On Monday afternoon, I nearly merged two clients into one
Last Monday, I opened my client spreadsheet and found company emails mixed into two similar profiles. I casually asked AI which entry was genuine. Its answer was confident, but I wasn’t: was this a duplicate client, or the same company using a new contact email? The list became messier as I cleaned it, and I was stuck for half an hour, still afraid to merge anything.
If you’ve faced records that “look identical but can’t be confirmed,” don’t immediately hand everything to your most expensive model. I used to do exactly that: the budget was fine, but my clients paid the price for my mistakes.
Let cheap models screen first, stronger models review the exceptions
This is called “tiered review.” The first tier uses lower-cost large language models to check obvious matches and conflicts, handling ordinary records directly. Only suspected duplicates, incomplete profiles, or records where the rules conflict get passed to a stronger, more expensive model.
HighTouch described this approach in its identity-resolution review product, claiming it could deliver accuracy close to Opus at about one-quarter of the cost. Lin Ran, a 24-year-old independent developer in Guangzhou, uses a similar setup to clean client databases. At 2 p.m., she sent 800 ambiguous records into the first tier; only 63 remained for an Opus-assisted, human-style review. “Not every record is worth the most expensive judgment,” she says.
This doesn’t turn AI into an infallible answer machine. Whether sensitive fields are redacted, how deletion history is preserved, and whether final merges happen automatically still need confirmation. For payment, medical, or legal identity data, I definitely wouldn’t overwrite the original record based on a single judgment.
Reproduce it today: the budget isn’t low, but the test can be small
Money: Using the pricing logic in the article, 1,000 records would cost about $8. Actual costs will vary with the model, record length, and number of retries. Time: The first setup takes roughly two hours; each later batch takes about 10 minutes. Technical barrier: I need to copy spreadsheets, organize two columns of data, and know how to preserve the original version. I don’t need to write code. First step: Create a sample table with 30 records, choose “New workflow” or “Upload data” in an AI tool, and select “Extract company name, email, and website for each entry.” Don’t let it merge records yet.
I also got this wrong. My first attempt enabled automatic changes and merged two clients with the same name. I switched to “recommendations only, don’t modify the source table,” added a “Needs review” tag, and only then felt comfortable continuing. If I’m not ready to try it now, running one round on redacted fake data is much safer than touching real client names immediately.
Match the setup to my current stage
If I’m just starting out, with few clients and simple data, I’d manually check the key fields rather than rush into a workflow. Reviewing duplicates once a month would be enough. Once I have one or two stable clients and same-name companies begin appearing in my list, I’d select 30 ambiguous records, run two rounds of judgment, compare the cheaper model with the stronger one, and then decide whether to scale. If I’m expanding and large volumes of leads arrive automatically every day, I’d keep automatic screening while logging the reviewer, timestamp, and original value for every change, so anomalies can be traced and reversed.
Not everyone needs the tooling. With only a few records, spending an extra five minutes checking manually is often better. The parts worth automating are the high-frequency, repetitive tasks that are easy to overlook.