AI implementation for SMEs: from a first process to part of the daily work

You have licences for ChatGPT or Copilot, a few enthusiastic colleagues and a management team that wants to know what it delivers. Meanwhile the real work carries on as always: reading email, retyping data, putting quotes together. Implementing AI in a mid-sized organisation means making one of those tasks run differently for good. Not a demo, but a fixed part of the work, with an owner and a measurement.

This article shows how to approach it: from choosing the first process to the moment you decide to stop, improve or scale up.

What does implementing AI actually mean?

An employee who occasionally asks ChatGPT to rewrite a text is using AI. That is useful, but it is not an implementation yet. You are implementing when AI gets a permanent place in a business process. A workflow that sorts incoming customer emails, looks up the right data, drafts a reply and puts it in front of a colleague for review is an implementation.

Such an implementation always has the same parts:

• a clearly scoped problem in one process;
• an owner who is responsible for quality;
• input that is available digitally;
• a workflow that talks to your existing systems;
• agreements on when a person checks the output;
• success criteria you set in advance;
• management after go-live.

Why do so many AI projects never get past the demo?

Almost never because of the technology. The model usually does fine in the demo. Things go wrong around it:

• The project starts with a tool instead of a problem. "Let's do something with AI" is not a brief.
• The first pilot is too big. "Automate customer service" stalls, "sort and route incoming email" does not.
• There is no owner. Management sponsors it, but nobody steers it week by week.
• Data quality disappoints. Cleaning and unlocking data takes longer than building.
• Nothing was measured. Without a baseline, every review becomes a difference of opinion.
• Staff don't know what will happen to their work and keep their distance.

More on this pattern in why AI pilots stall. The steps below are designed to avoid those pitfalls one by one.

Step 1: choose one process

Start with a list of tasks people spend a lot of time on. Think in tasks, not departments: reading and routing email, gathering quote information, checking invoices, logging call notes in the CRM, enriching product data. Talk to the people who do the work. They know where it gets stuck; management often doesn't.

A task is a good first candidate if it happens often, mostly follows the same steps, has digital input, produces an output you can check quickly and has a clear owner. How to weigh candidates against each other is covered in which business processes suit AI.

Step 2: measure how it works today

Record how often the task occurs, how many minutes it takes each time, how often something goes wrong and how long work waits before someone picks it up. A sample of a few weeks is enough. The page on the cost of manual work shows how to set up that calculation.

Keep time saved and money saved apart. An hour that frees up is only a saving if the work really disappears. Often it is capacity: the same team can handle more. That is valuable too, but it is something different. How to calculate it honestly is explained in the AI business case.

Step 3: decide what role AI plays

Not every problem needs a language model. Broadly, there are four forms, from simple to complex:

• Fixed rules: if the amount exceeds a threshold, the invoice goes to a second approver. No AI needed.
• Recognising and extracting: pulling the subject, customer number and urgency from an email.
• Generating: writing a summary or draft reply based on your own information.
• An agent: combining several steps and systems, such as reading a request, checking data and preparing a draft record in the CRM.

Choose the simplest form that solves the problem. An agent is not automatically better than a short workflow. The difference is explained in AI agent or regular automation.

Step 4: sort out data, integrations and permissions

Before you build, find out where the input lives, which systems need to be read and updated, and whether they have an API. Establish which personal data is involved and which actions must never happen without human approval. Give the application only the permissions the task needs.

Look at the rules too. The GDPR applies as soon as personal data is involved, and the EU AI Act sets requirements depending on the application and your role. What to arrange in practice is covered in the AI Act and GDPR: what do you need to sort out now?.

Step 5: build a small pilot

A good pilot is small enough to test quickly and big enough to measure something. Limit it to one mailbox, one document type or one team. Agree up front what the pilot may and may not do. Classifying, summarising, suggesting fields and drafting: fine. Making payments, making commitments or deleting data on its own: not yet.

Set three success criteria before you start. One on quality (how many suggestions are right first time), one on time (minutes per item) and one on risk (for example: not a single message sent to the wrong customer).

Step 6: test with real examples

Don't only test with the tidy cases. Include emails with missing data, duplicate requests, badly scanned attachments, other languages and messages that try to trick the workflow. Check for each example whether the outcome is right, whether the source is visible and whether the workflow is honest when it isn't sure.

Step 7: keep a person in the loop

Let the pilot start by making suggestions that a colleague approves or corrects. Record who reviews, which errors are escalated straight away and how corrections are stored. Make reviewing quicker than doing the work yourself. Otherwise people work around it or approve everything unseen.

Only when the score is stable do you switch categories to automatic, one at a time. Where there are financial, legal or personnel consequences, a person often simply keeps signing off.

Step 8: measure and decide

Compare the pilot with your baseline on four points: quality, speed, usage and risk. Then there are three honest outcomes.

• Stop, if quality stays too low, volume disappoints or the risks can't be managed. That is not a failure: you have avoided a bad investment.
• Improve, if the basics work but a few errors keep coming back.
• Scale up, if the figures have been stable for a while, ownership is in place and you know what it costs at higher volume.

Put that decision point in the diary in advance. A pilot without an end date stays half-alive forever.

Who needs to be involved internally?

At least three roles: a business owner who sets the goals and decides on budget and scope, a process expert who knows the exceptions and supplies test cases, and a technical owner who manages integrations, logging and incidents. In a small team one person can hold two roles, as long as all three responsibilities are named. No IT department of your own? Read implementing AI without an IT department.

What does implementing AI cost?

It depends on the process, the number of integrations, security requirements and volume. An internal workflow on one mailbox is not the same as an agent updating three systems. Don't forget your own hours: the process expert who tests and the owner who steers. An indication per type of project is on what an AI solution costs.

Where do you start if you have several candidates?

Most organisations have more ideas than capacity. Then it helps to put the processes side by side on volume, time, error rate and how easy they are to integrate. That is what the AI Opportunity Audit does: we map your processes and put the opportunities in a reasoned order. If you already know which process it will be, AI implementation starts there directly.