Why isn't my AI agent finding the right answer?

If your AI agent returns the wrong answer, an incomplete answer, or no useful answer, start by checking where the information is supposed to come from.

Most answer problems can be traced to one of these areas:

• The information is missing
• The wrong node is being used
• The source is outdated or unclear
• The request is routed to the wrong path
• AI Instructions conflict with the source
• A file or Article is not ready or selected
• Structured data does not contain the expected record
• The published version does not include your latest changes

1. Identify the expected source

Ask:

Where should the correct answer come from?

For example:

Policy
→ Document Search

Product price
→ Data Search or Products

Current weather
→ Live Web Search

Calculation result
→ Calculations

Appointment availability
→ Calendar through Scheduling

If you do not know the intended source, troubleshooting becomes harder.

2. Confirm the information actually exists

Open the source and verify the answer is present.

Do not assume the information is there.

Check:

• Exact policy wording
• Product record
• Data row
• Article
• Uploaded document
• Calendar availability
• Calculation input or formula

If the source does not contain the answer, the AI agent cannot reliably retrieve it.

3. Check whether the right node is being used

A common problem is using the wrong type of source.

For example:

• Long policies belong in Document Search
• Structured prices belong in Data Search
• Current web information belongs in Live Web Search
• Business rules belong in AI Instructions

If exact structured data is stored only in a long text document, retrieval may be less reliable than using Data Search.

4. Check Router behavior

If Router sends the request to the wrong path, the correct source may never be reached.

For example:

A pricing request should go to Data Search but is routed to Document Search.

Test several versions of the same customer intent.

If routing is inconsistent, adjust the route definitions.

5. Check AI Instructions

Instructions can accidentally block or distort a correct answer.

Look for rules that:

• Conflict with each other
• Tell the AI to avoid a topic
• Force a specific response
• Require information that is not available
• Tell the AI to use the wrong source

Keep important instructions clear and specific.

6. Check Document Search sources

If the answer should come from Document Search, confirm:

• The correct file or Article is selected
• The source is ready
• The source contains the answer
• The content is not outdated
• The wording is clear
• Important exceptions are near the main rule
• Conflicting documents are not both active

Then test the question using different wording.

7. Check Articles

If the source is an Article, confirm that:

• The Article is published
• The latest version is saved
• The Article is available to Document Search
• The correct Article is selected
• The content is still current

Draft Articles should not be relied on as published Document Search knowledge.

8. Check Data Search

If the answer should come from structured data, verify:

• The record exists
• Column names are clear
• The correct file is selected
• Values are formatted consistently
• The lookup field contains the expected value
• Related datasets join correctly when multiple files are used

Compare the returned result with the source spreadsheet.

9. Check Products

If the AI agent should show a Product Card, confirm:

• The Product exists
• Name and description are accurate
• Category is appropriate
• Price is current
• Options are correct
• The Product Card skill is enabled
• The Product intent is being recognized

Test both specific and broader product requests.

10. Check Live Web Search

If the answer requires current public information, confirm that Live Web Search is enabled.

If web access is restricted, make sure the approved websites contain the information.

A restriction that is too narrow can prevent the AI agent from finding a current answer.

11. Check calculation inputs

If the wrong answer is numerical, verify:

• The formula
• Required variables
• Input values
• Lookup factors
• Units
• Default assumptions
• Whether any required value was missing

Compare the result with a known correct calculation.

12. Check Calendar availability

If the problem is appointment availability, verify:

• Calendar is open for appointments
• Weekly hours are correct
• Timezone is correct
• Appointment length is correct
• Gap between appointments is correct
• Minimum notice is correct
• Booking window is correct
• The requested slot is actually free

The AI agent should not invent availability outside the Calendar rules.

13. Check whether the latest version is published

You may have fixed the workflow but still be testing an older published version.

After making important changes:

• Save the configuration
• Publish the updated version
• Start a fresh conversation
• Test again

Old conversation context can also make debugging confusing.

14. Test with a fresh conversation

Start from the beginning.

Ask the same question without previous conversation history.

This helps determine whether the problem comes from:

• Current configuration
• Old context
• Previously collected information
• A routing decision made earlier

15. Ask the same question in different ways

Try:

“What is your cancellation policy?”

“Can I cancel?”

“What happens if I cancel tomorrow?”

If one wording works and another does not, the issue may be related to routing or how the source is structured.

16. Test a question you know should work

Use a simple known example.

For example:

• A product you know exists
• A policy with a clear heading
• A known price
• A known Calendar slot
• A known calculation

This helps separate a general configuration problem from one difficult edge case.

17. Remove conflicting sources

If two sources disagree, decide which one is correct.

For example:

Document A

Refunds take 5 business days.

Document B

Refunds take 10 business days.

Do not expect the AI agent to determine which business rule is authoritative.

Update or remove the outdated source.

18. Improve unclear source content

If the source itself is ambiguous, improve it.

For example:

Instead of:

Late cancellation: 50%.

write:

Cancellations made less than 24 hours before the appointment are charged 50% of the service price.

Clear source content usually improves retrieval.

19. Check whether the request is actually in scope

Sometimes the AI agent is correctly refusing or redirecting because the request is outside its intended role.

Review:

• Scope
• AI Instructions
• Router paths
• Enabled skills
• Human Handoff rules

Make sure the workflow is supposed to handle the request.

A quick troubleshooting order

When an answer is wrong, check in this order:

1. What source should provide the answer?
2. Does that source contain the correct information?
3. Is the right node being used?
4. Is Router sending the request to the right path?
5. Are AI Instructions interfering?
6. Is the latest version published?
7. Does the same issue happen in a fresh conversation?

This resolves many common problems quickly.

Tips for troubleshooting answers

• Start with the source of truth.
• Verify the answer exists before changing instructions.
• Use the right node for the information type.
• Check Router before debugging the destination node.
• Remove conflicting or outdated knowledge.
• Compare Data Search results with the source file.
• Compare calculations with a known correct result.
• Compare Scheduling with Calendar availability.
• Publish changes before re-testing.
• Use fresh conversations for clean tests.

Related guides

For more information, see:

• Should I use AI Answer, Document Search, or Data Search?
• How should I prepare documents for better AI answers?
• How should I write AI Instructions?
• How should I test my AI agent before publishing?
• How do I route different customer requests?