AI processautomation for SMEs
A large part of office work consists of reading and typing: invoices, delivery notes, order confirmations, enquiries by email. We set up processes in which AI reads these documents, checks the details against your ERP or CRM and puts forward a draft. Your staff check and approve.
ERP · CRM · Document management · Mailbox · Run locally, in the EU or as a mix
What AI process automation is
Classic automation follows fixed rules and fails as soon as a document looks different from what was expected. AI process automation adds language models that can read and classify texts, documents and enquiries. Calculations, mandatory fields and posting rules stay in fixed, verifiable program code.
For most processes, an AI workflow with fixed steps is enough. It is cheaper and easier to check. An AI agent, which decides for itself which tool to use next, only pays off when there are many exceptions. We explain the difference on the AI agents page.
How an automated process is structured
Each step has its own task. This makes it possible to isolate errors and test changes one at a time.
Input
An email, a PDF, a form or a ticket triggers the process.
Read and classify
The AI recognises the type of document and extracts the relevant information.
Check against your systems
The information is checked against master data, orders and rules in the ERP or CRM.
Draft or proposal
The result is a draft quote, a booking proposal or a reply. Uncertain cases go straight to a person.
Approval and action
Anything with consequences, such as dispatch, a posting or an order, only runs after approval by a person. Every step is logged.
Typical processes
You will find detailed descriptions with metrics among the use cases.
Incoming invoices and documents
Capture them, match them against the order and goods receipt, suggest a cost centre, flag discrepancies.
Customer enquiries and quotes
Extract requirements, look up items and prices in the ERP, create a draft quote.
Service and complaints
Sort requests, retrieve the status, find similar cases, prepare a reply.
Contracts and tenders
Find clauses and deadlines, flag deviations from the standard, create a checklist.
HR and onboarding
Check documents for completeness, trigger tasks, provide information for each role.
Which systems we connect to
To all systems with a documented interface: ERP, CRM, document management, ticket systems and Microsoft 365. Before we build anything ourselves, we check what your ERP, Microsoft 365 or SAP can already do.
If there is no interface, software robotics (RPA) can close the gap. This is a program that replicates the inputs a person would otherwise click in the user interface.
Depending on the data, the language model runs in an EU cloud, on your own hardware in the AI Cube or as a mix. Models can be swapped later without rebuilding the process.
Control and security
The AI only gets the rights the process needs. Actions with consequences require approval, uncertain results go to a person, and every step is logged.
Documents and emails can contain hidden instructions. Experts call this prompt injection, and OWASP lists it first in its Top 10 for language model applications. We therefore separate data and instructions technically and limit what the AI can trigger. An instruction in the system prompt alone is not enough.
How we measure success
Before the start we measure how long a case takes today and how often rework is needed. In the pilot we measure the same again, including the time for checking and correction. Typical metrics are minutes per document, errors per field and the share of cases that go through without rework.
A worked example: with 1,000 documents a month and four minutes saved per document, around 67 hours a month are freed up. What that time is worth depends on what your team uses it for.
When automation is not worth it
If a process only occurs a few times a month, the effort rarely pays off. The same applies if the data is still on paper or nobody is responsible for the process. In that case we start with digitalisation or advise against it.
How a project works
Initial consultation
About 30 minutes, free of charge. You describe a process, and we assess whether automation makes sense.
AI Potential Check
Assess up to three processes, calculate the benefit with your figures, name the prerequisites. You receive the result in writing.
Pilot
One process, implemented with real cases and measured against today’s baseline. Often four to six weeks if data and interfaces are available.
Rollout
Connection to production systems, permissions, error handling, documentation and training for everyone involved.
Operation
Monitoring, updates and adjustments within a scope agreed in advance.
Frequently asked questions about AI process automation
Which systems can be connected?
All with a documented interface, that is ERP, CRM, document management, ticket systems and Microsoft 365. Without an interface, software robotics (RPA) helps.
What is the difference between RPA and AI process automation?
RPA operates programs according to fixed rules, like a person with mouse and keyboard. AI process automation additionally reads unstructured content such as emails and PDFs. The two are often combined: AI reads, RPA transfers the data into systems without an interface.
How reliably does AI read invoices and delivery notes?
That depends on the quality of the documents; it is never error-free. That is why the system checks every field against the order and master data and passes discrepancies to a person. In the pilot we measure the errors per field. Structured e-invoices are processed directly.
What happens if the AI recognises something incorrectly?
The case remains a draft until someone approves it. If the check against the ERP reveals a discrepancy, it is flagged and forwarded to the person responsible.
Does automation replace employees?
No. The AI takes over reading, typing and preparation. Checking, deciding and resolving exceptions stay with your people.
How long does it take until the first automated process?
A pilot often takes four to six weeks if data, interfaces and a contact person are available. Before that, the AI Potential Check clarifies which process is worthwhile.
Start with the most time-consuming document
Describe in two or three sentences which document costs your team the most time today. In the free initial consultation we tell you whether automation is worthwhile.
