AI for small businesses: where to start (hint: not a chatbot)

When a small-business owner tells me “I want to bring AI into my business,” they’re almost always picturing a chatbot. Something that answers WhatsApp around the clock.

They might need one. But in my experience, that’s rarely the right place to start.

I spent more than ten years in sales. Today I sell to restaurant owners every day and build AI systems for small and mid-sized businesses in Colombia. Here’s what I’ve learned about where to begin.

Step 1: find the number that hurts

Before talking about AI, ask: which number in my business do I want to move?

• How much money is sitting in inventory I can’t account for?
• How many hours a day does someone spend keying in, checking or reconciling?
• How many orders get lost or go out wrong?
• How long does my team take to answer an inquiry?

No number, no project. When I run a diagnosis and can’t find that number, I tell the owner. I’d rather do that than sell something that won’t move anything.

Step 2: look at your data before the AI

AI runs on data. If yours lives in a notebook, in one person’s head or across five different spreadsheets, AI won’t perform miracles.

So for many small businesses, step one isn’t an agent. It’s getting the information in order:

• What comes in (purchasing).
• What goes out (sales).
• What’s left (inventory).
• Who made each movement.

That’s why I think inventory is a great entry point. It’s less exciting than a chatbot, but it’s where almost every business loses money without seeing it.

A real example: a multi-location restaurant group

I recently spoke with a restaurant group running several locations. Their software wasn’t delivering and they wanted out. Their wish list:

• Stability.
• Every dish sold deducts its ingredients from inventory.
• Purchasing.
• A customer base for deliveries.
• A screen that’s easy to use.

Notice: nobody said “AI”. They named concrete problems. AI comes in later, where it saves real work. For example, an agent that reads a supplier invoice and drafts the purchase for a person to confirm, instead of someone typing every line.

Step 3: decide what you won’t build

This is the advice that relaxes owners the most.

Don’t build a POS or tax-compliant e-invoicing from scratch. In Colombia, as in much of Latin America, electronic invoicing is regulated and certified providers already exist. Use them. What’s worth building is the layer your POS handles poorly: recipe-level inventory, purchasing, per-location daily close, clean data for your accountant, deliveries.

Knowing what not to build saves months and money.

Step 4: start small and measure

My recommendation for almost any small business:

1. One location or one process. Not the whole company at once.
2. 30 days of measurement on real data.
3. One clear target. For example: 100% of sales and purchases recorded and reconciled with the register.
4. Decide with those numbers whether to expand.

The risk stays small for you, and if it works, you have the proof for every other location.

Step 5: a person confirms what matters

AI prepares the work. A person decides.

• A purchase drafted from an invoice gets confirmed by someone in the warehouse.
• A customer brief prepared by an agent gets reviewed by whoever handles the case.
• Where there’s legal or financial risk, there’s always a human in the loop.

That’s not distrust of AI. It’s how you build a system your team will actually use.

Step 6: ask what it costs to run

An AI system isn’t just a setup fee. It has a monthly cost: hosting, model usage, maintenance.

When I designed a routing system for a distributor, the obvious option would have cost about USD 3,600 a month in API fees alone. That would have wiped out the savings. I built a custom engine instead, at a fraction of the cost.

Always ask for that number on day one. If the person selling to you doesn’t know it, that’s a red flag.

Questions for your own diagnosis

To start today, answer these with your team:

• Which task repeats every day, and who does it?
• How long does it take, and how many errors come out of it?
• What do you use today for the register, invoicing, purchasing and inventory? What can you export?
• How many locations and users? What does each one need to see?
• What happens when the internet goes down?
• Who approves the investment, and what result would make a pilot a success?

With those answers, you have half a proposal.

And the chatbot?

It comes, at the right time. Once your data is in order, a WhatsApp agent can check real stock, take orders that land straight in your system, and hand off what it can’t resolve to a person. Without clean data, it just answers nicely.

The short version

1. Find the number that hurts.
2. Get your data in order.
3. Decide what not to build.
4. Start small and measure.
5. Let a person confirm what matters.
6. Ask what it costs to run.

If you run a business and want to see what this could look like for you, read the inventory system case study or message me on LinkedIn. I’ll show you a demo on your case before quoting anything.