up: which approach fits you?
In practice, we see two fundamentally different approaches companies use to start automation. Both work, but for different situations.
Bottom-Up
Employee-driven
In their daily work, employees spot tasks that can be automated and submit them. A central team (or a partner) assesses, prioritises and automates.
Typical characteristics
- Many small-to-medium automations with low complexity.
- Quick early wins, high acceptance among the workforce.
- Relatively low benefit per automation (€10,000 – €50,000 per year).
- Scales well to 40+ processes, but the administrative overhead grows complex.
When it makes sense: you don’t yet have an automation culture and want to build broad acceptance. Or: you have a heterogeneous middle layer with many small, painful processes.
Top-Down
Strategy-led
Senior management defines strategic goals, cycle-time reduction, cost-out, scaling, and identifies the end-to-end processes that most strongly influence those goals. These processes are deeply automated, often with Agentic AI.
Typical characteristics
- A few large processes with high complexity.
- Longer project duration (6 to 18 months).
- High benefit per automation (€200,000 – €1,000,000 per year).
- Requires strategic commitment and cross-functional coordination.
When it makes sense: you already have bottom-up experience and now want to pull strategic levers. Or: you have a clear pain point (e.g. “39-day turnaround in order management”) that only end-to-end automation can solve.
The hybrid approach: what we recommend
In practice, a combination works best: strategically identify one to three large processes and automate them deeply, while building a bottom-up channel in parallel that creates broad acceptance and quick wins. At a global automotive supplier, we did exactly this, bringing 18 use cases into production and saving €708,000 in the first year.
Part C · 03 Functional Areas
Where automation concretely pays off: Finance, HR, Procurement, Customer Service and IT Operations. Five areas, three candidates each, thresholds, compliance and a real case for depth.
Area overview
Where automation concretely pays off
We cover the five functional areas that exist in almost every organisation with more than 1,000 employees, where we run most of our projects. Each area follows the same structure: status quo, three concrete automation candidates with thresholds, typical results, disqualifiers, compliance aspects and a real case for depth. See also the selected case studies from the Lunatec practice.
Area 01
Finance & Accounting
Finance is the area with the highest automation maturity in the mid-market. Incoming invoices, accounts-payable upkeep, travel expenses, and audit documents: The processes are rule-bound, high-frequency and transactional. Even so, a substantial share of the work in most accounting departments is done manually. The reason: data arrives unstructured, every supplier uses its own invoice format, and SAP screens are keying-intensive.
Three concrete automation candidates
Typical results in Finance
- Handling time per document: −80 to −95 per cent (e.g. from 5 minutes to 30 seconds)
- Error rate: typically 0 to 1 per cent (before: 3 to 8 with manual processing)
- Early-payment-discount capture: markedly improved through faster approval workflows
- Payback: typically 4 to 9 months at medium volume
When it doesn’t pay off
- You have fewer than around 200 incoming invoices per month; the economics tip over.
- An ERP migration is due in the next 12 months; the entire setup will change.
- Invoices arrive exclusively via EDI; then you need an integration solution, not RPA.
Area 02
Human Resources
HR is under pressure in 2026 as rarely before: the skills shortage makes every hire harder, sickness rates are rising, and regulatory requirements (the AI Act, whistleblower protection, new minimum-wage laws) create additional documentation work. At the same time, many HR processes in most companies are still highly manual, with Outlook, Excel and PDF attachments as the main tools.
Three concrete automation candidates
Typical results in HR
- Process run-time: −70 to −89 per cent (example: sick-note processing at an airline)
- Data quality: noticeably improved, because there are no more transcription errors
- First-time-right rate of the AI in Document Understanding: 90 to 95 per cent
- Staff relief: measurable in full-time-equivalents freed for higher-value work
When it doesn’t pay off
- You have fewer than 200 employees and correspondingly low HR volumes.
- The HRIS is outdated and will be replaced in the coming months: wait.
- You want to automate personnel decisions: that is dangerous and in part unlawful.
Area 03
Procurement & Supply Chain
In procurement, unstructured incoming data (supplier invoices, order confirmations, delivery notes) meets structured target systems (SAP MM, Oracle Procurement, ERP systems). This is the ideal use case for Document Understanding and Agentic AI, and it is usually the area with the highest euro leverage per automated process.
Three concrete automation candidates
Typical results in Procurement
- Processing time per document: −85 to −95 per cent
- Error rate: 0 to 1 per cent (before: 5 to 10 per cent with manual entry)
- Staff relief: buyers can focus on supplier relationships instead of data entry
- Common secondary effect: better early-payment-discount capture and more accurate demand forecasts
When it doesn’t pay off
- You already use EDI with all relevant suppliers: then an integration solution is the better route.
- The supplier base is very small (e.g. under 50 active suppliers): the volume isn’t enough.
- The ERP will be replaced in the next 12 months.
Area 04
Customer Service & Operations
Customer service is where the value gap between classic RPA and Agentic AI shows most clearly. Pure rule-based bots fail with natural language, unclear requests, and special cases. Agentic AI, by contrast, can understand and classify requests, gather information from multiple systems and generate context-aware answers, or support a human with prepared data in seconds rather than minutes.
Three concrete automation candidates
Typical results in Customer Service
- Turnaround time: −60 to −85 per cent
- First-response time: often from hours to minutes
- Customer satisfaction (CSAT): typically +15 to +25 per cent
- Employee satisfaction: often +20 per cent, because repetitive tasks fall away
When it doesn’t pay off
- There are fewer than 500 tickets per month, or a very heterogeneous request landscape.
- Requests are predominantly complex and require human judgement (e.g. legal advice).
- The CRM is outdated and is being replaced: wait.
Area 05
IT Operations & Service
IT is usually the first department to come into contact with automation, not least because it builds the infrastructure for it. Even so, internal IT processes are often the least automated: user provisioning by email, access reviews in Excel, manual tests at every release. This is exactly where the fastest wins lie.
Three concrete automation candidates
Typical results in IT Operations
- Standard-change turnaround: often from days to minutes
- Audit readiness: complete traceability at any time
- Test coverage: often a doubling or tripling of the paths tested
- Release cycles: markedly accelerated through automated regression
When it doesn’t pay off
- There are fewer than 100 IT service tickets per month: the volume isn’t enough.
- You are in the middle of an ITSM tool migration: wait.
- The IT landscape is heterogeneous, with many legacy systems due to be replaced over the next few years, in which case it is only worth investing in transitional solutions.
Part D · 04 Implementation
From idea to production: six phases, three make-or-buy options and the twelve questions that filter for a partner who delivers rather than just selling slides.
Implementation
The path from idea to production
Once you decide a process should be automated, implementation usually runs in six phases. We have used this framework for years — not because it is especially innovative, but because it works and prevents the most common pitfalls.
Realistic overall timeframes: a simple pilot takes 6 to 8 weeks to the first production bot. A medium process takes 3 to 4 months. A complex end-to-end process with Agentic AI takes 6 to 9 months. Anyone who promises you a complex process live in 4 weeks is ignoring either reality or phases 1, 2 and 4.
Make or Buy
Build, buy or partner
One of the most important strategic decisions made early in automation: do you build an internal team, engage an external partner, or a hybrid of the two?
Build an internal team
Makes sense if you expect at least 5 to 10 parallel initiatives, want long-term ownership of the automation architecture, and are willing to invest in recruiting and training. Realistic build-up time: 9 to 18 months until the team delivers independently. Risks: learning curves, staff turnover, lack of experience with unusual use cases.
External partner
Makes sense if you need results quickly, can’t justify building a specialised team for 3 to 5 processes, and want to draw on experience from many comparable projects. Risks: dependency, know-how staying external, switching costs if you change provider.
Hybrid approach
Most often successful in practice: an external partner builds the first 10 to 20 processes while, in parallel, an internal Centre of Excellence is established that takes over operations after 18 to 24 months and develops new initiatives itself. The partner remains for more complex use cases (Agentic AI, new domains) and for peak load.
At M-net we took the hybrid route: building an internal RPA Competence Centre over 4 years, scaling from 5 to 75+ processes, with Lunatec as the external partner for methodology, training and complex use cases. The internal developers were trained in a structured 4-phase programme.
Partner check
The 12 questions to ask an automation partner
If you decide on an external partner, here are the twelve questions you should ask. In our experience they filter out the providers who only sell slides from those who also deliver.
If a provider dodges any of these questions or stays vague, that is a signal. The most honest answer to the first two questions is usually a number, a name and a service level, not a slogan.
Frequently asked questions
Automation FAQ: cost, timelines and the AI Act
Do I need UiPath to start?
No. There are several strong platforms: UiPath, Microsoft Power Platform, Automation Anywhere, Blue Prism. Which is right depends on your existing IT landscape (Microsoft shops often benefit from Power Platform), the use case (Document Understanding is particularly strong on UiPath), and your scaling ambitions. We are a UiPath Diamond Partner, but we also build with Microsoft as needed: The choice follows the requirements; it is not a precondition.
What happens to the employees whose work is automated?
Across more than 120 Lunatec projects, we have not seen a single case in which employees were let go directly because of automation. What happens instead: employees move to higher-value work, buyers tend to supplier relationships rather than data entry, accountants do analysis rather than routine reconciliations, HR staff advise managers rather than sorting documents. The bottleneck in the mid-market today is almost never surplus headcount, but the skills shortage. Automation addresses exactly that bottleneck.
How long does a first pilot take?
A well-scoped first pilot typically runs 6 to 8 weeks from kick-off to live operation. That assumes the process is clearly defined, IT access is available promptly, and a clear business owner is available as the point of contact. If these three conditions are missing, it takes longer, regardless of who the implementation partner is.
Roughly what does it cost?
For a simple first RPA process, budget €15,000 to €30,000 for implementation plus €3,000 to €8,000 per year to operate. For more complex use cases with Document Understanding, €40,000 to €80,000 plus €8,000 to €20,000 per year. Highly complex Agentic AI projects come in at €100,000 to €250,000 plus €15,000 to €40,000 per year. These figures are orders of magnitude, not fixed prices; the actual amount depends heavily on process complexity, interfaces and data quality.
Can we start with a pilot, without a big upfront commitment?
Yes, in fact we recommend it. A well-defined first use case with clear success metrics is the best way to test the collaboration before you commit strategically. At Lunatec, we offer a 60-minute analysis workshop in which we jointly identify 3 to 5 processes with the highest automation leverage. No obligation.
What about data protection and the GDPR?
GDPR compliance is a standard topic on every project, not a special case. In concrete terms: data is hosted in the EU, access rights are documented and audit-proof, and personal data is minimised. For especially sensitive use cases (e.g. healthcare, HR) we carry out a separate data protection impact assessment. We typically run our solutions in the UiPath Automation Cloud (EU region) or on-premise in the client’s data centre.
What about the AI Act?
The EU AI Act enters into force in stages, with high-risk applications from August 2026. As of July 2026 (the document’s date) it is already partly in effect. It distinguishes between high, medium and low risk levels of AI applications. For most business automations (incoming invoices, document classification, data transfer) the requirements are manageable. Applications that make or prepare personnel decisions are high-risk and require corresponding documentation and human oversight. We build compliance into the system design as standard.
The Decision Toolkit for Automation
A self-assessment suitability scorecard, a quick ROI calculator and a 12-question checklist for partner selection. As a PDF, 30+ pages, in the same design as our case booklet.
Thank you! Your toolkit is ready.
The download starts via the button below. If you have any questions about using it, do get in touch.
Open the toolkit
Lunatec · Agentic Automation · 2026
Which processes in your organisation have the greatest potential?
In a 60-minute analysis call, we jointly identify three to five processes with the highest automation leverage, with no obligation, based on your specific situation. You get an assessment of the effort, the benefit and the payback period — with no sales pressure.