OMR Reviews recognises diva as a top-rated ERP system
With no jury, based solely on user ratings: diva receives the
According to a survey, 63 per cent of companies that do not use AI cite costs that are difficult to estimate as a barrier, Bitkom Association study (September 2026, 603 companies with 20 or more employees). For CFOs and IT managers in SMEs, this means that whilst there are usually plenty of ideas, things become difficult when it comes to budgeting for AI as a whole. AI only becomes quantifiable for SMEs when you calculate the cost of a single project using specific units.
The ROI of an AI project in a small or medium-sized enterprise is calculated by comparing the net benefit to the total cost of the project over a fixed period: benefit minus cost, divided by cost. Benefits include cost savings per transaction, shorter processing times, higher availability and fewer errors. Total costs include cloud consumption and licences, integration, data, change management and operations. The ROI can only be calculated on a per-use-case basis, not for artificial intelligence as a whole. This article illustrates the five cost categories, the four benefit categories and the calculation method leading to payback, using a real-world project as an example.
Use of AI: Between 2022 and 2024, 20 per cent of small and medium-sized enterprises in Germany used AI; according to KfW Research, this is around five times the proportion recorded between 2016 and 2018.
Project accounts: The ROI of AI in small and medium-sized enterprises can only be calculated on a case-by-case basis, in terms of processes, minutes and euros per process.
Cost categories: According to Bitkom, AI users cite infrastructure (51 per cent), data preparation (50 per cent) and system integration (41 per cent) as major cost drivers, whilst licences account for just 21 per cent.
Practical example: At Carl and Carla, around 90 per cent of returns processes are automated; according to the case study, the cost per call is less than one euro.
Prioritisation: The ROI is determined before the project begins; the AI Discovery Workshop produces a shortlist, prioritised according to business value and effort, in around three days.
20 per cent of small and medium-sized enterprises in Germany used artificial intelligence between 2022 and 2024 – just under 780,000 businesses. This is shown by KfW Research in the KfW SME Panel (Fokus Volkswirtschaft No. 533, February 2026), a study on German SMEs with just over 6,800 responses. Between 2016 and 2018, the proportion stood at 4 per cent; since then, the use of AI has increased approximately fivefold. This means that four out of five SMEs do not yet use AI.
The proportion increases with company size: 36 per cent among medium-sized firms with 50 or more employees, and 19 per cent among micro-enterprises with fewer than five employees.
Easily accessible applications dominate the field of AI technologies: 14 per cent of all small and medium-sized enterprises use natural language generation, 10 per cent use text recognition, and 3 per cent use data analytics, according to KfW Research in ‘Fokus Volkswirtschaft’ No. 554 from July 2026.
According to a study by the Bitkom trade association, the main obstacle to the adoption of AI is not so much a lack of ideas as a lack of knowledge and data. 85 per cent of companies not using AI cite a lack of technical know-how as a barrier, 66 per cent cite legal barriers, 63 per cent cite costs that are difficult to calculate, and 51 per cent cite a lack of human resources.
For users, too, data protection requirements (66 per cent) and costs that are difficult to estimate (54 per cent) remain the biggest challenges. If you cannot estimate the costs, you cannot calculate the ROI.
This article explains whether your company meets the requirements for using AI how companies can make good use of AI and where they would be better off waiting a little longer. It also covers the current status of the digitalisation of your processes.
Studies on the potential of the technology for the economy as a whole say nothing about an individual AI project in a small or medium-sized enterprise. According to KfW Research (Focus on the Economy No. 554), which draws on the 2025 Joint Economic Forecast, the projections range from 0.05 to 2.8 percentage points of additional productivity growth per year over the next ten to 15 years, depending on the country. This range is not suitable for a budgetary decision. Only a calculation using figures from one’s own company can be relied upon.
The ROI formula for an AI project is: Benefit minus costs, divided by costs, based on a fixed period. The payback period answers a different question: after how many months do the cumulative benefits exceed the project costs?.
These figures are available in most SMEs, provided that their processes have been digitised to a sufficient extent: in their telephone system, ERP system or ticketing system.
Transactions per day: calls, enquiries, documents, searches
Minutes per task before automation
Cost per task, based on minutes and an hourly rate
Error or enquiry rate per transaction
Anyone familiar with these four units for a particular use case can convert any benefit block into euros. Without them, any AI implementation lacks a target value.
Licences and token usage account for the smallest part of the bill. According to Bitkom association study from September 2026 AI users most frequently cite infrastructure such as the cloud and computing power (51 per cent) as a major cost factor. This is followed by data preparation (50 per cent) and integration into existing systems (41 per cent); 21 per cent cite subscriptions and licences, whilst only 8 per cent mention token consumption.
Guides on introducing AI in small and medium-sized enterprises usually only cover the first stage: the technology itself. There are five factors that contribute to the total cost of an AI implementation.
Cost section | What’s included | Cost drivers | Measurement parameter |
|---|---|---|---|
1. Cloud usage and licences | Azure OpenAI is categorised into three main categories in Microsoft Foundry, according to Microsoft Learn (As at August 2026): Standard by token, provisioned with a monthly or annual reservation, batch of 50 with a % discount within 24 hours (Azure pricing page, accessed October 2026); on collana in the customer’s Azure environment; Microsoft processes the data as a data processor; OpenAI has no access (Generative AI in Azure) | Transaction volume, load profile | euros per 1,000 transactions |
2. Integration with ERP, CRM and telephony systems | APIs for ERP systems such as Microsoft Dynamics 365 Business Central; telephony via speech-to-text and text-to-speech; integration with DMS and CRM | Number of systems, real-time requirements | Person-days: Integration |
3. Data preparation and quality | Consolidate documents, master data and spare parts data | Scattered sources, paper collections | Person-days; proportion of usable documents |
4. Change, training, process adaptation | Involving staff, adapting processes; according to Bitkom, 42 % of AI users cite a lack of acceptance as a barrier | Roles and locations affected | Training days; utilisation rate |
5. Operations, Monitoring, Compliance | Governance, monitoring, cost control; data protection measures (below) | Data protection requirements | Hours per month; audit workload |
Data protection also entails predictable costs, which can be quantified in concrete terms. Four measures should be included in the project plan:
EU Data Zone: In the case of a Data Zone deployment, processing remains, according to Microsoft Learn in the EU Member States, including EFTA states where applicable. Prompts and responses are, according to Microsoft’s privacy notices regarding Azure OpenAI (As at May 2026) they are not made available to other customers or to OpenAI; they do not train foundation models without the customer’s permission or instruction.
Offline models for particularly sensitive data: Local LLMs run on-premises or in the private cloud; the cost involves maintaining your own GPU infrastructure rather than pay-as-you-go billing (Offline AI and Local LLM).
Pseudonymisation or anonymisation prior to the prompt: Pseudonymised data is, according to Article 4(5) of the GDPR which can no longer be linked to any individual without additional information stored separately. Anonymisation goes further: an AI solution from collana automatically redacts personal information in documents.
Data Processing Agreement in accordance with Article 28 of the GDPR including, amongst other things, the subject matter, duration, nature and purpose of the processing.
The second compliance issue is the AI Regulation, which, according to Article 113 of the AI Regulation has been in force since 2 August 2026. According to the Amending Regulation (EU) 2026/1744 Under Article 113, as amended on 24 July 2026, the obligations for high-risk AI set out in Annex III shall take effect on 2 December 2027 and those set out in Annex I on 2 August 2028. The transparency obligations under Article 50 have applied since 2 August 2026; providers of generative AI systems that were already on the market prior to this date have until 2 December 2026 to label AI-generated content (paragraph 2).
Whether you licence AI tools, integrate them into existing systems or develop them yourself, it merely shifts the balance between the different elements. This article compares the three options The right AI strategy: buy, integrate or develop in-house.
Four areas of benefit can be measured when it comes to AI in small and medium-sized enterprises: cost per transaction, processing time, availability and quality. The greatest potential for improvement often lies in Search time, downtime and loss of knowledge, i.e. items that do not appear in any cost centre. The practical examples are taken from collana case studies in the areas of customer service and technical support.
Benefits section | Measurement parameter | Data source within the organisation | Example from collana cases |
|---|---|---|---|
Cost per transaction | euros per call, enquiry or document | Telephone system, ticketing system, cost accounting | Less than €1 per call with Carl and Carla |
Processing time | Minutes per operation, diagnosis time | Call transcripts, service reports | Call duration halved; document searches take seconds rather than minutes at Hoval |
Availability | Hours per week, languages | Service hours, call statistics | 24/7 and multilingual at Carl and Carla’s; three languages at Bucher Municipal |
Quality and consistency | Enquiry rate, error rate | Complaints, follow-up enquiries | Please cite your sources for every answer on Hoval |
At Bucher Municipal, much of the technical documentation – particularly for older vehicle models – was only available in decentralised locations or in paper form. This resulted in time-consuming searches, long downtimes and inefficient processes. The chatbot on Azure OpenAI now answers questions in German, Spanish and Chinese for service technicians worldwide.
The report cites the following results: Bucher Municipal Case Study (As at October 2026) Faster maintenance call-outs, shorter diagnosis and repair times, and lower operating costs – no figures given. The app was developed by Option 4.0, a member of the collana Group, in collaboration with Bucher Municipal.
“Our collaboration with Option 4.0 is going very smoothly. Our point of contact is far more than just a technical service provider – he acts as an integral part of our in-house IT team. With flexibility and speed, he has steered the project to success and impressively demonstrated to us the possibilities offered by AI.” Christian Johansson, CIO at Bucher Municipal.
Hoval, a manufacturer of heating and indoor climate control systems, is bringing together in the Hoval Case Study (As of September 2026) scattered, multilingual product documentation and spare parts data in the Azure-based assistant „Hoval Bro“. Each answer includes a source reference. According to Patrick Renz, Global Product Manager at Hoval Services, searching for documents now takes seconds rather than minutes.
„What used to take minutes — finding the right document for the right product variant in the right language on the right page — now takes seconds. And all it takes is a single chat message.“ Patrick Renz, Global Product Manager at Hoval Services.
Both cases represent Generative AI in knowledge management: It makes scattered information findable. The metric for this benefit block is: search minutes per call-out multiplied by call-outs per year multiplied by the number of service engineers. Neither case publicly quantifies these figures; please use your own values from the service logs. Hoval demonstrates that the benefit stands or falls with the database – in other words, with cost block 3.
Using four metrics from a real-world project, the business case for an AI agent in customer service can be calculated in just a few lines: call volume, degree of automation, call duration and cost per call. These figures are taken from the Case Study: Carl and Carla (As at October 2026).
Background: Carl und Carla operates a van hire business in Germany, with over 1,000 vehicles in more than 40 towns and cities. Particularly at weekends and outside business hours, the company received up to 200 calls a day regarding vehicle returns.
The solution is the AI agent ILAI, developed by collana (Impact Labs) in collaboration with Carl and Carla, as a cloud-based voice agent on Microsoft Azure. The system can be run either in the cloud or entirely on-premises, with integration into Business Central via APIs.
Within six weeks, the following took place: around 90 per cent of all returns processes Fully automated. The average call duration has been halved, and the cost per call is less than one euro. This corresponds to the operating costs per automated transaction. The service has since been available round the clock and in multiple languages.
Find out more about the AI agent at the Product page ILAI.
„Thanks to the AI-based call centre agent, our customers can now return their vehicles in just a few minutes and without any waiting time. This is a huge relief for our team, as routine enquiries are processed automatically and reliably. We were particularly impressed by the seamless integration into our processes: Around 90% of returns are processed completely automatically - this saves time and ensures satisfied customers.“ Gregor Wendt, CTO at Carl and Carla.
The number of transactions per day multiplied by the number of working days gives the annual volume. In this case, the peak was 200 calls a day; use your own average figure.
The annual volume multiplied by the degree of automation gives the number of automated processes. In this case: around 90 per cent of returns processes.
The costs per task to date (your minutes multiplied by your hourly rate) minus the AI operating costs per task (less than 1 euro in this case) give the saving per task.
The savings per transaction, multiplied by the number of automated transactions, gives the annual benefit.
The one-off project costs from the five cost categories, divided by the monthly benefit (annual benefit divided by twelve), gives the payback period in months.
The cost per call before the change remain your placeholder; collana does not give a figure for this. The six weeks leading up to full-scale operation shorten the period during which the system is not being used. The article describes how voicebots work in practice AI in customer service.
The ROI is determined before AI is implemented, during the selection of the use case. In its ‘Fokus Volkswirtschaft’ report No. 554, KfW Research identifies the identification of initial, practical AI applications as a key hurdle for small and medium-sized enterprises. According to a study by the Bitkom trade association, 30 per cent of companies see no use cases without AI. Those who prioritise use cases according to business value and effort start with the case that, the one that pays off the quickest.
Two dimensions are sufficient to assess the potential of a use case: business value, calculated as transaction volume multiplied by the savings per transaction, plus availability and quality; and effort, calculated based on the depth of integration, data availability and compliance requirements. This results in four areas:
This is how an AI strategy for SMEs is developed step by step: start with a measurable use case and apply the approach to other business processes.
The AI Discovery Workshop The collana process takes around three days – including preparation and follow-up – to move from the initial inspiration to a concrete roadmap. It prioritises use cases according to benefit and feasibility, and sets out the top use cases in detail, including effort, timelines and responsibilities. Three deliverables:
a shortlist of prioritised use cases, weighing business value against effort for the top use cases
A technical overview with recommendations on Microsoft and Azure technologies
a roadmap setting out the next steps for the most promising use case
The first step is a free, 30-minute initial consultation. If there is no data set available for measuring the benefits, the Data Intelligence Workshop the step before that.
An AI project is only worthwhile if the key performance indicators continue to be monitored after go-live. Three factors ensure that the project remains viable in its second year:
Track the four benefit categories as key performance indicators on an ongoing basis: present the monthly trends in costs and minutes per transaction alongside consumption figures.
Maintaining the database: At Hoval, every response includes a source reference; this highlights any out-of-date documents.
Managing consumption: sporadic use via tokens, continuous usage as a monthly or annual subscription, time-insensitive batch processing with a 50 per cent discount according to Azure pricing page; Monitoring and cost control form part of the governance of the Azure environment.
The extent to which AI in small and medium-sized enterprises is linked to business success and competitiveness is demonstrated by KfW Research in ‘Fokus Volkswirtschaft’ No. 558 from September 2026, though only as a point of reference. SME AI users in 2024 a 1.1 percentage point increase in revenue growth . Added to this were a 0.9 percentage point increase in return on turnover and a rise of around 3.4 per cent in labour productivity.
KfW Research regards this as a snapshot and leaves open the question of whether the use of AI is the cause or whether more successful companies are more likely to adopt AI. For your project, therefore, your own measurements are what count.
The ROI of an AI project is the net benefit divided by the project costs over a fixed period, i.e. benefit minus costs, divided by costs. It is measured in units per transaction: transactions per year, minutes and euros per transaction. An ROI of 100 per cent means that the net benefit is equal to the investment.
There is no fixed target figure. A practical benchmark is for the investment to pay for itself within the period for which the budget was approved. More important than the amount is that the calculation is based on the organisation’s own units.
The payback period refers to the month from which the cumulative benefit equals the investment. The ROI refers to the net benefit per euro invested over the period under consideration, i.e. the rate of return.
Proven use cases from collana: telephone automation in customer service (Carl and Carla) and maintenance and service knowledge for technicians (Bucher Municipal, Hoval). Another solution offered, for which no case study has been published, is the anonymisation of documents containing personal data.
At Carl and Carla, around 90 per cent of all returns processes were automated within six weeks – a one-off case with a clearly defined procedure. The roll-out of ILAI takes place in five stages, from the use-case workshop through to go-live; the duration depends on the depth of integration and the data situation within the company.
The calculation set out in this article only applies to a specific use case. If you have not yet prioritised your use cases according to business value and effort, that is the first step. During the free preliminary consultation for the AI Discovery Workshop, you will identify which processes within your organisation deliver the four units. Arrange a free initial consultation now
For an overview of the AI capabilities of collana, from ILAI through to generative AI in Azure and offline AI, please see AI solutions from collana.
Would you like to apply what you’ve read to your business? In a free initial consultation, we’ll work through your processes together and show you which step is worth taking first.
With no jury, based solely on user ratings: diva receives the
Two bronze awards at the „ERP System of the Year 2026“ awards: b4dynamics impresses with
Several retail companies have recently upgraded their stock management systems with MAC IT-Solutions to