# AI automation: the mistakes to avoid when you start from processes

> The most common mistakes an SME makes when automating a process with AI, drawn from Power Automate, Make and Zapier documentation, and how to avoid them.

Source: Giallo Studio (giallostudio.it) · Page: https://giallostudio.it/en/blog/automazione-ai-errori-da-evitare · Published: 2026-11-02 · Updated: 2026-11-02

Five mistakes to avoid emerge from the Power Automate, Make and Zapier documentation when you automate a process with AI: turning on a flow without having tried it, ignoring the connections the program flags, writing vague requests, expecting the AI to do things it cannot do, and not reading how costs are counted. The official documentation hints at each of them, even when it does not call them "mistakes".

By AI automation we mean a flow that starts from an event (an email arrives, a form is filled in) and carries out a series of steps, at least one of which is handed to an artificial intelligence model: reading a text, classifying it, summarizing it, proposing a reply. The flow is repeatable, and for exactly that reason a setup mistake is repeated every time the flow runs.

A candid note: the sources used here are product documentation pages, not research on SMEs. They do not measure how often a mistake happens. They do tell us what the vendors recommend and which limits they declare, which is a good starting point for not repeating them.

<div class="key-points">
<p class="kp-title">In summary</p>
<ul>
<li>After saving a flow, the vendors recommend trying it before entrusting it with real work.</li>
<li>Connections that need attention are flagged with a red exclamation mark: they must be resolved, not bypassed.</li>
<li>A request written in the form "when X happens, do Y" gives more precise results than a generic description.</li>
<li>Zapier's AI does not learn from previous runs and does not read websites or internet addresses.</li>
<li>On Zapier and Make, cost depends on how tasks, tools, operations and credits are counted: read this first, not afterwards.</li>
</ul>
</div>

## Why start from the process and not from the tool?

An automation tool does what the process tells it to do. If the process is confused, the flow will replicate it faster and with fewer human eyes checking it. That is why the first job is not choosing the platform, but writing in plain words what happens, who decides and what must stay under a person's control.

On how to choose what to automate, our [practical guide to process automation in an SME](https://giallostudio.it/en/blog/how-to-automate-business-processes) is the reference on the topic. Here we focus on what goes wrong afterwards, when the flow already exists.

<div class="callout">
<p><strong>Limits of the sources.</strong> The materials gathered do not say which processes to automate first or what happens if the starting data is wrong. On these two points you will not find statements of fact on this page: they call for a check on your own case and, if you like, a conversation with a consultant.</p>
</div>

## Which mistakes emerge from the documentation when you automate a process with AI?

These are the five mistakes that emerge from the documentation. For each one we explain why it happens and how to avoid it.

### Turning on the flow without having tried it

**Why it happens.** A newly created flow seems to work: the editor shows no errors, the description is reasonable. The rush to "have it in production already" makes people skip the test.

**How to avoid it.** Microsoft, in its guide to creating a cloud flow with Copilot, recommends testing the flow once it is saved, as reported in the [Power Automate documentation](https://learn.microsoft.com/it-it/power-automate/create-cloud-flow-using-copilot). Zapier follows the same logic: in the right-hand panel you use Preview to try the request, then check the result and correct the request or the settings, as described in the [AI by Zapier guide](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data). Test with real messages or documents, including the awkward ones.

### Ignoring flagged connections

**Why it happens.** Connections are the links to email, spreadsheets and management systems. When one is not right, people in a hurry tend to treat it as a technical detail to fix "later".

**How to avoid it.** According to the [Power Automate documentation](https://learn.microsoft.com/it-it/power-automate/create-cloud-flow-using-copilot), connections that need attention are marked with a red exclamation mark. Treat it like a traffic light: as long as it is there, the flow is not ready.

### Writing vague requests and stopping at the first result

**Why it happens.** The process is described the way you would tell it to a colleague, with a lot left unsaid. If the first result is not right, it gets discarded instead of corrected.

**How to avoid it.** The Power Automate guide suggests writing the request in the format "When X occurs, run Y" to get more accurate results, and editing and rerunning it if the first outcome is not convincing: you can find this on [Microsoft's page](https://learn.microsoft.com/it-it/power-automate/create-cloud-flow-using-copilot). A useful detail on Zapier: versions of the request are saved only during the session and, once the editor is closed, only the current one remains, as explained in the [AI by Zapier documentation](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data). If a wording works, copy it into a document of your own.

### Expecting from the AI what it does not do

**Why it happens.** The names of the tools suggest a collaborator who learns and knows how to search. In reality, every product has precise limits.

**How to avoid it.** Read the declared limits before designing the process:

- Copilot in Power Automate is a technology still in development, optimized for English and with limited support for other languages, and it is available only in the new designer, not in the classic one ([Microsoft](https://learn.microsoft.com/it-it/power-automate/create-cloud-flow-using-copilot)).
- AI by Zapier does not learn from previous actions: on every run its knowledge returns to the initial state. It also cannot search for and extract information from websites or internet addresses ([Zapier](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data)).
- For PDFs, Zapier tries to extract the text and works better with documents created or exported as PDF than with scanned documents ([Zapier](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data)).
- Very large knowledge sources, from 1,000 pages or files upwards, may time out when they are added as a new source ([Zapier](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data)).
- The AI step may fail if the content violates the usage policies of the underlying providers, which Zapier lists as subprocessors ([Zapier](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data)).

The practical consequence: a flow in Italian on Power Automate should be tried with real Italian texts, and a process that requires memory between one case and the next should be designed with external storage, not by relying on the model.

### Not reading how costs are counted

**Why it happens.** The price of a plan is looked at straight away; the way usage is counted, much less.

**How to avoid it.** Before turning it on, write down next to the flow what it consumes. The detail is in the next section.

## How much can the way consumption is counted matter?

The sources do not report plan prices and figures, which change over time and should be checked on the official pages. They do, however, report the counting rules, which matter as much as the price list.

| Platform | What to understand before turning on | Source |
| --- | --- | --- |
| Make | Operations and credits are related but distinct concepts: operations count the activities performed in scenarios, credits are what you pay for those activities | [Make, credits and operations](https://help.make.com/credits-and-operations.md) |
| Zapier | A step with a standard model uses 1 task per run; with Advanced (3x) or Premium (5x) tasks are consumed at the rate of that tier | [Zapier, AI by Zapier](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data) |
| Zapier | If the step uses a tool, each call carries an additional cost | [Zapier, AI by Zapier](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data) |
| Zapier | On the free plan you cannot add tools or publish a Zap with an AI step | [Zapier, AI by Zapier](https://help.zapier.com/hc/en-us/articles/8496342944013-Use-AI-by-Zapier-to-analyze-and-return-data) |

This means that two flows that look identical can cost different amounts depending on the model tier chosen and the tools connected. For a broader view of costs you can read [how much it costs to automate a business process](https://giallostudio.it/en/blog/business-process-automation-cost).

## How do these mistakes show up in a real SME?

Here are two hypothetical examples, built to illustrate the mechanisms: they are not client cases.

**A professional firm.** A small consulting firm wants incoming email requests to be summarized and sorted. Whoever sets up the flow writes a generic request, saves it and turns it on. After a few days they notice that some emails in Italian are misinterpreted and that a connection to the shared mailbox had stayed flagged. The fixes: try the flow with real emails, resolve the flagged connection, and rephrase the request in the form "when X arrives, do Y".

**A distribution company.** A business that sells spare parts would like the AI to read price lists on suppliers' websites. The step does not work, because AI by Zapier does not extract information from websites or internet addresses. The solution lies in the process: have the price lists sent as files and have them read from there, checking that they are not scans. You can find the same theme, from another angle, in [automate a management software without API](https://giallostudio.it/en/blog/automate-business-software-without-api).

If your process touches customers and sales, [AI for CRM and sales](https://giallostudio.it/en/blog/ai-for-crm-sales) also shows where a setup mistake affects your relationship with people.

## How do you apply all this, step by step?

Before you start, check these prerequisites.

<ul class="checklist">
<li>You have described the process in words, with the starting event, the steps and the person responsible.</li>
<li>You know which tool you will use and in which version of the editor (for Power Automate's Copilot you need the new designer).</li>
<li>You have read how the platform counts tasks, operations or credits.</li>
<li>You have real examples available on which to try the flow.</li>
</ul>

Then proceed in this order.

<ol class="steps">
<li><strong>Write the request in the form "when X happens, do Y".</strong> One action at a time, with concrete terms.</li>
<li><strong>Resolve the flagged connections.</strong> No red exclamation mark should remain.</li>
<li><strong>Test and read the result.</strong> Use Preview where available, then test the saved flow with real cases and edge cases.</li>
<li><strong>Correct and try again.</strong> If the outcome is not convincing, edit the request and rerun it; keep the good version outside the editor.</li>
<li><strong>Turn on with a reduced scope.</strong> A person checks the first results before the flow is left on its own.</li>
</ol>

### What can go wrong even then

A flow can stop working after weeks: a connection expires, a content item is rejected by the model's providers, a file changes format. Zapier notes that you can analyze run errors by having the AI describe them in plain language, with suggested fixes, as indicated in the [Zapier AI quick start guide](https://help.zapier.com/hc/en-us/articles/18590756459277-AI-quick-start-guide-in-Zapier). Make also lets you be alerted: we cover it in [getting alerts when an automation fails on Make](https://giallostudio.it/en/blog/make-event-subscriptions-automation-errors).

If you work in the Microsoft ecosystem, [Copilot Studio and Power Automate for agents and workflows](https://giallostudio.it/en/blog/copilot-studio-power-automate-workflows) clarifies how the two tools divide the tasks. To understand where each platform is the better fit, see also the overview of [automation with Make](https://giallostudio.it/en/blog/make-com-automation) and the one on [automation with Zapier](https://giallostudio.it/en/blog/zapier-automation).

## Why is it not enough to replace a person with AI?

Because every automatic step shifts control from "who does" to "who checks". The point also holds outside written flows: in voice conversations, for example, the same principles of testing and supervision appear in the [mistakes to avoid with voice AI](https://giallostudio.it/en/blog/voice-ai-mistakes). And if your doubt is whether you really need AI or a simpler automation is enough, [what RPA is and when it makes sense](https://giallostudio.it/en/blog/what-is-rpa) helps you tell the cases apart.

## Where to go from here

Keep the sequence in mind: clear process, precise request, connections in order, tests with real cases, consumption understood. If you want a conversation about your process before building the flow, you can see how we work on [automation of business processes](https://giallostudio.it/en/business-process-automation).

## Frequently asked questions

### Can I switch on a flow created with Copilot in Power Automate right away?

Better not to. Microsoft recommends testing the flow after saving it. Try it with real cases, and also check the connections it flags, before letting it run on its own.

### How should I write the request to get a more precise flow?

Use the form "When X happens, do Y", which Microsoft indicates for more accurate results. If the first outcome is not what you expected, edit the request and run it again.

### Does Zapier's AI remember previous runs?

No. According to the documentation, the AI's knowledge returns to its initial state on every run. If the process requires memory, you need storage outside the model.

### Can Zapier's AI read the content of a website?

No, the documentation states that it cannot search for and extract information from websites or internet addresses. It is best to give it the content as text or as a file.

### On Make, are operations and credits the same thing?

No. They are related but distinct concepts: operations count the activities performed in scenarios, while credits are what you pay for those activities. Read how your platform counts them before turning the flow on.

### Which processes should be automated first?

The sources used in this article do not answer this question. To get your bearings, you can start from the general guide to process automation in an SME and check your case with a conversation about your own work.
