What to Do When AI Makes a Wrong Decision? The 6-Step Response and the 5 Most Common Causes
By Pedro Sonda, dooop
When AI makes a wrong decision, the company should follow six steps: contain (suspend or limit use to prevent the error from recurring), correct for those affected, communicate transparently, investigate the cause, prevent recurrence, and record the learning. The company is responsible for the error, not the tool. Investigations usually find one of five causes: data problems, model limitations, use outside the intended scope, failure in human review, or system manipulation. Responding well to an error preserves trust; trying to blame the "machine" usually destroys it.
Updated October 2026.
This content does not replace legal advice.
What Real Cases Show
| Case | What Happened | Source |
|---|---|---|
| Air Canada (2024) | The chatbot provided incorrect refund policy information; the company argued the chatbot should be responsible for itself, but the tribunal ruled the company must compensate the customer | ABA Business Law Today, 2024 |
| Deloitte Australia (2025) | A report delivered to the government contained references with errors attributed to generative AI use; the company refunded part of the amount and delivered a revised version | The Register, 2025 |
| Klarna (2025) | After automating much of customer service focusing on cost, the company acknowledged a drop in quality and reinvested in human customer service | Customer Experience Dive, 2025 |
Interpretation: AI errors generate financial, contractual, and reputational liability. The best outcomes occur when the company assumes responsibility, corrects the issue, and changes the process.
dooop Framework: The 6-Step Response
| Step | What to Do | Indicative Timeline |
|---|---|---|
| 1. Contain | Suspend the system, limit use, or activate full human review until the problem is understood | Hours |
| 2. Correct for Those Affected | Identify who was affected and reverse or compensate the impact | Days |
| 3. Communicate | Inform those affected and, if applicable, authorities clearly about what happened and what is being done | According to severity and legal requirements |
| 4. Investigate | Find the root cause, not just the symptom | Days to weeks |
| 5. Prevent | Change data, configuration, scope, human review, or the use itself | Weeks |
| 6. Record and Learn | Document the incident, decisions, and changes; share the learning | At the end |
Step 2 in Detail: The Right to Review
When a decision affects a person based on automated processing of personal data, the LGPD guarantees the right to request review (art. 20) (LGPD). Having an active contestation channel before an error occurs greatly facilitates step 2.
dooop Framework: The 5 Causes of Error
| Cause | How to Recognize | Typical Prevention |
|---|---|---|
| 1. Data | Incorrect, outdated, or incomplete source information | Correct the source; assign data ownership |
| 2. Model Limitation | Plausible but invented or wrong response in a certain case type | Restrict use to tasks where performance was tested; human review |
| 3. Use Outside Scope | System used for something other than intended | Define scope clearly; block unintended uses |
| 4. Human Review Failure | Review planned but rushed or not performed | Redesign review process; reduce reviewer workload; logging |
| 5. Manipulation | Malicious input caused improper behavior | Security controls; limit actions and access |
Cause 4 is frequent and little discussed: when AI is correct most of the time, human review tends to become a rubber stamp. See How to Create Human Review Mechanisms for AI.
What Not to Do
- Blame the tool publicly or to the customer.
- Fix the individual case without investigating the cause.
- Hide the error from those affected.
- Permanently disable AI without assessing whether the problem was design or usage related.
What to Do Now
- Define who can suspend each moderate or high-risk AI system.
- Create a contestation channel for people affected by AI decisions.
- Prepare a communication template for errors impacting customers.
Complete incident preparation is available at How to Prepare Your Company for AI-Related Incidents. To structure governance that reduces and addresses errors, see dooop’s AI Governance practice.
Frequently Asked Questions
Is the company always responsible for AI errors? Generally, the company using AI is responsible for outcomes toward clients and third parties. Division of liability with suppliers depends on contracts and specific cases and should be evaluated with legal counsel.
Is every AI error an incident? No. Minor errors corrected during review are part of operations and should be measured. Errors reaching clients or affecting people, money, or data must be treated as incidents.
Should we disclose that the error was caused by AI? Transparency about the cause usually preserves trust if accompanied by what the company did to correct and prevent it. What should not be done is using AI as an excuse.
Also Read
About dooop
dooop is a consultancy based in Caxias do Sul (RS) that prepares leaders and organizations to transform artificial intelligence into business capability through strategy, governance, and training.