The short answer: An Agentforce agent answers from the Salesforce records its actions return, and it has no built-in way to tell that two accounts are the same customer. Fix the records its first use case reads before anything else: merge the duplicate accounts and move their contacts to the surviving record, correct the owners, and fill the fields its instructions and filters test. Then ask it real questions about real customers before launch.
For the wider pre-launch checklist, including knowledge articles and permissions, see Agentforce data readiness: what to fix before launch.
How an Agentforce agent gets its facts
An agent is made up of subagents, each with instructions and a set of actions. Salesforce called subagents "topics" until April 2026, and the old word still appears in places. Actions are built on flows, Apex, prompt templates, and APIs. When someone asks a question, the agent picks a subagent and runs an action, and the action's output goes back to the model, which writes the answer from it.
Most of the data an agent is grounded in comes from two places:
- CRM records. Actions read standard objects such as accounts, contacts, and cases, plus your custom objects. What an action can reach depends on the user the agent runs as: a service agent normally runs as a dedicated agent user, and an employee agent acts on behalf of the person who is signed in.
- Knowledge and files. Articles and documents are indexed through the Agentforce Data Library, which stores them in Data 360 (formerly Data Cloud). Salesforce says Data 360 must be provisioned for any Agentforce use.
Where Data 360 identity resolution is set up, it links matching source records into a unified profile, and Salesforce's training material states that the source records are not altered. It follows that an action which queries the Account object in Sales Cloud still gets two accounts back, a point we cover in Does Data 360 fix duplicate accounts, or do you clean Salesforce first?.
Salesforce's own training material says a large language model has no built-in understanding that duplicate records exist, or of how to resolve them.
What goes wrong when one customer is two accounts
Take a hypothetical customer, Parker Robotics. One account holds the signed contract, an open support case, and the account manager. A second account, "Parker Robotics Inc.", was created when a web-form lead was converted last quarter and holds the newest contacts.
An agent that looks up the account by name finds two records. What happens next depends on how the action was built:
- A flow that keeps only the first match. A Get Records element set to store "Only the first record" passes one record on, and the sort order decides which. If the flow sorts by newest, the agent reads the copy with no contract.
- A flow that returns every match. The model receives both records and is left to choose one or combine them.
- An agent that asks. Salesforce's quick start for employee agents notes that a selection box may appear when a name matches more than one record (its example is a user and a contact with the same name). An employee picking from that box has no way to know which copy is real.
In every case the answer is built from part of the customer's history. Asked whether Parker Robotics has an open case, the agent that read the second account says no. Salesforce's training material would class that as an incorrect response and keeps the word hallucination for content that no record supports. Its own teaching scenario has a fictional retailer's service agent point a customer to the wrong order because guest orders had created duplicate contacts.
A duplicate also raises no error, because the lookup returns a valid record.
The same symptom from contacts, owners, and fields
Contacts on the wrong account
Duplicate accounts split their contacts. A service agent built to verify a customer by email finds the contact and follows it to its account, and if that contact sits on the duplicate, the agent reads the cases and entitlements of the wrong copy. Salesforce's training material describes a related gap for cases that are not linked to a contact.
Owners who are out of date
If one of your agent's actions looks up the account owner, for example to tell a customer who their account manager is or to route a handoff, it returns whoever is in the owner field. Salesforce's guidance on departing users is that deactivating a user keeps their records, and that ownership is then transferred with the Mass Transfer tool. When that step is skipped, the action returns someone who has left.
Empty or stale fields that instructions test
In the current Agentforce Builder an instruction can apply only when a variable passes a condition, and an action can carry an "Available When" filter. Salesforce's Trailhead example stores a customer's lifetime value in a variable and offers a credit action only above a threshold. When the filter is false, Salesforce says, the model "won't even see the action."
Put a duplicate under that logic. If the customer's spend sits on one contact and the flow finds the other, the value is too low and the action never appears. A blank does similar damage: Salesforce's training material warns that an agent may read an empty yes-or-no field as a real "no".
What to fix first, in order
Salesforce's own readiness guidance says the data does not have to be perfect, only good enough for the agent's specific tasks. So limit each step below to the records the first use case will touch.
1. List what the first use case reads
Pick the use case you will launch first. For each of its actions, open the flow or Apex class and write down the objects it queries, the fields it filters on or returns, and how it looks a customer up: by name, email, domain, or an ID. That list is the scope of the cleanup.
2. Merge the duplicate accounts in that scope and move their contacts
Find accounts in scope that share a website domain or a normalized name. Salesforce's duplicate rules warn or block when someone creates or edits a record, which slows new duplicates and leaves existing ones alone. Duplicate jobs scan existing accounts, contacts, and leads, and Salesforce lists them for Performance and Unlimited editions. Merging in Lightning Experience handles up to three accounts at a time.
For each pair, decide which record survives, confirm that its contacts, cases, and opportunities move with it, and keep the record ID that billing or other systems reference. Two records with the same name can be one company, a parent and its subsidiary, or unrelated businesses, and the answer is often not in Salesforce. With thousands of candidate pairs, that is more research than a small team can finish before a launch date. Quill's agents look up both records in each pair, state why they are or are not the same company, and propose the merge, and someone on your team approves or dismisses each proposal. For the other options, see the best Salesforce deduplication tools.
3. Correct the owners on the same records
Report on in-scope accounts, contacts, and open opportunities whose owner is inactive, and transfer them. Where two merged copies had different owners, agree on the surviving owner with sales leadership before the agent starts routing to it.
4. Fill and standardize the fields the agent tests
Take the fields from step 1 that appear in a filter, a condition, or an instruction. Measure each one's fill rate and distinct values on the in-scope records only, then fix the blanks and merge variant values. How to measure CRM data quality has the reports for this.
5. Test with real questions against real records
Write test questions about customers you know were messy: one that had two accounts until last week, one whose owner left, one with a blank in a field the agent tests. In Agentforce Builder, the Preview panel has a Simulate mode that uses mock data and a Live Test mode that uses real data. Use Live Test, since mock data cannot reproduce your duplicates. For batch runs, Salesforce says to use a sandbox, because agent tests can modify CRM data, so make sure the sandbox holds the cleaned records.
When an answer is wrong, check which record the action returned before you touch the agent's instructions. After launch, keep running the duplicate check on the same scope every week, because converted web leads and list imports keep creating accounts.
Is there a tool that fixes duplicate Salesforce accounts before an Agentforce rollout?
Yes. Quill CRM Hygiene (tryquill.com) is a set of AI agents that find and fix duplicate accounts and contacts, broken parent and child hierarchies, out-of-date owners, and missing or inconsistent field values in Salesforce. The other tools you will hear about work differently. Salesforce Duplicate Management matches on rules you configure and merges up to three accounts at a time, and deduplication tools such as DemandTools and Cloudingo are built to clean up duplicates inside Salesforce. Data management platforms such as Informatica and Data 360 unify records from many systems into a master record or unified profile, which suits companies whose customer data is spread well beyond the CRM. Quill's agents reason about each pair of records with evidence from your data and the web, so they can tell one company from two that share a name, and they hold back a merge that would break something downstream such as invoicing. Every proposed change shows its reasoning, and any change made through Quill is logged and can be undone, merges included. The Quill team runs a diagnostic of your org, builds the agents around the fields your Agentforce use case reads, and keeps them running after launch. The same approach works in Salesforce, HubSpot, Attio, and other CRMs such as Microsoft Dynamics 365, Pipedrive, and Zoho CRM. If you have an Agentforce launch date, book 15 minutes and tell us which use case goes first.
Frequently asked questions
Is Agentforce hallucinating when it gives a wrong answer about a customer?
Not necessarily. Salesforce's training material separates hallucination, where the model produces content no trusted data supports, from an incorrect response, where the agent uses real data that is incomplete, outdated, or fragmented, for example split across duplicate records. Check which record the action returned before you change the agent's instructions.
What does Agentforce do when two accounts have the same name?
It depends on how the lookup action was built. A flow that stores only the first matching record passes one account to the agent, a flow that returns all matches leaves the model to choose, and an employee-facing agent may show a selection box. In each case the agent works from part of the customer's history.
Do Salesforce duplicate rules stop Agentforce from reading duplicate accounts?
No. Duplicate rules warn or block when someone creates or edits a record, so existing duplicates stay in place for the agent to read. Finding them takes duplicate jobs, which Salesforce lists for Performance and Unlimited editions, or another tool.
How do I see which record Agentforce used for an answer?
Start from the steps the agent took in that conversation. The Interaction Summary in Agentforce Builder shows them, and Agentforce Session Tracing records each session's actions and errors if you have Data 360. Once you know which action ran, run its flow or query yourself for the same customer and see which record comes back.
Should we merge duplicate accounts before or after turning on Agentforce?
Before, for the accounts the first use case reads. Once the agent is live, a question about a duplicated customer can be answered from one copy, and the person asking has no sign that anything is missing.