Playbook

How to implement AI in your business: a 6-step playbook for SMEs

Most AI projects fail because they start with a tool, not a problem. The step-by-step approach I use to implement AI workflows in small businesses.

Most businesses don't fail at AI because the technology doesn't work. They fail because they start with the tool ("we need a chatbot") instead of the problem ("we lose ten hours a week re-typing orders"). The result is an expensive demo that impresses in a meeting and is never used again.

The good news for small and mid-sized businesses: you don't need an AI team or a big budget to get this right. You need one well-chosen workflow, a pilot on real work, and honest numbers.

This is the playbook I use when a company asks me where to start. It is the same way I build AI in production as Head of AI at Herald: pick one real workflow, prove it on real work with a person in the loop, measure, then scale.

Step 1: Find the work that eats your week

Before talking about models or tools, list the repetitive tasks your team does every week. For each one, write down four things:

TaskTimes per weekMinutes eachHours per month
Answering booking requests by email60416
Typing supplier invoices into accounting40616
Sending quote follow-ups151010

You will usually find two or three tasks that quietly cost a full day of someone's time every week. Those are your candidates.

Good first candidates share a few traits:

  • High volume. The task happens many times a week, not twice a year.
  • Text-heavy. The input is emails, documents, forms, PDFs or messages. This is where language models are strong.
  • Clear definition of done. You can tell whether the result is right.
  • Cheap mistakes. An error is easy to catch and fix before it causes damage.

Bad first candidates are rare tasks, tasks that need deep expert judgment, or anything where one mistake is expensive and nobody would notice it in time.

Step 2: Pick one workflow and define success in numbers

Choose one task. Not three. Then write down what success looks like, in a way you can measure:

  • "Reply to 80% of booking requests within 5 minutes, day and night."
  • "Cut time spent on invoice entry from 16 hours to 3 hours per month."
  • "Follow up on every open quote within 48 hours."

Measure where you are today before you change anything. Without a baseline, you will never know whether the project worked, and neither will anyone you need to convince.

Step 3: Choose the simplest tool that works

Not everything needs an AI agent. There are three broad options, and the cheapest one that solves the problem is the right one:

  • Rule-based automation (tools like Zapier, Make or n8n) when the steps are always the same: "when a form is submitted, add a row and send this email."
  • An AI agent when the input is messy and someone has to read it and decide: "read this supplier email, find the order, check the price and update the system."
  • An assistant or chatbot when people need answers: "what is our refund policy for this case?"

In practice, the best setups combine them: automation does the plumbing, AI reads and decides on the messy part, and a person approves what matters. I wrote a full comparison of agents, automations and chatbots if you are unsure which one you need.

Step 4: Prepare your data and access

An AI system is only as good as what it can see and do. Before building, answer these questions:

  • Where do the inputs live? A shared inbox, a folder of PDFs, a form tool, a CRM?
  • Which systems does it need to update? And do those systems have an API, or only a screen?
  • What does a good result look like? Collect 20 to 50 real past examples of the task done well. They are the best test set you will ever have.
  • What are the privacy rules? If personal data is involved, you need a legal basis under GDPR, a data processing agreement with every provider, and access limited to what the system really needs. Sensitive data, like health or financial information, needs extra care.

Step 5: Run a pilot with a person in the loop

Start in "draft mode". The AI prepares the reply, fills in the form or proposes the update, and a person on your team approves it with one click. This does three things:

  1. Nothing goes out unchecked while you are still learning how the system behaves.
  2. Your team sees exactly where it is reliable and where it is not.
  3. Every correction becomes an example that makes the system better.

Run the pilot on real work for a few weeks. Track how often the draft is accepted as is, how often it needs a small edit, and how often it is wrong. Then, case by case, let it act alone where it has proven reliable, and keep a person on the rest.

Step 6: Measure, then scale

At the end of the pilot, compare against the baseline from step 2. Look at the full picture:

  • Time saved per week, and what your team did with it.
  • Quality: response times, error rates, customer feedback.
  • Total cost: building, running (model and hosting costs) and maintaining the system. I break these down in what AI really costs a small business.

If the numbers are there, keep it running, document how it works, and move to the next workflow on your list. If they are not, you have learned that cheaply, on one workflow, instead of on a company-wide project.

Common mistakes to avoid

  • Starting with the tool. "We need a chatbot" is not a problem statement.
  • Automating a broken process. If the workflow is messy on paper, AI will make it messy faster. Fix the process first.
  • No owner. Someone on your team needs to own the workflow, review the results and say when something is off.
  • Measuring nothing. If you cannot show the time saved, the project will not survive the next budget discussion.
  • Going fully autonomous on day one. Trust is earned on real cases. Start with approval, then remove it where it is safe.

Where to start this week

Take 30 minutes with your team and fill in the table from step 1. Circle the task with the most hours per month that also has a clear definition of done. That is your first AI project.

Not sure what qualifies? These AI workflow examples show the twelve workflows businesses most often automate first.

More on how I approach this in practice: AI workflow automation.

About the author

Martin Michaux is Head of AI at Herald and was named to Forbes 30 Under 30. Before Herald he co-founded NOX Energy. He writes about what he learns building AI agents in production.

Profile AI workflows More articles
Reply

Book a call with Martin