Ask an AI agent to find potential customers and it can return a convincing list. The harder part starts when that list meets your CRM: two names for the same company, missing emails and a revenue figure with no reporting period. Decide how to handle those cases before you let the agent create records.
The first useful output is a selection you can explain. At Innovadata, that means a country, supported filters and company identifiers that survive the import.
Give the agent a brief it can check
Suppose you sell software to Estonian accounting firms. “Find companies that need our product” gives the agent little to work with. Give it a concrete starting point:
Find the activity codes for accounting in Estonia. Show the available employee filters, then count companies with at least five employees and an email address. Show the codes and filters you used. Don't retrieve company records yet.
This is an example brief; it doesn't predict who will buy. An activity code narrows the registered business activity, while a size filter helps you choose accounts your team can serve. The agent should ask when the brief needs a filter the country doesn't support, rather than silently substitute something else.
Use list_filters and list_activities before count_companies. These are the discovery and count tools in the Innovadata MCP reference. Counts and metadata are free; retrieving company rows on a metered account uses credits.
Review five records before importing a list
Once the selection looks right, ask for five records and check them against your brief. Five is a review batch, not evidence that the whole list has the same contact coverage. A public Excel sample helps you inspect the column structure; it is a separate snapshot, not the output of your filters.
Look for failures you can recognise:
- The company fits the registered activity, but its website describes a business outside your target segment.
- The email field is empty. Keep it empty; an address the agent guesses from a domain belongs in a separate research field, if you keep it at all.
- The revenue value is present but its period differs from the period in your sales brief.
- The CRM already contains the company under an older name.
A general inbox also needs a different approach from a named contact. Don't let the import turn one into the other.
Match companies before creating accounts
Use country plus registration code as the external company key. Keep the identifier as text so an import cannot drop a leading zero. Company names are useful for reading the list; use the stable identifier for matching.
Define an overwrite rule for each CRM field. For example, refresh a registry address while leaving a salesperson's account notes untouched. An empty incoming contact field should not erase a contact your team has already verified unless you deliberately choose that rule.
Innovadata retrieves company data; writing it to a CRM requires a separate CRM connector or API. First show a preview of the proposed changes, with existing accounts and new accounts separated. Approve that batch before writing.
Make the second run safe
Save the filter definition outside the chat. Log the retrieval time and the identifiers your CRM accepted, and keep failed rows for review. Re-running an import should update the same accounts instead of creating them again.
Set the row and spending limits in your integration, where the job can enforce them. A sentence in a prompt helps communicate intent; the code still needs to stop the job when it reaches the limit.
Start with one market and a batch your team can review. Inspect the country data, then try the same selection in search. Outreach comes after the import review; a public email address alone doesn't settle permission to send marketing.