TOP 10 mistakes when using AI and how to fix them in your business

Ten mistakes in under two minutes

Cloudata's video summarizes this checklist. Use it to start a conversation with your team, then explore the examples and practical steps in this guide.

This TOP 10 is an editorial selection, with no statistical ranking of frequency or severity. The examples set in Peru are fictional and illustrative; they do not describe clients or measured results. Each correction's priority depends on the consequences for your operation.

1. Believing an answer because it sounds confident

A convincing answer can include false figures or references. For example, an assistant compares suppliers in Lima and attributes an unverified certification to one of them.

What to do: open the original sources, check dates, and verify the claims supporting the decision. Ask the assistant to identify missing information. The NIST Generative AI Profile [1] recommends checking sources and citations during testing and monitoring.

Limit: displaying a link does not prove that the source supports the claim. Someone needs to read it.

2. Asking for results without explaining the purpose

“Write a summary” leaves length, audience, and content open to interpretation. A warehouse manager in Arequipa might need outstanding deliveries and receive a sales overview instead.

What to do: specify the task, audience, available information, and format. For example: “Summarize these orders for the warehouse manager. List the promised date, product, and outstanding item. If a date is missing, flag it without estimating.” OpenAI's prompting guidance [2] recommends precise instructions and relevant context.

Limit: a clear request reduces ambiguity, but it cannot fix incorrect data or guarantee accuracy.

3. Using outdated or incomplete sources

An assistant might prepare a quote in Peruvian soles using last month's price list. Even if the answer accurately reproduces the document, the price is no longer current.

What to do: identify the authoritative source, its version, and its date. Assign an owner to update it and check that changes reach the retrieval system. For systems that retrieve documents to generate answers, known as RAG, Microsoft's documentation [3] covers index updates and relevant content selection.

Limit: connecting documents does not guarantee completeness or retrieval of the correct version. Test both.

4. Sharing more sensitive information than necessary

To identify reasons for complaints, a team might upload a full spreadsheet containing names, phone numbers, and identity document numbers that the task does not need.

What to do: prepare a copy with essential fields, remove unnecessary identifiers, and use tools approved by your organization. Review who can access the information and how it is retained. OWASP [4] recommends limiting access to necessary data and controlling inputs and sources.

Limit: removing names does not automatically anonymize a file. Free text and other columns can still identify people.

5. Granting permissions the assistant does not need

An assistant checking stock does not need to delete products or change prices. Keeping those permissions allows an error to affect operational records.

What to do: start with read-only access, limit tools to the task, and enforce permissions in the systems executing operations. Remove connections that are no longer used. OWASP's guidance on excessive agency [5] distinguishes unnecessary functionality, permissions, and autonomy.

Limit: writing “do not delete anything” in the instructions is not equivalent to removing deletion permission.

6. Following instructions hidden in documents

A supplier's PDF could contain an instruction to ignore the original request or disclose information. This manipulation is known as prompt injection.

What to do: treat external documents as content to analyze. Separate their text from system instructions, control sources, and validate actions before execution. The OWASP RAG Security Cheat Sheet [6] explains these measures.

Limit: preventive wording alone cannot guarantee protection. Combine access controls, validation, and adversarial testing.

7. Automating important decisions without review

Preparing a proposal and committing the company are separate steps. An assistant could draft a customer response and send it with an unauthorized discount.

What to do: define who approves payments, contractual changes, and high-impact communications. Before approval, the reviewer should see the specific action, its recipient, and its consequences. OWASP [5] recommends human approval for high-impact actions.

Limit: review loses value when it becomes an automatic click. The reviewer needs sufficient information and the ability to stop the action.

8. Testing only easy cases

A demonstration with clear questions misses everyday situations: an order without a reference code, conflicting dates, or a connection failing during a lookup.

What to do: combine typical, ambiguous, incomplete, and adversarial cases. Define in advance when the system should answer, ask for clarification, or refer the task to a person. OpenAI's evaluation guidance [7] includes typical, edge, and adversarial cases; testing connection failures extends the review to the full workflow.

Limit: passing a test suite demonstrates performance on those cases. Add errors encountered during real use.

9. Measuring only whether it responds

A support team in Trujillo could produce more drafts while spending more time correcting them. Counting responses leaves that problem out of the measurement.

What to do: compare requests of similar difficulty with the previous process. Track the quality of accepted results, total time, and cost per resolved case, including review and rework. Microsoft's design principles for AI workloads [8] cover quality, performance, and cost measurement.

Limit: a fast, inexpensive response can still fail the objective. Define an acceptable result first.

10. Launching it and forgetting about it

A change to the model, instructions, or documents can alter a previously useful result. For example, a new returns policy requires checking the assistant's answers.

What to do: assign an owner, record incidents without retaining unnecessary data, and set alerts with a defined response. Repeat evaluations after changes and keep a way to stop the workflow or return to the previous process. OpenAI [7] recommends continuous evaluation and expanding tests with new cases.

Limit: a dashboard surfaces signals; someone must investigate them and address the cause.

Start with one specific process

Choose a task your team already knows and complete this review:

• What counts as a correct result, and who checks it?
• Which sources does it use, and who keeps them current?
• Which data and permissions does it actually need?
• Which actions need approval, and what happens when information is missing?
• How will you compare quality, time, and cost with the current process?

Address the risks of information exposure or unauthorized actions first. Then test the complete workflow and use evidence of the results to decide whether to expand its use.

To discuss your case, explore Cloudata's approach to data, AI, and automation [9]. These recommendations need to be adapted to each organization and do not guarantee security or regulatory compliance.