Two records with the same name are sometimes two companies
Duplicate detection by string similarity flags subsidiaries, companies that have changed their names, and other companies that happen to share a word. Teams turn the checks off, because merging those records would be worse than leaving the duplicates. The data then continues to get worse.
If you are deciding whether two records refer to the same entity, you have to go beyond the CRM fields. Whether the domains are the same, or whether one company acquired the other last year, is not in the fields being compared.
Contradictions work the same way. A field that contradicts another might be an error, or it might be a deliberate override for a reason that was never recorded.
Gaps only matter if you group the data by industry. An empty industry field is significant then. If you are not grouping by industry, it does not matter. That is a business question, not a data question.
How the agent is built in Agent Studio
Automatic merging does not happen. Put that rule in both the system prompt and the flow. The agent produces two lists, one that is safe to merge and one that needs a person, and never acts on them.
The system prompt also specifies the fields your segmentation actually depends on, so the gap report is sorted by consequence rather than by missing content.
MCP connections provide external checks that let the agent distinguish a subsidiary from a duplicate.
A Start Scheduled flow drives the sweep. Salesforce or HubSpot supplies the records. A Loop node processes the suspicious pairs, and the Agent Chat node then requests a verdict for each pair.
Both lists include the reasoning. The Slack node posts only the counts, not the contents, because a hygiene report should be opened on purpose rather than skimmed.
Run it against a copy first, then compare the results with your own judgment on 50 records. Trust comes from a sample, not from a statement in a system prompt.
What this agent is built from
- System prompt: no automatic merges or deletions, and the fields your segmentation genuinely depends on.
- MCP connections: external checks, which are how a subsidiary is distinguished from a duplicate.
- Salesforce (Integration): supplies the records. HubSpot fits the same slot.
- Loop (Control) plus Agent Chat (Util): one verdict per suspicious pair on a schedule.
- Google Sheets (Integration): receives the safe-to-merge and needs-a-person lists with reasoning.
- Slack (Communication): posts count rather than content, so the report is opened deliberately.
- When an ActionFlow is enough: If you are filling blank fields with standard formatting, that is a workflow. The agent is for calls a rule cannot make.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.