TellParrotEvery organisation with more than a handful of systems has duplicates. The same customer in billing, in the CRM and in the service desk, spelled three ways, with three addresses, two of them stale. That is the visible problem, and it is the one most master data tools are sold to fix.
The invisible problem is what happens when the fix is wrong.
A confident engine is not a correct one
A matching engine that merges records on a score is confident by construction. It cannot know that two records with the same surname at the same address are a father and a son. It cannot know that the account it just folded into another belongs to a subsidiary that must be reported separately. When a silent merge is wrong, the damage arrives later and somewhere else: a customer who receives a stranger's statement, a supplier paid against the wrong terms, a subject access response that discloses a person who was never part of the request.
And because the merge was silent, nobody can say when it happened, who allowed it, or what the records looked like before.
Let the machine find, let a person decide
TellParrot runs matching at the scale a real estate needs. Rules are configured per field, exact, fuzzy, phonetic or token based. Blocking keeps large datasets tractable. Candidates come back as scored duplicate clusters with a survivorship preview, and stewards can work through them in mass review or leave background matching runs to build the queue.
What the engine never does is merge on its own. A steward approves or rejects every merge. The decision lands on the ledger with who made it, in what role, and against what evidence. Merges are never silent, so the question "why is this one record" always has an answer that a person put their name to.
Undo has to mean undo
Approval is only half of the safety. The other half is being able to reverse a decision exactly. Most systems "unmerge" by recreating the original records from whatever they remember, which is usually less than the whole story.
TellParrot's unmerge is a true inverse. When records merge, the platform keeps split receipts describing what came from where. Unmerge restores the originals from those receipts rather than reconstructing them, so the estate returns to the state it was actually in, and that reversal is itself on the ledger.
Every field explains itself
A golden record in TellParrot carries field-level provenance: for each value, which source it came from and why it won. Where no rule decides survivorship, field-scoped source trust weights do, and every change to a weight is ledgered with a required reason. A per-record explain log and a plain-language narrative sit on the record for anyone who needs to read it rather than query it.
Why this matters more now
A model prompted or trained on merged records inherits every silent mistake in them, and it inherits them confidently. Human-approved merges are not a slower way to do master data. They are the only way to hand a record to an AI system and still be able to say what it is.