How to run AI agents on your own server with Ollama and AIK
An AI agent needs a model to think with. Usually that model is a big cloud service, and every document the agent reads travels to it. Ollama is a free program that runs open models — on your own computer, or, as a hosted service, on Ollama's servers. AIK works with both through the same switch. Here is how, and what we can honestly say about each.
Two ways to use Ollama
Ollama Cloud. You create an account at ollama.com, add credits, take an API key and paste it into AIK's Settings. The open models run on Ollama's machines. Since 9 November 2026 Ollama Cloud is pay as you go: its free plan ended, so you pay for what you use, and AIK bills nothing on top. It is a cloud — the pages do leave your building.
Ollama on your own machine. The same program, installed locally; the models never leave the room. That needs hardware: a 7–8 billion parameter model wants about 8 GB of memory and is slow on a plain CPU, quick with a graphics card of 8 GB or more. In AIK it is the same dropdown — a local model appears in the list the moment Ollama has it.
The three steps in AIK
- Get a model. Cloud: pick one from the Ollama Cloud list. Local: an administrator installs it from Admin → Ollama Models, or in a terminal with
ollama pull <model>. - Set the default. Settings → AI/LLM — the model every agent uses unless told otherwise. Settings → Document Extraction chooses the model that reads documents; a model that can see pictures is needed there for scans and photos.
- Choose per agent. Each agent has a mode: Auto (the default), Local (a model on your machine), Cloud (a provider on your own key). One agent can read invoices locally while another writes English e-mails with a bigger cloud model, and the run log names which model each step used.
What we know, and what we do not yet
We ran Gemma 4 on Ollama Cloud every day until its free plan ended, and it was good enough for the invoice pipeline, the audit reports and the assistant. What we have not yet measured ourselves is a small model running fully locally on an office PC — how many scanned invoices a 7B model gets right, and how the local embedding model that powers document search behaves on real Bulgarian files. We will publish those numbers when we have them rather than quote someone else's.
What does not depend on the model: AIK's checks. Totals that do not add up, VAT that is not 20 % of net, an EIK that fails its checksum, a supplier seen for the first time — those are in the code, and a doubtful row waits for a person whichever model read it. That is what makes a weaker model usable at all.
Trying it
You can try AIK for free at aikautomation.com with your own model key. To have it on your own computer or server — with local Ollama set up and the first agents pointed at it — you buy a licence and we install it for you. Get in touch, or read the local-LLM page first.
More from AIK
- How to extract invoice data from PDF to Excel automatically (no code)
- A self-hosted alternative to cloud automation tools for AI workflows
- EU AI Act Article 50: how to mark AI-generated documents and calls
- How to build a phone answering AI assistant
- Invoice fraud: how a swapped IBAN gets caught before payment
- What is a "process audit" for automation, and what does it tell you
- How much does an AI agent cost to run per month
- RPA vs AI agents: which processes belong to which
- How to give Claude access to your own automation platform (MCP)
- Как да автоматизирам въвеждането на фактури в счетоводството
- AI асистент, който вдига телефона на български — как работи
- Как да си направя чатбот за сайта без програмиране
- Автоматични отговори на имейли с изкуствен интелект — безопасно ли е
- Какво е on-premise AI и защо счетоводна кантора би го предпочела
- Регламент за изкуствения интелект (AI Act) — какво трябва да пише на документите
- Как да проверя дали PDF фактура е променяна
- Алтернатива на UiPath за малки процеси
- AI за счетоводители: какво поема изкуственият интелект и какво остава за вас
- Автоматизация на счетоводството: откъде да започне малката фирма
- Изкуствен интелект за малък бизнес: пет задачи, които да дадете на AI още този месец
- Автоматизация на офис процеси: как да намерите кое си струва
- AI accounts payable automation: from inbox to approved invoice
- GDPR compliant AI automation: five questions to ask before an AI touches personal data
- AI agents for small business: what they really do, and what to give them first
- Ready-made AI agents: start with one sentence, not a blank canvas
- Готови AI агенти: започнете с едно изречение, не с празен лист
- An AI operating system for business: one place where AI work is asked for, done and checked
- AI операционна система за бизнеса: едно място, където AI работата се поръчва, върши и проверява
- Model-agnostic AI agents: OpenAI, Claude, Gemini and a local model in one platform
- AI с OpenAI, Claude, Gemini и локален модел — в една платформа, без абонамент на заявка
- Custom tools for AI agents: when the ready-made ones are not enough
- AI агент със собствени инструменти: когато готовите не стигат
- Secure AI agents: sandbox, permissions and a human in the loop
- Сигурни AI агенти: изолация, права и човек, който одобрява
- AI agent integrations: connect the tools you already use — and add your own
- Интеграции на AI агенти: свържете инструментите, които вече ползвате — и добавете свои
- Intelligent document processing without an enterprise project
- Интелигентна обработка на документи с AI — без голям проект
- Одит на процеси за автоматизация: кои процеси да автоматизирате първо
- A WhatsApp chatbot for business that also answers the phone and your website
- Как работи обработката на фактури в AIK — стъпките, проверките и предпазителите
- Как AI асистент записва часове във Viber — и кога предава на човек