AI FOR SMALL BUSINESSES IN GERMANY

A practical event for a market that is ready to move

Artificial intelligence has moved from curiosity to operating priority for German companies. At the Munich event, the discussion was noticeably more mature than the familiar “What is ChatGPT?” debate. The presentations and start-up pitches focused on implementation: where to begin, when a standard tool is enough, when a custom integration is justified, how to protect company data and how to keep a promising pilot from dying in the proof-of-concept stage.

That shift is visible in the wider market. KfW Research reports that 20% of German SMEs used AI in the 2022–2024 period—almost 780,000 companies—up from only 4% in 2016–2018. Even among firms with fewer than five employees, the share reached 19%. The momentum accelerated further in 2026: an ifo Institute survey in May 2026 reported that 54.5% of German companies were using AI in business processes, compared with 40.9% a year earlier. A further 16% planned to start and 21.6% were discussing potential applications. The figures use different samples and definitions, but both point in the same direction: AI adoption is accelerating, while the hard work is shifting from experimentation to integration.

Source context: KfW Research on AI use in German SMEs and ifo Institute on AI adoption in German companies.

Seven highlights from the presentations

1. Problem orientation beats tool orientation

A recurring warning was that companies often begin with a product decision—“We are introducing Copilot” or “We want an OpenAI solution”—before defining the process they want to improve. A stronger starting point is to locate high manual effort: inboxes that must be sorted, reports that are repeatedly rewritten, knowledge that is difficult to find, or data that is copied between systems. The business problem should determine the architecture, not the other way around.

2. Standard tools and custom AI are complementary

Standard tools are attractive because they are immediately available, relatively inexpensive and suitable for generic work such as drafting emails, summarising meetings, editing documents or preparing presentations. Custom AI becomes compelling when the process is a core capability, occurs frequently, depends on proprietary data or requires deep integration with ERP, CRM, laboratory, quality or document-management systems. Most companies will ultimately use both.

3. The 80/20 mindset is a major cultural change

Traditional software projects are designed to behave deterministically. Generative AI is probabilistic. Several talks highlighted the value of accepting that AI may automate 80% of a task while a specialist reviews the final 20%. That can still create an excellent business case. Requiring perfect autonomy too early often prevents useful automation from entering production.

4. Data quality matters—but perfection is not a prerequisite

Company information is commonly spread across drives, mailboxes, databases and outdated templates. Clean, governed data improves results, yet the advice was not to postpone all experimentation until a multi-year data programme is complete. A narrow knowledge assistant or document workflow can expose where the data problems really are and create a concrete roadmap for improvement.

5. Proof of value is more useful than proof of concept

A technically impressive prototype is not enough. The event repeatedly returned to business measures: minutes saved per transaction, fewer manual entries, shorter response times, higher throughput, improved quality or reduced energy consumption. A proof of value should test the smallest scope that can demonstrate an operational and financial benefit.

6. Adoption is a training and change-management challenge

Simply purchasing licences and announcing that everyone may use them rarely changes behaviour. One practical learning was that users need role-specific examples, recurring office hours, short demonstrations and a safe place to ask questions. Each team requires different prompts, agents and workflows. AI literacy is also a compliance issue under Article 4 of the EU AI Act.

7. Keep the model replaceable

The leading model can change rapidly. A durable enterprise architecture separates the workflow, permissions, data connections and evaluation framework from the underlying model. This reduces dependency on a single vendor and allows companies to switch between providers as capability, cost, latency or sovereignty requirements evolve.

The most interesting business AI use cases discussed

The strongest examples were not generic content-generation demos. They showed AI acting as a structured layer between unorganised inputs and existing operational systems.

Use case

What the AI actually does

Multichannel order intake

A brewery received orders by email, voicemail and even fax. An agent read or transcribed the incoming request, extracted the order data and prepared the ERP entry. The employee’s role changed from typing to checking and releasing.

Private knowledge assistant

Employees ask questions in a familiar chat interface, but answers come from approved company documents. Each answer includes a source reference, and the system declines to answer when the evidence is missing.

Meeting-to-CRM workflow

With the customer’s consent, a meeting is recorded and transcribed. AI produces a structured protocol, identifies actions and buying signals, stores the note in the CRM, creates a deal activity and proposes the follow-up date.

Laboratory report generation

Historical laboratory reports and reporting rules are used to draft new reports from measured values. A laboratory specialist remains responsible for quality control and final approval.

Technical and expert reports

Engineers and consultants can focus on inspections, measurements and customer conversations while AI drafts the repetitive narrative portion of the report using approved examples and templates.

Inbox classification and routing

Incoming requests are classified, key details are extracted and the next processing step is prepared. This is a practical entry point because the process is frequent, measurable and often painfully manual.

AI telephone assistant

A voice agent can answer recurring calls, collect structured information, route urgent cases and provide service outside normal office hours, while handing uncertain or sensitive conversations to a person.

Energy optimisation

AI can combine tariffs, consumption forecasts, batteries and operational constraints to decide when electricity should be purchased, stored or used—turning cost management into a continuously optimised process.

Quality and environmental management

AI can organise evidence, support audits, compare documentation against requirements and identify missing or inconsistent records, provided that accountability remains clearly assigned.

Website and search visibility agents

An agent can monitor content performance, identify gaps, prepare updates and adapt pages for traditional search and emerging AI-based discovery. Human editorial control remains essential for accuracy and brand voice.

A decision framework for SMEs: buy, configure or build?

Approach

Best when

Typical examples

Use a standard tool

The task is generic; the process is not strategically differentiating; speed and low entry cost matter most.

Email drafting, meeting summaries, presentation outlines, document editing, basic research.

Configure an existing platform

The company needs its own knowledge, permissions, templates or simple integrations, but not a completely new product.

Internal knowledge assistant, document workflows, CRM copilots, role-specific agents.

Build a custom integration

The workflow is a frequent core process with proprietary logic, high volume, significant risk or a strong competitive advantage.

Order-to-ERP automation, lab reporting, expert-report generation, operational optimisation.

A 90-day implementation roadmap

Weeks 1–2: Select one painful process

Choose a frequent task with a clear owner. Document the inputs, outputs, exceptions, current time requirement and error rate.

Weeks 3–4: Establish the baseline and risk level

Measure the current cost and quality. Identify personal data, confidential information, legal consequences and decisions that require human approval.

Weeks 5–7: Run a narrow proof of value

Use representative data and real users. Test whether the proposed workflow saves time and maintains acceptable quality—not whether the demo looks impressive.

Weeks 8–10: Integrate and govern

Connect only the systems needed for the pilot. Define permissions, logging, model choice, retention, escalation and the human-in-the-loop checkpoint.

Weeks 11–12: Train, evaluate and decide

Provide role-specific training, compare results with the baseline and decide whether to stop, redesign or scale. Record the lessons so the next use case becomes faster.

Governance is part of the business case

Data sovereignty was a central concern at the event. Modern enterprise implementations can keep data in Germany or the EU, operate in private or on-premises environments, prevent customer data from being used for model training, inherit existing system permissions and record every automated action. Those options do not remove the need for governance; they make it possible to design governance into the workflow.

For companies operating in the EU, AI literacy is already a concrete responsibility. IHK Munich’s AI Act guidance notes that Article 4 has required providers and deployers to ensure an appropriate level of AI competence among staff since 2 February 2025. A practical programme should include an inventory of AI tools, defined responsibilities, role-specific training, documentation of participation, privacy rules and clear escalation paths.

Read the current guidance: IHK Munich — AI Act rules for companies.

Conclusion: start small, but design for production

The Munich event made a convincing case that small companies do not need a grand AI transformation programme before they can create value. They need a clear problem, a responsible owner, a measurable baseline and the discipline to test a narrow workflow with real users. The most promising opportunities are often hidden in unglamorous work: copying data, writing repetitive reports, searching for documents, documenting meetings and sorting incoming requests.

The competitive advantage will not come from choosing one “winning” model. It will come from learning how to redesign work around AI while keeping expertise, accountability and trust in the hands of people. Companies that can repeatedly identify a pain point, prove value and integrate the solution safely will build a capability that outlasts any individual tool.

Authoritative links and further reading

EDF Non-Thematic SME Calls 2026: The Overlooked Entry Point into Europe’s €1 Billion Defence Fund

Leave A Comment