Process automation and AI integration for SMEs
Weber Development automates recurring workflows in SMEs and integrates language models (LLMs) where they replace manual work: answering enquiries, reading documents, moving data between systems. The yardstick is not whether something contains AI, but whether less gets retyped, chased and copied afterwards. Developed in Bichwil SG by Seya Weber, with experience from a chatbot platform in production and multi-stage AI pipelines.
Last updated:
01
Who this is for
Businesses where the same work is repeated by hand every day. Typical situations:
- Similar enquiries arrive daily by email or WhatsApp, and 80 percent of the answers are the same.
- Quotes, reports or orders arrive as PDFs and are retyped into the ERP or accounting system.
- Two systems have no interface, so data is entered twice.
- An AI chatbot was tried out and now produces high API bills or wrong answers.
- An existing automation in Zapier, Make or Power Automate has hit its limits, and nobody knows exactly what it does.
Industries: manufacturing and service, retail and wholesale, practices and clinics, administration.
02
What is included
- Process analysis: which steps, how often, by whom, with which exceptions
- Assessment per step: fixed rule or language model, and what pays off relative to the effort
- Workflow automation: APIs, background jobs, scripts, connections to email, ERP, CRM or accounting
- Chatbots for customer enquiries on your website or WhatsApp, connected to your own data
- LLM integration into existing software: classification, extracting data from documents, summarising, drafting replies
- Cost control: model choice per task, caching, limits, monitoring of API spend
- Quality assurance: test cases from real examples, human approval for anything that reaches customers or moves money
- Data protection: which data may leave the building and which may not, with providers chosen accordingly
- Documentation, handover, operation
03
How it works
- 01First call You show the workflow that annoys you most, ideally on screen. About 30 to 45 minutes, free of charge.
- 02Process analysis and feasibility I watch, count and test with real examples whether a rule is enough or a language model is needed. Typically 1 to 2 weeks.
- 03Pilot on one process The automation runs alongside the current way of working, and the results are compared. Typically 2 to 6 weeks.
- 04Measure and roll out When accuracy and cost are right, the pilot becomes the default and the next process follows.
- 05Operation Monitoring of cost and error cases, adjustments when the process changes.
04
What it costs
Pilot at a fixed price for one clearly bounded process.
Further development on a time-and-materials basis or as a monthly retainer when several processes follow.
Ongoing API costs of the language-model providers run on your own account without mark-up. Before the pilot you get a monthly estimate based on your volume.
05
References
FlowlyChat
Multi-tenant WhatsApp chatbot platform for SchulzMedia GmbH (since 2025) with Swiss German support, RAG via Cloudflare Vectorize and a four-stage abuse-protection system; around 65 % lower LLM costs through model routing across several Claude tiers.
Everyday Explainer
Self-initiated, self-improving AI pipeline for short-form YouTube explainers, in which specialised Claude Code subagents write scripts, fact-check claims through web research, review quality and learn from analytics.
WebGuide
Support platform in which Workers AI classifies incoming tickets and drafts articles, embedded in a ticket system with email routing.
06
Frequently asked questions
Do we need AI at all, or is ordinary automation enough?
Often a rule is enough. If the workflow is always the same (file arrives, values are transferred, email goes out), fixed automation is cheaper, faster and never wrong. A language model pays off when inputs vary: free text, different PDF layouts, customer questions in their own words. The process analysis decides this step by step.
What happens to our data when a language model processes it?
That is settled before the pilot, not afterwards. Personal data and trade secrets can be masked before processing, or the step runs through a provider with a suitable data location and contract.
What does running a chatbot cost per month?
It depends on volume: number of conversations, their length, the chosen model. Before the pilot you get an estimate based on your current enquiry numbers, and in operation limits are set so there is no surprise on the bill.
How do you stop the chatbot from talking nonsense?
Three ways. The bot answers only from the data you have provided and otherwise says it does not know. Anything binding (prices, appointments, commitments) goes to a person or a fixed system rather than to the model. And before launch it is tested with real enquiries from your inbox, not invented ones.
What if our process changes?
Rules and templates can be adjusted without starting over. Small changes go through maintenance, larger ones as a stage. If the process changes fundamentally every few weeks, automation is not yet the right step, and I will say so.
Start a conversation