A company shows up in your CRM labeled a mid-market SaaS business. It is actually a 900-person agency that resells software. One field, quietly wrong.
Nobody notices, because nobody reads the field. They read what the field feeds. And what it feeds is almost everything.
This is the part of AI-driven enrichment that no one demos, because it is not the exciting part. The exciting part is watching a model read a website and fill in a column. The part that decides whether you can trust the column is what happens after.
This guide is about that part. How you take an AI that labels your accounts and make it reliable enough to run a go-to-market motion on. Not "usually right." Reliable, measured, and provable.
A wrong field does not stay contained
Start with what that one mislabeled account actually costs. Not in the abstract. Trace it forward through the systems that read the field and never question it.
The same story runs through employee count bucketed one tier too low, an industry read from a stale label, a headquarters pulled from the wrong office. None of these are dramatic on their own. All of them are load-bearing.
Here is the uncomfortable version. Every scoring model, routing rule, and payout you build sits on top of fields you did not verify. The model is only ever as good as its ground truth. If the field is a guess, so is everything downstream of it.
You cannot trust what you cannot measure
The instinct, when an AI gets a label wrong, is to reach into the prompt and reword it until the one bad case flips. That fixes the case in front of you and quietly breaks two others you were not looking at.
Prompting by anecdote is not reliability. It is whack-a-mole with a language model. To make a field trustworthy you need the one thing anecdotes cannot give you: a score.
So you start where every serious measurement starts. With records whose answer you already know.
We call it the golden set. A few dozen to a few hundred accounts, each hand-labeled by a human who was right, chosen on purpose to include the boring middle and the genuinely hard edges. It is not a sample of your data. It is an answer key.
Once you have an answer key, "is the classifier good" stops being an opinion. You run the classifier across the golden set, compare its label to the known-true label on every row, and read a number: how often it agrees, and exactly which cases it misses.
Meet the label the AI gets wrong
Account type is a closed question. A company is SaaS, or e-commerce, or a digital agency, or a traditional business, or one of a couple of others. A short menu, one right answer per account.
Take the account from the top of this guide. On paper it is a textbook trap. It has a product page, a pricing table, a login, and the word "platform" in the hero. Read the surface and any model calls it software.
Read two more pages and the truth shows up. The button says book a call, not book a demo. The pricing is a retainer. The "platform" is how they deliver a service, and the case studies describe done-for-you work, not a product a customer logs in and runs. It is an agency wearing a SaaS website, which is, as it happens, exactly the kind of business we run.
A human who knows your market gets this in about fifteen seconds. A first pass from a model does not, because it stops at the signal that is loudest, not the signal that decides. This one account, sitting in the golden set with its true label attached, is now a test the classifier has to pass.
The bench is where you fix it
The bench is the loop where the classifier gets hardened against the golden set before it ever touches your live data. It has two halves, and the second one is the one people skip.
The score is not one number, it is a breakdown. For each possible label you see how often the model was right, and every case it got wrong, held next to the answer you know is true.
That agency-as-SaaS miss is now visible, with its evidence attached. So you do the real work: you sharpen the rule that let it through. A "Book a demo" button is not proof of software on its own. A "Book a call" CTA, retainer pricing, and done-for-you case studies are the signals of a service. You write those into the rule as the positive and negative signals that separate the two labels, so the classifier stops reading the loudest cue and starts reading the deciding one.
Then you run the entire golden set again. The point of re-running everything is that you can see, immediately, whether the fix helped one case and hurt three others. You keep the change only if the number goes up across the whole set, not just on the case you were staring at.
You loop until it passes. Not "looks better." Passes, at a bar you set in advance, on an answer key that includes the hard cases on purpose.
Passing the bench is not the same as being right
Here is the trap that catches most teams who get this far. The classifier passes the golden set, everyone exhales, and it goes to production as though the job is done.
It is not. The golden set is a few hundred rows you thought to include. Production is every account you have, including the shapes you never imagined and could not have written a test for. A bench score tells you the classifier learned its lessons. It does not tell you the lessons covered reality.
So the second loop is a different exercise, with a different question. Not "does it match the answer key" but "on real accounts, with a human looking, is it actually right, and where is it wrong in a way the bench never saw."
A person samples the live output, opens the actual company next to the label the model gave it, and renders a verdict. This is where you discover the failure modes you could not invent: the holding company that owns businesses in four categories, the rebrand mid-acquisition, the shell with no website at all.
Production QA is how you earn the right to say a number out loud. Not "the AI is usually fine," but "we sampled the live run, checked it by hand, and here is the measured accuracy of this field." That sentence is the whole product.
Every miss makes the answer key harder
The two loops are not a line, they are a circle, and one move connects them. When production QA catches a real miss, that account does not just get corrected. It gets promoted into the golden set, with its now-known-true label attached.
Now the answer key contains a case that fooled the classifier in the wild. The next version has to pass it to ship. The set that started as a few dozen obvious accounts slowly becomes a museum of every hard case your real data has ever thrown, and the bar climbs every time.
That is the flywheel. The classifier does not just get fixed once. It gets harder to fool with every pass, because the thing it is measured against keeps absorbing the ways it has failed. Reliability stops being a hope and becomes a number that only goes one direction.
Why this is the whole game
Come back to the field that started this. Account type, employee band, industry, region. Small columns, quietly load-bearing, feeding the routing and the scoring and the payout that your revenue actually runs on.
Anyone can point an AI at a website and get a label. The label is cheap. What is expensive, and what almost no one builds, is the machine that proves the label is right and keeps proving it as your data changes underneath you.
That machine is the difference between an AI that fills your CRM and one you can build a go-to-market motion on. A golden set to measure against. A bench to fix against it. Production QA to catch what the bench could not imagine. And a loop that turns every miss into the next test.
Ground truth is not a feature you turn on. It is a discipline you run. When you run it, the field stops being a guess, and everything downstream of it can finally be trusted.
That discipline is the work we do at Sculpted. We build the data and the loop that proves it, so your team can trust the CRM and run on it.