Costs

How much does AI cost for a small business? What you actually pay for

AI doesn’t have to be expensive. What small and mid-sized businesses really pay for with AI workflows, what drives the cost up, and how to keep it low.

"How much does AI cost?" is the first question most business owners ask me. The honest answer: less than you probably fear, if you start the right way. The expensive AI projects are almost always the ones that started too big.

Here is what you actually pay for, what makes AI expensive, and how to keep it affordable without cutting corners.

The three costs of an AI workflow

1. Building it (one-off)

Designing the workflow, connecting it to your tools, writing the instructions and checks, and testing it on your real data. This is usually the largest part of the cost, and it depends mostly on three things:

  • How many tools need to be connected. Reading one inbox and writing to one CRM is simple. Five systems with no API is not.
  • How messy the input is. One invoice layout is easy. Two hundred suppliers with their own formats takes more work.
  • How many exceptions there are. Every "except when..." in the process needs a rule, a check or a person.

A focused pilot on one workflow is a small project. A company-wide "AI platform" is a big one, and rarely a good first step.

2. Running it (ongoing)

Every time the AI reads an email or a document, you pay the model provider for the text it processes. For most everyday business tasks, that comes to somewhere between a fraction of a cent and a few cents per email or document. Add modest hosting costs for the workflow itself.

To put that in perspective: a workflow that reads 2,000 emails a month at a few cents each costs roughly tens of euros a month to run. Compare that with the hours of work it replaces.

3. Maintaining it (ongoing, small)

Your tools change, your processes change, and AI models get updated. Plan for a little time each month to check results, handle new edge cases and improve the workflow. It is far less than the build, but it is not zero, and anyone who says otherwise is not being straight with you.

What makes AI projects expensive

  • Starting too big. "Let's automate customer service" is a programme. "Let's draft replies to order-status emails" is a project.
  • Buying a platform before knowing the problem. Per-seat licences for tools nobody uses add up fast.
  • Training your own model. For almost every small-business task, existing models from the big providers do the job. Training your own is slow, expensive and usually unnecessary.
  • Automating a messy process. If nobody can explain the workflow on one page, AI will not fix it. It will just make the mess faster.
  • Measuring nothing. Without a baseline, you cannot tell a good project from a bad one, so bad ones run for months.

How to keep it affordable

  • One workflow, one number. Pick one task and one metric, like hours per month or response time.
  • Use the tools you already have. The AI should work in your email, CRM, accounting software and agenda, not in a new platform you have to pay for and learn.
  • Use existing models. Pay per use, switch provider if a better or cheaper one appears.
  • Pilot with a person approving. You get value from week one, without the risk of a fully autonomous system going wrong.
  • Fix the scope and price of the pilot. You know the cost before you commit, and you decide on the next step with real numbers.
  • Reuse what works. Proven patterns for reading documents, checking data and routing exceptions do not need to be invented again on your budget.

How to know if it pays off

A simple way to estimate it before you start:

Example
Hours spent on the task per month16 hours
Loaded hourly cost of the people doing it€40
Monthly cost of the task today€640
Share the AI workflow can take over75%
Time value given back per month€480
Running cost of the workflow per monthTens of euros

These numbers are only an illustration. Fill in your own. If the monthly value is clearly higher than the running cost, the question becomes how quickly the build pays for itself, and for a well-chosen first workflow that is usually a matter of months.

Not every benefit shows up in hours, either. Faster replies win more customers, fewer typing errors mean fewer costly corrections, and your team spends its time on work that actually needs people.

Questions to ask any AI provider (including me)

  1. What exactly will be automated, and how will we measure success?
  2. What will it cost to run per month at our volume?
  3. What happens when the AI is unsure?
  4. Can we switch model providers later, or are we locked in?
  5. Where does our data go, and who can see it?

If a provider cannot answer these clearly, keep looking.

The bottom line

AI is affordable for small and mid-sized businesses when you start with one real problem, measure it, and build only what you need. That is also the most reliable way to get it working.

For the full approach, read how to implement AI in your business.

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