AI agents and process automation for SMEs

The term AI agent is used a lot and not always consistently. We use a working definition that works in practice: an AI agent is an autonomous piece of software that gets a goal, decides itself which steps are needed, and executes those steps across multiple systems.

It performs tasks, not just answers questions

A chatbot replies: "Your invoice was sent on 12 May." An agent says nothing, it creates the invoice, checks it, sends it, and logs it in your bookkeeping. The difference is in the execution, not the conversation.

It decides within bounds

A regular workflow follows a fixed sequence: if this, then that. An agent can judge. Which quote does it follow up, which does it leave. Which mail does it handle itself, which does it forward to a human. Within bounds you set in advance.

It works across multiple systems at once

An agent reads from your mailbox, writes to your ERP, checks your CRM and posts to Slack. Not as separate integrations you have to make yourself, but as a coherent work process. That is exactly where manual work in SMEs gets stuck: between systems that do not talk to each other.

Honest story: not every process needs an AI agent. For much recurring work in SMEs a regular workflow is cheaper and more reliable. An agent is only needed once something has to be understood or a choice made.

When a regular workflow suffices

Order in, then create invoice, then mail to customer. Fixed steps, no judgement needed. Boring, reliable, cheap. For most administrative work in SMEs this is the right route. We also build such workflows, with n8n where handy, with custom code where it needs to be more stable.

When an AI agent adds value

Once text has to be read, a mail assessed, a document understood, an exception classified. An incoming quote request in free text. A service ticket with handwriting. A mail that could be both a complaint and a new order. Then AI comes in, because a fixed rule cannot catch that.

During scoping we first check whether an agent is really needed. Often a combination is most practical: a workflow for the mechanical part, an agent for the part where judgement or understanding is needed.

Mailbox and communication agents

An agent that triages your mailbox or customer communication. It reads incoming mails, recognises the message type (quote request, invoice question, complaint, support, spam), extracts the relevant information, and routes to the right person or system.

Quote and sales agents

Agents that follow up open quotes at the right moment, send a follow-up in the right tone, and give a summary of the customer history to the salesperson. No lead disappearing among the mails.

Work-preparation and production agents

An agent reads an order, finds comparable earlier orders in the ERP, checks material availability and prepares a draft work order. A work planner who normally spends three hours per order checks the draft in ten minutes.

Document and administration agents

Agents that process incoming documents: service tickets, invoices, contracts, certificates. Read, validate, prepare for booking. At a fleet owner in heavy transport we built this for service-ticket processing: sixty percent less manual entry.

We build agents ourselves, no no-code wrapping

Many AI agent providers connect no-code platforms (n8n, Zapier, Make) to an LLM API and call that an agent. That works for simple flows. For production environments where agents must run stably, we prefer to build in Laravel or Python. Transparent about when we do what.

Direct LLM integration, no intermediate layer

Most agencies build with LangChain or LlamaIndex. For proof of concept that works fast. In production we regularly see issues: slow responses, unexpected behaviour on updates, poor error handling. We work directly with the APIs of Claude, GPT-4o and open-source models.

Multi-agent architecture as an option, not hype

We have our own multi-agent platform (VoidOS) where we build, test and roll out agents before they go to clients. That gives us experience with problems that never appear in marketing material: context handover between agents, memory management, error recovery.

Self-hosted on your own infrastructure

Agents often work with sensitive data: mailboxes, quotes, customer data. With us the orchestration layer runs on your own server. No data via a third party, no vendor lock-in on an agent platform. GDPR and EU AI Act compliant.

Industrial context as the basis

Patrick, our founder, worked eighteen years in production and maintenance. The agents we build for manufacturing are not the standard office agents. But work-preparation agents, quality-control agents, sanctions-list-check agents, class-file agents. Specific work processes, not generic automation.

• A defined workflow without AI component (n8n or custom code): from 3,000 euro
• A single AI agent on a defined process (mailbox triage, quote follow-up, document processing): from 7,500 euro
• Multi-agent platform with multiple cooperating agents and own orchestration: from 25,000 euro

Besides the build cost there are usage costs: API costs of the LLM provider. For a typical mailbox agent reckon on fifty to two hundred euro per month. Our rates and approach are transparent on the approach page.

View the demo →