Data quality
Duplicate finder
Tells a true duplicate from a look-alike, like a hospital and its foundation, and shows the reasons for every call.
Loading page…
Commercial Real Estate Firm
An AI investigator for bad company records, with a person making every call
Fixing a single bad company record meant a manual investigation across several systems and the public web, so backlogs grew and key-client data drifted.
Built an investigation agent with read-only data access, a structured findings format, and a review queue where a person approves every fix before anything changes.
Record investigations that took hours now arrive as findings ready to approve
The situation
This firm's client and deal data depends on company records being right: the correct parent company, the correct match to a purchased company database, current details and clean name variants. When a record went wrong, someone on the data team investigated it by hand, querying several tables and checking the web, then fixed it. A small batch could take most of a day. We built an agent that does the investigation and hands a person a clear finding to approve.
Company records attached to the wrong parent or matched to the wrong entry in a purchased company database
Stale enrichment data and inconsistent name variants spread across systems
Each investigation done by hand, so checking even a handful of records could take most of a day
The delivery
A job can be a single record, a plain-English question or a whole list uploaded as a CSV, such as a key-client list. The agent reads the relevant tables (read-only), follows leads and checks the public web, then a second, smaller model turns what it found into a structured finding: the issue, the current value, the suggested fix, the evidence and a confidence level, with the record's deal count so reviewers can prioritise. Reviewers approve, dismiss or send a finding for manual review in one click, and the fix is made in the firm's own admin tool. The agent never writes to the system of record, merges that can't be undone always go to a person, and every decision is logged.
Built an investigation agent with read-only data access, a structured findings format, and a review queue where a person approves every fix before anything changes.
The result
Whole key-client lists investigated in bulk from a single upload
Every finding comes with its evidence, a suggested fix and a confidence level
Nothing changes in the system of record without a person signing off
A clear path from assisted review today to auto-fixing low-risk issues
Try it yourself
We rebuilt parts of this engagement as working demos you can talk to. The data is invented, so nothing from the client appears, but the problems are the real ones.
Tells a true duplicate from a look-alike, like a hospital and its foundation, and shows the reasons for every call.
Scope
Contact
Tell us where the business is stuck. We'll shape the strategy, enablement, and engineering around the outcome that matters.