AI automation for SMEs: 6 processes you can hand over today

How much of your week goes into work where you think: a sharp colleague could do this? Reading an inquiry and judging whether it is serious. Building a quote out of a messy email. Retyping an intake form into your system. Turning the monthly numbers into a short story for your team. That is exactly the kind of work where AI automation for SMEs now pays off. Not as a replacement for your people, but as someone who does the groundwork.

In this piece we walk through six processes you can hand over today. One of them, lead qualification, we work out in full, so you can see what a system like that looks like on the inside. And we are honest about what you are better off keeping.

What AI adds to ordinary automation

We wrote earlier about automating a business process as an SMB. That piece is about work with fixed rules: if this, then that. It also said that work involving a lot of judgment is hard to automate. For a long time that was true. A script cannot read an email and understand what the customer actually means.

That is where things have changed. A language model can read unstructured input: an email, a form with free text, a PDF, meeting notes. It can summarize, sort, make a first assessment and write a draft. It still makes mistakes, so a human stays in the loop. But the groundwork, the reading and the first draft, you can hand over.

Rule of thumb: messy text in, something structured out? Then it is a candidate.

1. Lead qualification

An inquiry comes in through your site or your inbox. Someone has to read it, judge whether it fits, look up some extra information and decide who picks it up. With a handful of inquiries a week, that is doable. With more, things start to slip, and it is the good lead that ends up waiting three days for a reply.

We work this one out step by step further down.

2. Quotes

A customer emails what they need, in their own words. Someone translates that into lines in your quote template. AI can read that email, pick the right products or services from your own price list, flag questions that are still open and prepare a draft quote. You check it and send it.

The prices come from your system, not from the model's memory. That matters. A model can choose and fill in, but it does not work out your rates on its own.

3. Intake

New customer, new case file, new tenant. Documents and forms come in and someone copies the details into the system. AI can read the documents, fill in the fields and point out what is missing or does not add up. Your colleague checks instead of retyping.

4. Reporting

The numbers usually already live somewhere. What takes time is the story that goes with them. What stood out this month, where did things fall behind, what does the team need to know? AI can write a first version of that commentary from your dashboard or export.

One caution: do not let the model calculate the numbers itself. Let your system do the math and the model do the explaining.

5. Content

Product copy, job ads, newsletters, social posts, answers to common questions. AI is good at a first version in the right shape. Rarely at the final version. This is how we use it ourselves: the model delivers an outline, and a person turns it into something that sounds like your business. Anything that goes out under your name gets read by someone first. We described how we handle that in AI in the loop, not on autopilot.

6. Follow-up

A quote goes out and then it goes quiet. A customer had a question and nobody wrote back. AI can keep track of what is open, prepare a draft follow-up at the right moment that refers to what was discussed before, and flag when something has been sitting too long. Sending stays behind an approval.

Worked out: a lead qualification pipeline

Here is what a system like that looks like in practice.

Step 1: intake. An inquiry comes in through the contact form, an email or a chat.

Step 2: read and enrich. The model reads the inquiry and pulls out the core: who it is, what they are asking, whether a budget or timeline is mentioned. Where that is allowed, the system looks up extra public information, such as the company's website. Anything not in the inquiry gets marked as unknown, not made up.

Step 3: score against your criteria. This is where the real work sits, and it is not technical. You decide what a good lead is. Which sector, which size, what kind of question fits what you do? We write those criteria down together. The model judges each inquiry against them and gives a score with a short explanation. That explanation matters: it shows you why, so you can correct course when the model weighs something differently than you would.

Step 4: draft reply. Every inquiry gets a draft ready. For a strong lead, a personal reply with a suggestion to talk. For a maybe, a few pointed questions. For something that does not fit, a friendly no, possibly with a referral.

Step 5: a human approves. Someone on your team sees the list: inquiry, score, explanation, draft. Approve, edit or reject. Only then does anything go out.

Step 6: into your CRM. The lead goes into your CRM or your own system with all the details and the score, so the follow-up continues from there.

What does that get you? The person who looks at your leads only has to decide, not dig. And a good inquiry no longer sits for days.

There are a few things we always set up in a pipeline like this. Every decision gets logged, with the input and the explanation, so you can see afterwards what happened. There is a fixed set of test inquiries we use to check that a change to the instructions does not make things worse somewhere else. And the criteria do not live in someone's head. They are written down and kept in version control, like the rest of the code. Why that matters so much, you can read in prompts are code too.

Does the pipeline need to talk to the software you already use, like your CRM or your planning tool? That works through ordinary integrations, and more and more through a standard built for exactly that. We explain how in MCP: how to connect AI to the software you already use.

What you are better off not handing over

Not everything belongs with a model:

• Decisions with big consequences for people. Who you hire, or who gets credit. Let AI do groundwork there at most, and be extra careful when it does. The AI Act looks at this more closely too.
• Anything where money leaves the building directly. Payments, discounts, price agreements. A model can make a suggestion, not make the call.
• Work you rarely do. One tricky report per quarter? Then the effort of setting it up and testing it does not outweigh the time saved.
• Processes that are not clear yet. If your team does not agree on what a good lead is, a model is not going to settle that for you. It just makes the mess faster.

Where do you start with AI automation for SMEs?

Pick one process. Ideally one that happens often, clearly takes time today, and where you can tell afterwards whether it went well. Lead qualification and intake are often good first candidates, because the outcome is easy to check.

Measure the current state first. How many inquiries a week, how long until the first reply, how many hours does it take? Without that baseline you will not know later whether it paid off.

Start small, with a human approving everything. A first working version is usually up in 2 to 4 weeks. Then you let it run alongside your team, see where the model gets it wrong and sharpen the instructions. Only once it runs reliably do you consider letting simple steps go through without approval.

What something like this costs depends on how many systems are involved and how much of your own data the model needs. In what does it cost to build an AI agent in 2026? we lay out the ranges. If you would rather add AI to a system you already have, read adding AI to your existing web application: where to start?.

Since 2017 we have been building apps, web applications and AI systems for SMEs and start-ups. What we keep seeing: the gain rarely comes from the smartest model. It comes from a well-defined process, with someone keeping an eye on it.

Shall we look at which process comes first?

Not sure which process would pay off most for you? Or wondering whether your lead flow is big enough to build something for? Book a call. We will look at how things run now, and we will say so plainly if a simple tool or a better team agreement is enough.

Frequently asked questions

Can AI handle my leads fully on its own?

Technically it can, but we advise against it. Let the model read, score and prepare a reply, and let a person approve. Only once you have seen over time that it runs reliably should you let simple steps run automatically.

Which process should I automate with AI first?

One that happens often, clearly takes time today, and has an outcome you can easily check. Lead qualification and intake are often good candidates. Start with one process and measure how it goes now, so you know later what it delivered.

How long does it take before something like this works?

A first working version is usually up in 2 to 4 weeks. A version that runs in your daily work, with integrations, logging and tests, takes more like 6 to 12 weeks, depending on how many systems are involved.

How is this different from ordinary automation?

Ordinary automation follows fixed rules. AI can read unstructured input, like emails and documents, and make an assessment of it. For work with fixed rules, ordinary automation is often cheaper and more reliable.