Waterfall Enrichment for Candidate Sourcing

What is waterfall enrichment?

Waterfall enrichment is the practice of passing each record through several data providers in sequence. Provider one runs on the full list, provider two runs only on what provider one failed to find, and so on down the chain. The result is higher coverage at lower cost, because the expensive sources only ever see what is left over.

The technique comes from B2B sales, where it is used to find buyers. Every major explainer on the subject, from Clay to Apollo to ZoomInfo, teaches it that way. Almost nobody writes about it for candidate sourcing, which is a problem, because the requirements are genuinely different and the differences are the parts that decide whether a shortlist is usable.

Why does one data source only find about a third of your candidates?

Because every provider builds its database from overlapping but incomplete sources, and none of them publishes where the gaps are. In a specialist niche, a single provider commonly returns coverage in the region of a third of the real market. You cannot tell from the inside, because the rows that come back look complete.

A list of 200 enriched candidates from one source looks like 200 candidates. It is actually 200 out of a market you never measured, and the 400 you did not see include the ones your competitors have not contacted either.

I tested this on a UK project controls agency. They had 18,195 candidate records on file, built over three and a half years. When I mapped the live market for the specific skill they needed to hire and checked it against their database, 14% of that market was already visible to them and 63% had never been seen by anyone. Their previous sourcing engine had contacted 5,056 people and surfaced none of the qualified ones, because it filtered on job titles and the thing they needed was a skill written in the body of a profile.

What makes a candidate waterfall different from a sales waterfall?

Three things. The personal email matters more than the work email. A qualifying step sits in the middle of the chain, where you read profiles before paying for contact data. And your own database is the first place you look. A sales waterfall optimises for a verified work address. A candidate waterfall that does the same thing produces a list you cannot safely use.

The work inbox is the wrong channel

When you approach a passive candidate you are asking them to consider leaving the employer whose inbox you would be writing to. That inbox is often monitored. Most enrichment providers return work addresses by default, which means a sales-shaped waterfall hands you the one channel least likely to get a reply and most likely to cause a problem for your client.

Qualification happens mid-waterfall, not at the end

In sales, the title is usually the qualifier. A VP of Finance is a VP of Finance. In recruitment, a large share of what you are asked to find is a skill, not a title: a certification, a security clearance, a named modelling tool, a specific methodology. No database holds those as a filter, because they live in free text that the person wrote about themselves.

So you read the profile text before you spend anything on contact data. On the role I ran, a title-led search returned roughly 60% people from a completely different profession who happened to share a word in their job title. A keyword-led search that read profile bodies returned a 44% genuine hit rate.

How do you build a candidate sourcing waterfall?

Six steps, in this order. Each one only runs on what the previous step could not resolve, and nothing paid runs until a candidate has passed the qualification screen. The order is the entire technique, so resist the urge to reshuffle it for convenience.

Here is the job spec I am working on: [paste it, or paste the LinkedIn
profile of someone who would be perfect for the role].

Before you look anywhere outside, search my own ATS and tell me:
1. Who in there genuinely matches this spec, with the line from their record
   that proves it.
2. Who we have already contacted about anything, so we do not approach them
   twice.
3. Anyone who replied positively in the past and was never followed up.

Give me the proof for each one. Do not include anybody you cannot evidence.

Here are the profiles you found: [attach the file].

Read the full text of each one, then score every person against this job
spec. A candidate only passes if they meet ALL of these:
1. [your first hard requirement]
2. [your second]
3. [your third]

For every person you keep, give me a word-for-word quote from their own
profile that proves each requirement. No quote, no row.
Tell me how many you rejected and the most common reason.

The five layers at a glance

The three-layer version, if five is too many

You do not need the full chain to beat what you are doing now. Build this instead, and add the discovery and profile-reading layers once the habit is established.

What does a waterfall enrichment setup cost?

Less than most recruiters expect, because the structure is designed to keep spend off the records that will be rejected. The only fixed per-unit cost in the chain above is the profile scrape. Everything else runs on a plan you already hold, and only touches whatever is left over.

The single biggest cost control is the sequencing rule: never enrich a candidate you are about to reject. Qualification is step 3 and contact data starts at step 4 for exactly that reason. Agencies that enrich first and qualify second pay for the whole market instead of the shortlist.

What stops the list being wrong?

One rule and two review passes. The rule is that every candidate on the list carries a verbatim quote from their own profile proving why they qualify, and no quote means no row. The review passes are two independent checks that run before a human sees the pack, and the tool that built the list is never the one that approves it.

The quote rule does two things at once. Your consultant can verify any line against the public profile in seconds, which is what makes the list defensible in front of a client. And the automation cannot quietly invent a reason, because it has to cite the sentence.

Three failures worth borrowing from mine

1. Evidence has to come from the current role to be claimed as current practice. On one run, 11 of 17 profiles I had recorded as current programme work were actually describing a previous employer. Still valid candidates. Wrong claim in the outreach.
2. Constrain the outreach copy to quoted facts. Left unconstrained, 64% of my first-pass openers contained a factual error, including inventing a programme and describing a risk register as a modelling tool. A deliberately general opener does no damage. A confidently wrong one costs you the relationship.
3. Never append rows to a finished list. I added 18 candidates after quality control had already passed. Those 18 were rejected at 44%, against 6% for everything that went through the full pipeline in order.

One concrete thing the second review pass catches: a shared team mailbox such as risk@theiremployer.com, sitting in the file marked verified. It passes every verification check precisely because it is a shared inbox that accepts anything sent to it. Sending to it means a poaching approach arrives in a team-wide inbox at the candidate's own employer, which costs you the client.

What did this return on a real role?

On one test role the waterfall returned 46 of 50 candidates contactable, 33 of them with a personal email, and surfaced 102 qualified candidates the agency had never seen. That was a UK project controls agency searching three cities for a skill no database holds as a filter. Here is the full output, with the method stated so the numbers can be checked.

Be clear about where the speed actually is, because this gets oversold. The longlist is the part that goes from days to minutes. Searching, reading profiles and filling in contact details across hundreds of people is a few minutes of processing. On a separate build we scraped 606 lawyers from one firm's directory, matched 89% of them to a LinkedIn profile, and verified email addresses for 604 of the 606, in under eight minutes. Profile matching and email discovery run as independent layers, which is why the email count is higher than the LinkedIn match count.

The shortlist still takes a person a few minutes to read and approve. There is no step where that disappears. The scoring gate runs automatically, then someone reads the top of the list, and what used to be a week of scrolling becomes a scored shortlist you sign off over a coffee, because the system did the scoring.

Where to go next

If you want the wider context around this, the candidate sourcing automation guide covers how the engine fits a full recruitment workflow, the sourcing automation playbook covers the campaign layer that sits on top of a waterfall, and the AI tools breakdown covers what else belongs in a recruitment stack in 2026.

Frequently asked questions