Small Steps Big Impact
+32 474 13 78 13
AI IN BUSINESS / AI 6 MIN READ

What a local AI model can and cannot do for a small business

SWSenne WellensFounder

Most AI assistants run in the cloud: what you type goes to the provider's servers. A local model runs on your own computer or server instead. Open models such as Llama (Meta) or the open models of Mistral can be run that way, for example with the free tool Ollama.

That sounds like the answer to every privacy worry. It is a good answer to some of them. This is what to expect.

What it does well

Well-defined tasks on one document at a time: summarising, sorting, and pulling fixed fields out of a document such as a delivery note or an order form.

And the data stays where it is. For a company that handles other people's plans, files or contracts, that can be the whole point.

Where it falls short

The models that fit on ordinary office hardware are smaller than the largest cloud models. You notice that in tasks that need reasoning across several documents at once, and in text that goes to customers, where tone matters.

Like every language model, a local one can also give a wrong answer with full confidence. A human check stays necessary.

What it costs

No subscription, but not free. Count three things.

  • Hardware: a machine with enough memory and preferably a strong graphics card. The bigger the model, the more it needs.
  • Setup: installing the model, connecting it to your documents or workflows, and testing.
  • Maintenance: updates, new model versions and someone who looks at it when it stops. Without an owner it dies quietly.

When it is worth it

When there is a real reason why documents may not leave the building, and the task is narrow and repetitive.

If the worry is mainly 'will they train on our data?', look first at the business plans of the cloud assistants. Several of them do not train on business data by default, and that is far less work than running your own model.

Three tips to start

  • Pick one task, not 'a general assistant'. A local model is a tool for one job.
  • Test on real documents from the past, and compare the result with what a person would have filled in.
  • Decide in advance who maintains it and how many hours that may take.
More articles
[AI]NOV '26Free ChatGPT at work: how small businesses avoid a data leakStaff paste customer data into free chatbots. What happens to that data, and five measures you can take this week.[AI]JUN '26EU AI Act for SMEs: what applies to an ordinary service businessMost of the act is aimed at high-risk systems. What already applies, what comes later and what is simply good practice.[Automation]AUG '26Five n8n flows every SME should have before touching AILead intake, quote follow-up, invoice reminders, onboarding, weekly report. Boring, and the base everything else builds on.
Next step

See where the gains are.

Ready for your first small step?

Book an intro call with Senne. No strings attached.
Either way you leave with a few concrete pointers.

Senne Wellens of Small Steps Big Impact in a video call at his desk
PreviousFree ChatGPT at work: how small businesses avoid a data leak Next · 3/9Five n8n flows every SME should have before touching AI
FAQ

Questions we get a lot.

Not answered here?
01What we do

Small Steps Big Impact automates internal business processes with AI, connections between systems and custom software. We always start with an on-site screening. That way we only automate what saves time or earns money, and not what happens to be possible.

Mainly for SMEs with 5 to 30 employees where the same data is entered in several systems. Think of installation companies, technical maintenance, machine service and facility services. The more work happens on site or at the customer, the bigger the gain.

Usually quote follow-up, planning, registration of work orders, post-calculation per project or contract, invoicing based on actual work and product data from suppliers. Which process comes first we decide after the screening, based on impact.

Yes. Besides automation we give AI workshops and guide teams in using AI day to day. That way you get a return from AI without starting a full trajectory right away.

02Approach and trajectory

A screening is a two-hour session at your location. We map your processes, speak with the people who do the work and prioritise what pays off most. You get a report with findings and a 30-minute follow-up call. You are not committed to anything.

In the blueprint we set out exactly what the process, the data and the connections look like before we build. Blueprint and build always go together. That way what we build is really what you need.

Workflow automation connects existing tools and takes over recurring steps. Custom software is an application of your own, for example a field service tool, in which those automations are already built in. The screening shows which of the two fits.

One clearly defined process is often live in less than a month. A trajectory across several chains, such as sales, planning and stock, takes more like three to four months.

With a Care subscription we monitor your automations monthly, solve problems and adjust when a tool changes. Without maintenance a connection does not keep working forever either.

03Tools, team and cost

Not necessarily. We work independently of tools and prefer to connect the software you already use through APIs. We only build custom software when existing platforms do not carry your process well.

No. We build around the way your team already works today and speak with the people on the floor during the screening. That is why the solution feels to them like less work and not like an extra system.

That depends on the number of processes and on whether existing tools are enough or custom software is needed. After the screening you know exactly what we propose and what it brings in, so you decide based on your own figures and not on a standard package.

Yes. Your data stays yours and stays in your own systems: we work as much as possible within the tools you already use and follow the GDPR. Licences and access stay in your company's name where possible, we never share data with third parties, and access we no longer need is removed after the project.

Yes. We work for companies in Belgium, the Netherlands and beyond, from Limburg. Up to 90 minutes' drive we come on site, further away we work digitally or on site by request.

Not answered here?