AI Agents for SMEs: Run the Operational AI Audit Before You Automate
Florian Hödl
AI Expert
> Summary: SMEs have crossed the AI adoption threshold. According to the KI-Index Mittelstand 2026 by Salesforce and the German Mittelstands-Bund, 51.2% use AI, but only 16.6% run autonomous agents. The real value sits in the work between ERP, CRM, email, and PDF that is still done by hand. Dolibarr or another ERP stays the system of record, agents sit on top, and a human signs off the critical steps. The Operational AI Audit clarifies where your business stands across six areas before a single line of code is written.
Where Does the Mittelstand Really Stand on AI?
The 2026 numbers are clearer than many expect. According to the KI-Index Mittelstand 2026 by Salesforce and the German Mittelstands-Bund, 51.2% of SMEs use or test AI, up from 33.1% the year before. The share running autonomous agents nearly doubled from 8.7% to 16.6%.
At the same time, practice shows a wide gap between knowing and doing. Many companies recognise the relevance but have nothing in operational use. That is rarely a technology problem. It is a question of the first concrete step: which process, which data, which sign-offs.
Where Do AI Agents Actually Save SMEs Time?
The interesting part is not inside the ERP. An ERP posts transactions, holds master data, runs inventory, and settles invoices. What it cannot do is reason about an inbound email, read a PDF, or answer a question that spans three systems. In an SME, that work is done by people, by hand.
Bain & Company sized this area in 2026: automating the coordination work between systems is a market of around 100 billion US dollars in the US alone, roughly double with Europe, and more than 90% of it is still untapped. This is the daily glue work: pulling data from the ERP, checking it against a spreadsheet, interpreting a free-text vendor email, and deciding whether to escalate.
Typical candidates in an SME:
- Order capture from PDF: orders arrive by email, the data is typed into the ERP by hand. An agent reads the PDF, structures the line items, and creates a draft in the ERP.
- Invoice reconciliation: check incoming invoices against purchase order and delivery note, flag deviations.
- Quote drafts: turn an inquiry into a draft quote in the CRM that a human reviews and approves.
- Status requests: answer recurring questions about orders, dates, or stock from the systems.
The rule of thumb: high volume, clear rules, measurable effort. Those are the processes where an agent makes sense first.
Why the ERP Stays the System of Record
In almost every implemented project, the Mittelstand lands on the same architecture: the ERP stays the system of record and the audit trail, agents sit on top. This is not a stopgap. It is the only setup that fits the reality of German and Austrian businesses.
Three reasons drive it:
- GoBD and evidence duties. Tax-relevant records must be immutable and traceable. The ERP is the official record. An agent may prepare it, but never bypass it.
- EU AI Act and GDPR. Agents in procurement, HR, or production-adjacent processes can count as high risk. Decision logic, escalation paths, and documentation belong before the start, not after.
- Data sovereignty. With Dolibarr on infrastructure in Germany and agents working through the REST API, the data stays inside your perimeter. That is an advantage over purely US-based services.
The agent remains a tool with limits. A human confirms critical steps, and every action is logged. This is called human-in-the-loop, and for regulated processes it is the condition that lets agents reach production at all.
The Operational AI Audit: Six Areas
Before a line of code is written comes the assessment. The Operational AI Audit checks across six areas whether a business is ready to put agents on its systems.
| Area | Question | Output | |
|---|---|---|---|
| Process map | Which recurring work runs between the systems? | List of candidates with hours per week | |
| Data quality & systems | Where is the system of record, how clean is it? | Data picture and open issues | |
| Automation candidates | Which process fits first? | Priority by volume, rules, cost | |
| Governance & compliance | What may the agent do, when does a human take over? | Decision logic, escalation, EU AI Act rating | |
| Integration readiness | Are there stable APIs, webhooks, EU hosting? | Feasibility and integration effort | |
| ROI & pilot | What does a first pilot save, how is it measured? | Bounded pilot with clear metrics |
The output is not a concept for the drawer. It is a prioritised plan with a first pilot that can be measured.
Funding in Austria
Austrian businesses do not have to fund this step alone. The Chamber of Commerce and the Vienna Business Agency offer digitalisation grants for SMEs. A structured assessment like the Operational AI Audit is often the basis to support a grant application. EU AI Act compliance is also becoming a requirement in B2B and public-sector procurement. Building it in early is an advantage. Whether a grant applies depends on the individual case and is clarified with the respective body.
Conclusion
AI in the Mittelstand rarely fails on the technology. It fails on the first concrete step. The lever sits in the manual work between the systems, the ERP stays the system of record, and an agent takes over the recurring tasks with human sign-off. The Operational AI Audit makes that first step plannable. You can download the [Operational AI Audit checklist as a PDF](/downloads/operational-ai-audit-en.pdf) for free, 48 points to tick off across six areas. To see what a guided introduction looks like, visit our [AI consulting for SMEs](/en/services/ki-beratung/).
Frequently Asked Questions
Does an AI agent replace my ERP?
No. The ERP stays the system of record and the audit trail. The agent sits on top and takes over work between the systems, such as data from PDFs or emails. A human signs off the critical steps.
Does this work with Dolibarr?
Yes. Dolibarr provides a REST API and webhooks through which agents can read and write. Managed Dolibarr by Anexum combines hosting in Germany, maintenance, and this agent layer in one service.
How long does an Operational AI Audit take?
It depends on the number of processes and systems. Usually a few weeks are enough for the assessment, prioritisation, and a first pilot proposal with metrics.
Is this GDPR and EU AI Act compliant?
The architecture is built for it: operation on infrastructure in Germany, logging of every action, defined decision logic and escalation. The overall legal assessment stays with your legal and compliance advisers.
Where do I start?
With a single, well-measurable process that has high volume and clear rules. A bounded pilot proves the value before you automate more broadly.