AI for SMEs: where to start and what it delivers

You read everywhere that AI changes everything. In your own organisation you mostly see a few colleagues using ChatGPT while the work carries on as before. For SMEs, AI is no longer a technology question. The tools exist and they are affordable. The question is which process you tackle, how you measure whether it works and what agreements you make around it.

What holds businesses back is rarely money or technology. It is experience: nobody has done it before, so nobody knows where to start. This guide helps with that.

Three levels of AI in a business

It helps to separate three levels. They differ a lot in cost, lead time and risk.

Level 1: assistance

An employee uses an AI tool alongside their work: drafting, summarising, rewriting, translating, analysing an export. No integrations, no IT project. It goes wrong when there is no training and no house rules. Then half stop after a few weeks and the rest paste in data that doesn't belong there.

Level 2: configured applications

A recurring task gets a fixed form. A shared assistant that writes quote text in your house style, or one that categorises customer questions. This goes wrong when it is set up too broadly. One task done well works. "Everything around customer contact" doesn't.

Level 3: process automation with integrations

AI becomes a link in a chain of systems. Incoming invoices are read, checked and queued in the accounting software. Email is categorised and routed. Leads are enriched and scored in the CRM. This is where most of the gain is, and where it most often goes wrong on data quality and exceptions. The standard cases are sorted quickly. The exceptions decide whether it really saves time.

Most businesses do well to do all three, in this order. Level 1 builds trust and produces the examples you use to choose level 2 and 3.

What can AI do by type of business?

A few examples of applications that often come up first in different sectors:

• Professional services: draft quotes and proposals, meeting notes, making internal knowledge searchable.
• Wholesale and retail: enriching and translating product data, categorising customer questions, processing orders from email and PDF.
• Logistics and transport: document checks, customer status questions, preparing planning.
• Manufacturing and engineering: structuring work orders and reports, making maintenance history searchable.
• Construction and installation: preparing cost estimates, processing work orders, searching specifications. See also work orders and planning at an installation company.
• Administrative services: invoice and receipt processing, preparing checks, completing files.
• Healthcare and social care: supporting reporting and admin, with extra care because special categories of personal data are involved.

To assess whether a process is suitable, use the selection model in which business processes suit AI.

How do you choose the first application?

Four questions. If a process scores on all four, it is a good first candidate.

• Volume: does the task occur often enough to earn back the set-up?
• Repetition: does it mostly run the same way, or is every case different?
• Data: is the input already in a system, or only in someone's head?
• Review: can a person quickly see whether the outcome is right?

The last one is skipped most often and causes the most disappointment. One extra test: can you capture the current situation in a number? "This costs us about twelve hours a week" is a starting point you can measure against later. "This takes a lot of time" is not.

What does it cost?

Licences for generative AI tools are rarely the big item. Adoption is: teaching people to work with the tool, configuring an application, building integrations and the hours of your own people. The more systems involved, the more it costs. An indication per type of project is on what an AI solution costs. How to set the return against it is in the AI business case.

One lesson almost always holds: businesses that spend months comparing subscription prices lose more through delay than they save on licences.

Which rules apply?

Two frameworks side by side. The EU AI Act also applies to small businesses. If you use AI rather than supply it, it mainly comes down to three things: not using prohibited applications, making sure your people have sufficient AI knowledge and being open with people who deal with your AI. On top of that, the GDPR applies as usual whenever personal data is involved.

In practice that becomes one AI policy of a few pages: which tools are allowed, which data may go into them and who is responsible for what. See drafting an AI policy, AI literacy and the AI Act and GDPR: what do you need to sort out now?.

The five most common pitfalls

• Starting with the tool. You end up with licences nobody uses and a management team that concludes after six months that AI does nothing.
• Starting too big. Pick a task, not a department.
• No owner. Sponsorship from management is not the same as someone who steers it every week.
• Underestimating data quality. Much of the work is cleaning and unlocking data, not AI.
• Not measuring. Recording one number up front takes an hour and saves months of debate.

Do it yourself or bring in help?

Doing it yourself is obvious at level 1 and 2, if you have someone with a technical bent and there is no hard deadline. The learning curve is valuable in itself.

Bringing in help is wiser for integrations with your ERP or CRM, with sensitive data, under time pressure or when you don't know which process should come first. What to look for is covered in choosing an AI implementation partner. Either way, keep ownership in-house. A partner can build and speed things up, but you judge whether the outcome is good enough.

Where do you start tomorrow?

Pick one process, give it an owner and record how it runs today. The step-by-step plan is in AI implementation for SMEs. If you have several candidates and want to know which delivers most, the AI Opportunity Audit lines them up for you in a reasoned order.