CRM enrichment without the hallucinations

Half of what gets sold as 'AI enrichment' is a model making things up with confidence.

You ask for a company's headcount and the name of its CFO. Back come both values, formatted and certain. One of them is right. The other is a guess, and it just landed in your CRM looking exactly as solid as the fact beside it.

That's the problem with enrichment as most people buy it: nothing on the record marks which values it actually verified and which it made up. On the hard-to-find fields, the made-up share is big enough to hurt you.

A bigger model doesn't fix this; method does. Provenance on each value, a confidence score, and a human watching the fields that carry weight. It's dull to describe, and it's what holds up when an auditor starts asking where a number came from.

Why AI enrichment hallucinates in the first place

A language model is built to produce a plausible answer. Producing a true one is a separate task, and the two only line up most of the time.

Give it a fact it has seen a thousand times and it gets it right. Ask it for something it has never seen - a private company's exact headcount, a role that changed last month - and it still won't pause to admit the gap. It just returns whatever reads as most likely, and on enrichment work, likely and true come apart far too often.

So we're not chasing a model that never guesses. We want a system that can tell when it's guessing and handle that case differently.

Provenance: every field points back to a source

The first rule: no value enters the CRM without a source attached.

If the system writes '250 employees', it also records where it read that - the filing or the company page, with a link and a date. That source line is what lets anyone check the value in one click. Leave it off and there's nothing to check the number against.

If a field can't cite where it came from, it doesn't get written. That single rule clears out most of the risk on its own, because a fabricated value has no source to point at, so it never reaches the record. Every enrichment build we ship starts from it.

Confidence thresholds: let the system say it doesn't know

The second rule: the system scores its own confidence, and low confidence is not treated as fact.

A match on a clear, verified source is high confidence and can be written automatically. A weak match. A stale source - and sources go stale fast, with HubSpot's own data putting B2B contact decay at about 22.5% a year. Two sources that disagree. All low confidence, and none of it written as truth. It gets held and sent to a person.

Most enrichment tools run on a single setting: fill the field. A more honest one can do three different things with a value - fill it, hold it, or hand it to a human - choosing based on how sure it is.

You set that threshold yourself, and you can move it field by field. Where a wrong answer is cheap, let more through. For something you'll eventually report to a regulator, raise the bar and route more to review. Either way, you're the one deciding where the system is allowed to guess.

Human-in-the-loop on the fields that matter

People assume the human-review step is the slow, expensive part. Set up properly, it stays small.

The machine handles the fields it can verify - the overwhelming majority - on its own. Only the low-confidence handful reaches a person, already sorted, with the conflicting sources laid out side by side. At that point nobody is really enriching anything. They're making a call on a few records where a wrong answer would genuinely cost something.

This matters more the more regulated you are. I spent over a decade in regulated financial services, inside an eleven-country group, and a wrong field there was never just an inconvenience. It turned into a compliance problem that someone had to defend months later. Keeping a human on the fields that carry risk isn't caution for its own sake. In that world it's part of doing the work properly.

What to ask before you buy enrichment

Three questions tend to reveal whether an enrichment tool is honest about what it knows.

Where did this value come from - can it show the source for a given field. What happens when it isn't sure - does it write a guess or hold the field. And who checks the low-confidence ones - is there a human step, or does everything land as fact. A vendor who can answer all three is describing a real method. One who steers the conversation toward model size instead is hoping you won't press.

The method is the product

What you're paying for in good enrichment is the discipline built around the model. The model itself is close to the least interesting part.

Provenance, so every value can be traced back. Confidence scoring, so the system admits its own limits. Add a person on the fields that are worth one, and you have the whole method. None of it demos well, because the demo that sells is the one filling two hundred fields in four seconds without ever mentioning that a quarter of them were guessed. We hold to the same rule when training an agent on real data: keep it grounded in something checkable, and keep a person on the parts that matter.

Building a model that fills every field is the easy part. Getting one to flag the fields it couldn't actually fill is harder, and it's what makes the data worth trusting.

Ask the vendor which one they're selling.

Common questions