Data enrichment: filling the gaps in your CRM

Every CRM has holes. Someone filled in a form with an email and nothing else. A record from 2023 lists a job title that changed twice since. Enrichment is the family of techniques for closing those gaps automatically.

It gets sold as a data problem. It is really a decision problem. The question is never "can we fill this field" — a vendor will happily fill any field you pay for. The question is what you would do differently if the field were populated and correct.

What enrichment actually does

Three distinct jobs hide under one word, and they fail in different ways.

Your own first-party data — what someone did on your site, what they bought, what they wrote in a support ticket — is not really enrichment. It is data you already own and failed to join up. Do that first: it is free, it is accurate, and it beats anything you can buy. If records land incomplete because information is being retyped between systems, the fix is upstream — see how to automate data entry.

Why enriched data goes stale faster than people expect

Contact records decay because people move, and the scale of that movement is measurable: the Bureau of Labor Statistics recorded 62.8 million total separations across US employers in 2025, 38.0 million of them voluntary quits. Every one is a title, an email address and a direct line that changed in somebody's database and nowhere else.

So "we enriched the database last year" is not a state you can be in. Enrichment is a maintenance schedule, not a project, and the re-check cadence should follow how much the field matters. Company headcount can drift for a year without hurting anything. A direct dial on an account your team is actively working cannot.

Bad data has a price. Gartner's research puts the average cost of poor data quality at at least $12.9 million a year per organization — a figure drawn from large enterprises, but the mechanism scales down: wasted outreach, misrouted leads, decisions made on numbers that were wrong.

The rule that saves the most money

Enrich the field only if a documented decision depends on it.

Write the decision down before you buy anything. Not "we want better data" — the actual sentence. "If headcount is over 200, the lead routes to the enterprise rep instead of the self-serve sequence." "If the contact's title contains VP or above, we call within an hour."

Then check three things about each field:

1. Does the decision actually get made? Plenty of teams have a routing rule written down that everyone bypasses. Enriching the input to a rule nobody follows changes nothing.
2. Is the field's error rate tolerable for that decision? A 5% error rate is fine for prioritising a call list and unacceptable for setting a price.
3. Can you tell when it's wrong? Some errors surface immediately — a bounced email, a disconnected number. Others never surface at all, which is worse, because they quietly bias every report built on them.

Most fields on a vendor's feature list will not survive the first question.

How to run enrichment without breaking things

Match on the strongest identifier you have. Email domain is the usual join key and it is weaker than it looks — free-mail addresses match nothing, and large domains match the wrong subsidiary. Where you have a company website, a registration number or a verified account ID, use it. Score the match itself separately from the field.

Never overwrite a human-entered value silently. If a rep typed the title after a conversation, that beats a crawl. Write enriched values to their own fields. The moment you overwrite, you have destroyed the only copy of information a person actually verified.

Store provenance with every value — source, retrieval date, confidence, and the identifier it matched on. Four extra columns are the difference between a database you can audit and a pile of assertions.

Route low-confidence matches to a human queue rather than accepting or discarding them. Auto-accepting introduces silent errors; discarding throws away most of your coverage.

Test on a sample you can verify independently. Take 100 records you actually know the truth about and run them through the vendor before signing. Published coverage and accuracy rates are marketing; your sample is evidence. Same discipline as any other automation, as covered in CRM automation.

The costs nobody prices in

Enrichment vendors charge per credit or per record. That is the visible cost. The others:

• Deliverability. Enriched lists tempt teams into volume outreach on unverified addresses, which damages sending reputation. Google asks bulk senders — anyone sending close to 5,000 messages a day to personal Gmail accounts — to keep spam complaint rates below 0.1%, and to never reach 0.3% or higher. A list padded with guessed addresses is a fast route past that line, and the damage lands on your whole domain, not just the campaign.
• Legal exposure. Enriched personal data was, by definition, not obtained from the person. Under GDPR Article 14, that triggers an obligation to inform them — within one month, or at first communication if that comes sooner. US state privacy laws add their own disclosure and opt-out duties. Have an answer before the first regulator or prospect asks, and get it from a lawyer rather than a vendor's compliance page.
• Trust inside the building. The fastest way to make a sales team stop believing the CRM is to fill it with confident-looking data that is often wrong. Once they stop trusting the fields, they keep their own spreadsheet, and you are paying for a system nobody reads.

Frequently asked questions

Is data enrichment worth paying for?

It is worth paying for when a specific downstream decision changes based on the enriched field, and not otherwise. The test is whether you can write the rule as a sentence — "leads above 200 employees route to the enterprise team" — and confirm the rule is genuinely followed. Enrichment that only makes records look more complete produces no return, because nobody reads the extra fields. Start by enriching one field that feeds one rule, measure whether the routing or scoring improved, then expand. Most teams over-buy fields and under-use them.

How accurate is purchased contact data?

Accuracy varies widely by field and decays continuously, and no vendor's published figure should be taken at face value. Company-level attributes like industry and approximate headcount tend to hold up reasonably; individual job titles and direct phone numbers degrade fastest, because people change roles constantly — US employers recorded 62.8 million separations in 2025. The only reliable measurement is your own: assemble a sample of records whose truth you can verify independently, run them through the vendor before committing, and check both how many were filled and how many were filled correctly.

What is the difference between data enrichment and data cleansing?

Cleansing fixes what is already in the record — deduplicating accounts, standardising formats, correcting obvious errors, removing dead entries. Enrichment adds information the record never contained, drawn from an outside source. They run in that order for a practical reason: enriching a database full of duplicates multiplies the duplicates and the cost, because you pay to enrich the same company three times and then reconcile three conflicting answers. Cleanse first, establish a single record per entity, then enrich against a clean key.

Do I need consent to enrich personal data?

Consent is one lawful basis among several, and the requirements depend on jurisdiction and on what you do with the data — so this is a question for a lawyer, not a vendor. What is clear is that obtaining personal data from a third party rather than the person creates a transparency duty: under GDPR Article 14, the controller must inform the individual, at the latest within one month of obtaining the data or at the time of first communication with them, whichever comes first. US state privacy laws add their own disclosure and opt-out requirements. Decide your position before you buy, not after a complaint.

Can AI replace commercial enrichment vendors?

Partly, and unevenly. Language models are good at normalising and classifying data you already hold — mapping free-text company descriptions to industry categories, cleaning inconsistent job titles, inferring seniority from a title string. They are unreliable for retrieving facts they were not given, because a model asked for a phone number it does not know will often produce a plausible one. The safe division is to use models for transformation and classification over data you possess, and retrieval tools with a verifiable source for anything factual. Any inferred value should be labelled as inferred.

How often should enriched data be refreshed?

Set the cadence per field, based on how fast it moves and how much the decision costs when it is wrong. Company-level attributes — industry, rough headcount, location — can typically go six to twelve months. Attributes tied to employment, like title and direct dial, need checking before any campaign that depends on them. A useful pattern is refreshing on access rather than on a schedule: when a record enters an active workflow, check whether its enrichment is older than your threshold and re-verify then, which concentrates spend on the records you are actually using.