Skip to main content
Artificial intelligence

AI or automation: how to tell what you are being sold

Much of what is sold as “AI” is rule-based automation, and that is not a flaw. A simple test to tell the difference, and why it changes the price and the risk.

Digi4··6 min read

If the same document always gives the same result, and that result can be explained by a written rule, it is automation, not AI. And that is often good news: automation is more predictable, easier to test and generally cheaper to run. AI becomes useful when the input is too varied to describe in rules: free text, a document whose layout changes, a request phrased a thousand different ways.

The difference is not a matter of vocabulary. It changes the price, the risk of error, how you test, and what human checks you need to plan for. Here is how to spot it in a commercial proposal.

Updated on 17 November 2026.

What is automation, in practice?

A program that applies rules someone has written: “if the amount exceeds 50,000 DH, send it to the director”, “every evening, export orders to accounting”, “if the ICE field is empty, block the record”. The result is fully determined by the input and the rules. You can test it case by case, and when it is wrong, you find the faulty rule and fix it.

Approval workflows, automatic reminders, syncing between software, matching amounts, commission calculations, generating documents from a template: all of this is automation. It is the core of most of the business platforms we build.

What is genuinely AI?

A task where you cannot write the rule, because the input varies too much. Three families come up in companies:

  • Reading a document whose form changes: invoices from a hundred different suppliers, contracts, handwritten forms. See AI document reading.
  • Understanding free text: classifying a complaint, summarising an exchange, spotting that an email is a cancellation request. See bulk processing.
  • Judging similarity: do two labels refer to the same product, does this file look like past fraudulent files. See checks and validation.

In all three cases the result is probabilistic: the model proposes an answer, with a risk of error that depends on your documents. You do not test it case by case, you measure it on a sample, and you plan for a person to validate what is uncertain.

The five-question test

Ask the vendor, in writing if possible:

  1. Does the same document always give the same result? If yes, it is very likely automation.
  2. Can you show me the rule that produces this result? If the answer is a list of conditions, it is automation. If it is “the model learned”, ask what it learned from and how it is measured.
  3. What is the error rate, and on which documents was it measured? Real AI has an error rate. A vendor who announces none, or quotes a figure measured somewhere other than on your documents, has not done the work yet.
  4. What happens when the system is not sure? The right answer describes a screen where a person sees the document and corrects it. “It is always sure” is not a good answer.
  5. Where does my data go? AI often relies on a model hosted abroad, which raises the question of law 09-08. Rule-based automation usually runs on your own server.

Why does everyone say “AI”?

Because the word sells. A decision-tree chatbot, a form that calculates a quote or a dashboard that colours delays are regularly presented as AI. It is not necessarily dishonest, but it blurs two things you need to know before signing: how the system can go wrong, and what you are really paying for.

The reverse also happens: you are offered a language model for a task that a three-line rule would handle better. You then pay a volume-based usage cost, latency and a risk of error for a problem that did not need them. We say so in our checks offer: what can be written as a precise rule stays a rule.

How does it change the price?

NeedSuitable approachRange (DH excl. VAT)Assumption
Workflow, reminders, syncing, calculationsAutomation in a business module40,000 – 180,000Custom business module, rules written with you. Source: Digi4 price list
Finding out whether AI adds anythingAI audit15,000 – 30,000One or two processes reviewed. Source: Digi4 price list
Reading, classifying or summarising varied contentAI prototype on your documents12,000 – 60,000One use case, quality measured. Source: Digi4 price list
Putting AI into production in a workflowAI integration18,000 – 200,000Human validation and logging included. Source: Digi4 price list

On top of this, for AI only, there is a model usage cost that depends on the volume processed and is billed separately by the provider. Rule-based automation has no such variable cost: once built, its running cost is the server and maintenance (15 to 20% of the build cost per year for a platform we developed, an estimate published on our price list).

The right project often mixes both

In a real process, AI rarely takes more than one step. Take a supplier invoice: AI reads the document and proposes the fields; rules check that the total is the sum of the lines, that the ICE exists in the supplier master data, that the amount matches the purchase order; a workflow sends it for approval above a threshold. Only one step in four is AI. The rest is classic automation, and that is what makes the result reliable.

That is also why it is often more cost-effective to add an AI step to your existing software than to buy an “AI product” that redoes everything.

When neither is the right answer

If the process is not stable, automating it freezes the mess. If the volume is low (a few dozen documents a month), manual entry is often cheaper than any project. And if off-the-shelf software already does the job, a SaaS subscription or a configured standard ERP will be cheaper than custom development: we say so during the diagnostic, which costs 12,000 to 30,000 DH excl. VAT and is deducted from the scoping if the project goes ahead.

What our case studies show, and what they do not

Our published case studies are business platforms, not AI projects. Sportiq, for example, manages licences, transfers, competitions and treasury for sports federations: it is large-scale automation with business rules, and it is the foundation an AI step can later plug into. We would rather tell you that than relabel these projects as “AI”. Our AI projects are presented separately, with their limits: ScanClear, Isnad or MeetScribe.

To find out which applies to your case: describe your need.

Frequently asked questions

Is a chatbot always AI?

No. A chatbot that offers buttons and follows a question tree is automation. It becomes AI when it understands a free-form question, and then it can be wrong.

Is automation less modern?

No. It is more predictable and easier to audit. For many processes, it is simply the right tool.

Can you start with automation and add AI later?

Yes, and it is often the safest order: clean data and a stable process make the AI step easier to measure.

Newsletter

Get our upcoming articles straight to your inbox.

Message a Digi4 manager on WhatsApp