What Happens When an AI Agent Gets Conflicting Information?
One of the easiest ways to create a confusing AI agent is to give it information that disagrees with itself.
Imagine your knowledge contains:
“Returns are accepted within 7 days.”
Another document says:
“Returns are accepted within 14 days.”
A third file says:
“All returns require manager approval.”
The problem is no longer whether the AI has enough information.
It has too much inconsistent information.
So what should the agent do?
The first rule should be:
An AI agent should not silently choose whichever conflicting fact sounds most plausible.
The business needs an authoritative source or an explicit rule for resolving the conflict.
Where conflicting information comes from
Conflicts usually appear for ordinary business reasons.
You may have:
old PDFs
updated policy documents
outdated website copy
old pricing sheets
different departments using different terminology
duplicate knowledge sources
customer-specific information that differs from general policy
information that changes over time
The AI did not create the inconsistency.
Your information architecture did.
Static knowledge can become outdated
Suppose you upload:
“Our standard delivery time is 3–5 business days.”
Six months later, your operations team changes the policy to:
“Standard delivery is now 5–7 business days.”
If the older document remains connected to the agent, both statements may still be available.
This is why adding more knowledge is not always an improvement.
Your AI Agent Knowledge Troubleshooting Guide covers the importance of keeping knowledge sources relevant and current.
Not every conflict is equally important
There is a difference between:
“Our office is near the city centre.”
and:
“Refunds are available within 7 days.”
The first may have little operational impact.
The second can materially affect what the AI tells a customer.
You should pay particular attention to conflicts involving:
pricing
refunds
eligibility
legal or compliance policies
warranties
delivery commitments
product availability
discounts
payment rules
customer-specific account information
These are areas where a wrong answer can create real business problems.
The agent needs a source hierarchy
A useful approach is to establish which source wins when information differs.
For example:
Live customer/account data
↓
Current business system
↓
Approved policy knowledge
↓
General FAQ content
↓
Older reference material
The exact hierarchy depends on the business.
The important part is that it exists.
For example:
Order status should come from the live store.
Not:
a PDF uploaded three months ago.
A Shopify order-status agent, for example, should use the connected order data for the customer's actual order rather than guessing from general delivery information. See How to Build a Shopify Order Status AI Agent on WhatsApp.
Customer-specific information can override general information
Consider:
“Our standard cancellation policy allows cancellation within 24 hours.”
But a particular customer's order has a special contract.
The customer asks:
“Can I still cancel this order?”
The agent should not assume the standard policy is the final answer.
The customer-specific record may contain information that changes the outcome.
This is another reason to separate:
general business knowledge
from:
customer-specific data
Current state is different from history
Conflicts also appear when customer information changes.
Suppose the customer originally says:
“I'm interested in Basic.”
Later:
“Actually, I want Enterprise.”
Both statements are true historically.
Only one represents the customer's current state.
The system should distinguish:
Conversation history:
Customer previously considered Basic.
Current customer state:
Customer is now interested in Enterprise.
See What Should an AI Agent Do When a Customer Changes Their Mind?.
What should the AI do when it cannot resolve the conflict?
The safest answer is not:
“I think the policy is probably 14 days.”
Instead, the agent should avoid presenting an uncertain interpretation as fact.
For example:
“I don't have a reliable answer for that from the information available to me. I'll connect you with our team.”
That is much better than fabricating certainty.
ZazzyAgent can route such situations to a human through configured handoff rules. See How to Hand Off AI Conversations to Human Agents.
Do not hide uncertainty behind confident language
AI systems can produce fluent responses even when the information is inconsistent.
That creates a dangerous illusion:
confident wording = reliable answer
It doesn't.
Suppose the knowledge says:
Free delivery above ₹999.
Another source says:
Free delivery above ₹1,499.
A smooth-sounding response doesn't solve the conflict.
The underlying information needs to be reconciled.
A practical conflict-resolution process
Use this workflow:
Conflicting information detected
↓
Is one source clearly authoritative?
↙ ↘
Yes No
↓ ↓
Use authoritative Do not guess
source ↓
↓ Escalate / ask
Answer customer ↓
Resolve source
This is a better design than expecting the language model to act as your policy department.
Knowledge should have owners
One useful operational practice is to assign ownership.
For example:
| Information | Owner |
|---|---|
| Pricing | Sales |
| Delivery policy | Operations |
| Refund policy | Support |
| Product specifications | Product |
| Customer account status | CRM / backend |
| Promotions | Marketing |
Then establish which source is official for each category.
This makes updates easier and reduces the chance of contradictory documents accumulating.
How to avoid conflicts in ZazzyAgent
Keep each agent's knowledge focused on its responsibility.
A Sales Agent may need:
products
pricing
plans
qualification criteria
A Support Agent may need:
troubleshooting
returns
warranties
service procedures
There may be some overlap.
That's fine.
The goal is not to create completely isolated information.
The goal is to avoid giving every agent every document your business owns.
Separate policy from live data
This distinction is especially important.
Use Knowledge for:
“What is your return policy?”
Use live actions for:
“Has my return been approved?”
Use Knowledge for:
“How long does standard delivery usually take?”
Use an order lookup for:
“Where is my package today?”
Static knowledge explains the business.
Live systems tell you the customer's current state.
What about conflicting customer fields?
Structured data can also become inconsistent.
For example:
Lead Source = Instagram
Later:
Lead Source = Website
Before allowing the AI to update fields, define what the field means.
Is it:
original acquisition source?
most recent source?
current channel?
manually verified source?
A field without a clear definition can create conflicts even when the AI behaves correctly.
See ZazzyAgent Custom Fields Guide.
The agent should not overwrite history accidentally
Some information should represent current state.
Other information should represent history.
For example:
Current plan: Enterprise
Previously considered: Basic
Those should not necessarily be stored in one field.
Likewise:
Current status: Qualified
Qualification history: Qualified → Not ready → Re-engaged
A good data model prevents a later conversation from destroying useful context.
How to test conflicting information
Create deliberate test cases.
For example:
Knowledge A:
Returns within 7 days.
Knowledge B:
Returns within 14 days.
Customer:
“What is your return period?”
Then observe whether the agent:
chooses a source correctly
expresses uncertainty
escalates
or gives an unsupported answer
This should happen before the agent is exposed to customers.
The best solution is usually upstream
There is a temptation to solve every conflict through prompt instructions.
For example:
“Always choose the newest information.”
That helps only when “newest” can be determined reliably.
A better solution is to maintain a clean source of truth.
The AI should not be expected to perform document governance on behalf of the business.
The takeaway
When an AI agent receives conflicting information, the solution is not simply:
“Make the AI smarter.”
The business needs to define:
which source is authoritative
which data is current
which information is customer-specific
which fields represent current state
when the agent should ask
when it should use live data
when it should stop and involve a human
The strongest AI agent is not the one that always finds an answer.
It is the one that knows when the available information is not reliable enough to give one.