Enrich CRM data by choosing the decision you need to improve, identifying the missing fields, and testing those fields on a fixed sample before writing anything back. Start with the data provider you already pay for. Add another provider or a Clay workflow only when the sample exposes a specific gap that is worth the extra cost.
This is a proposed evaluation framework, not a report of customer results. Product documentation was checked on November 1, 2026. The sample sizes, review rules, and arithmetic below are illustrative operating choices; they are not vendor benchmarks or promises of revenue uplift.
Start with one decision
“Complete every CRM field” is a poor brief. “Route an account to the correct territory using its operating country” is a decision you can test. Write down the action, required fields, and consequence of a wrong answer before selecting an enrichment product.
Use a company domain plus your existing CRM record ID as the account anchor. A domain alone does not resolve subsidiaries, shared brands, or acquisitions. Send ambiguous matches to review; do not let a language model silently choose the account. For contacts, confirm the person and current employer before treating an email address as usable.
| Decision | Minimum evidence | Reject or review when |
|---|---|---|
| Territory routing | Confirmed company identity and operating country | Registered address and operating location conflict |
| ICP qualification | The specific attributes in your written rubric, with sources | A required attribute is unknown or only inferred |
| Contact preparation | Current employer and independently checked contact field | Employer is stale, identity is ambiguous, or validation is missing |
| Account research | Source URL, relevant excerpt, and check date | The excerpt does not support the proposed claim |
Choose the smallest row of this table that answers your immediate problem. Do not purchase phone data for an account-level classification exercise. Do not ask Claude to invent a revenue estimate because a required field is empty.
Use a staged enrichment policy
First normalize and deduplicate the input. Preserve the original values in separate columns, and store record_id, company_domain, and the reason each row entered the run. Exclude records already carrying current, accepted values for this decision.
Next, query one source. Retain the returned value, provider, check date, and evidence where available. “Provider returned a value” and “reviewer accepted that value” are separate states. A second source is useful only if it resolves a missing or rejected field without introducing more unresolved conflicts.
Then apply a fallback to the unresolved rows, rather than re-enriching the whole table. Clay conditional runs support gating actions on field values. Apollo’s waterfall documentation describes checking multiple providers. Neither feature proves that its output is correct for your segment; that is what the sample review measures.
Finally, use Claude for extraction or classification against supplied evidence. Keep factual extraction separate from a deterministic scoring rubric. An explanation that sounds plausible is not a source. When the source is inaccessible or insufficient, keep the result unknown and route it to a person.
The Clay and Claude lead-enrichment workflow provides a starting artifact for this separation. Its offline examples are synthetic. Use them to understand the checks, then run your own live pilot before connecting CRM updates.
Run a 100-record evaluation
Take a fixed 100-record sample from the intended workload, split across the regions, company sizes, and record ages you actually serve. Keep difficult and incomplete records in the sample. A convenient list of well-known companies hides the failures you need to find.
Freeze the input, required fields, date, and acceptance rules. Give the baseline provider and candidate workflow the same records. Review each returned field against an independent source where possible; when no reference can settle the answer, mark it unresolved rather than correct.
Record the results in a worksheet with these columns:
record_id, segment, required_field, baseline_value, candidate_value,
provider, source_url, checked_at, reviewer_decision, rejection_reason,
credits_used, external_cost, review_minutes
Do not put API keys or unnecessary personal data in the worksheet. Keep source access and retention aligned with your team’s approved process. One owner must decide which evidence wins when sources disagree; “latest response wins” is not an acceptance policy.
| Metric | Calculation | What it answers |
|---|---|---|
| Coverage | Returned required fields / requested fields | Did the source find anything? |
| Reviewed precision | Accepted fields / reviewed returned fields | Were the returned fields usable? |
| Accepted-row yield | Rows passing every required check / input rows | How many records can advance? |
| Cost per accepted row | Attributed run cost / accepted rows | Is the usable output worth its cost? |
| Review burden | Review minutes / accepted rows | What manual work remains? |
Count unknowns and conflicts separately. Coverage is not precision, and accepted-row yield is not a sales conversion rate. With zero accepted rows, report cost per accepted row as undefined and stop the rollout.
Price the complete run
Use run cost = allocated subscription cost + incremental usage + external API charges + review labor. Do not add the subscription’s included credits again as incremental usage. For an existing contract, show both the incremental cost of the pilot and the fully allocated cost of operating it.
Clay documents separate Actions and Data Credits. External API keys remove the corresponding Clay Data Credit charge, not the external provider’s bill or Clay’s Actions. Apollo’s usage report lets you inspect consumption by feature and request. Capture actual usage before estimating a monthly run.
Here is an invented example, not a Clay or Apollo measurement: 100 input records produce 80 accepted rows. Attributed platform and data cost is $24; review takes 30 minutes at an assumed $40/hour. Total cost is $44, or $0.55 per accepted row. A fallback adds $8 plus 15 minutes of review and yields 10 additional accepted rows. Its marginal cost is $18 / 10 = $1.80 per additional accepted row; blended cost becomes $62 / 90 = about $0.69.
Approve the fallback only if those additional rows justify $1.80 each for your decision. The increase in coverage alone does not justify it. Replace every example input with your recorded costs, reviewer rate, and accepted output.
Buy, compose, or leave it alone
Use packaged enrichment when it passes the required-field and export checks with acceptable review effort. Use Clay when you need explicit source order, conditional steps, or an evidence-based AI classification feeding an existing system. Read the Clay versus Apollo decision guide before buying both.
Keep the status quo when the missing field does not change an action, or when review consumes more effort than manual research. Build a custom script only when someone owns its retries, field mapping, access controls, and maintenance. Tool ownership is part of the cost, not an afterthought.
For the first rollout, require review of every proposed CRM change, preserve previous values, and write only approved fields. Stop on a wrong-company match or an unexplained cost spike. Recheck a sample when the input population, provider, prompt, or rubric changes. Refresh cadence should follow observed field decay and the next business decision; a blanket monthly refresh of every field is not a strategy.