Waterfall data enrichment fills an empty field by trying sources in order and stopping at the first one that answers with a value you can trust. It exists because no single provider covers every company, every person and every field, and because the second-best source for a Berlin fintech is often the best source for a Dallas roofing company. Run well, a waterfall roughly doubles the fill rate a single source produces and records where every value came from. Run badly, it fills fields with confident guesses from whichever source happened to be first.
This guide is the whole method: what a waterfall is, why single sources fail, how the run proceeds field by field, how to order sources, the provenance and conflict rules that keep the output honest, the verification pass that has to happen before anything is delivered, and what changes when an agent plans the waterfall per row instead of an operator wiring it once. It reflects how the enrichment agent actually works a list.
What waterfall enrichment is
Take one company and one empty field, say employee count. A single-source setup asks one provider and accepts whatever comes back, including nothing. A waterfall asks the first-ranked source for that field; if it returns a value above the confidence bar, the field fills and the run stops. If it returns nothing, or a value below the bar, the run moves to the second source, then the third, until the field fills or the sources run out. Every field in the row runs its own waterfall, which is why the honest unit of enrichment is the field and never the row. The short definition lives in what is waterfall enrichment; this guide is the operating manual.
The pattern generalizes beyond contact fields. Resolving a company name to its domain is a waterfall. Filling firmographics, technographics, funding, news and the buying committee on a CRM segment is a waterfall per family. Finding a verified email for a person is a waterfall whose last stage is verification. What the pattern buys is coverage without a single point of failure, and what it costs is the discipline to order, record and verify.
Why one source is never enough
Every data provider builds its coverage from somewhere - crawled sites, registries, partnerships, contributed address books, job boards, funding databases - and every source of coverage has a shape. One provider is strong on North American SaaS and thin in DACH manufacturing. Another has the best mobile numbers in sales roles and almost none in engineering. A third refreshes technographics weekly and firmographics annually. On a list that spans segments, the strongest single source still leaves a third to a half of the fields empty, and the empty share is never random: it concentrates in exactly the segments that were hard to find in the first place.
There is a second failure that single sources hide: the confident wrong value. A provider that returns an employee count for every company is returning a model's estimate for many of them. Without a second source to disagree, the estimate lands in the CRM as a fact. Waterfalls surface disagreement, and disagreement is information - it tells you which values deserve a lower confidence and which segments need a different order.
2xthe fill-rate gain a well-ordered waterfall typically delivers over its best single source
How a waterfall runs, field by field
| Field family | Confidence bar | Decay | What ends the waterfall |
|---|---|---|---|
| Identity (name to domain, person to company) | Exact match, or two sources agreeing | Slow | A resolved domain that the firmographic pass confirms |
| Firmographics (size, industry, revenue band, HQ) | Observed beats estimated; estimates need agreement | Quarterly | An observed value, or two estimates within tolerance |
| Technographics (stack, adoptions, removals) | Detected on the live site or a dated observation | Weeks to months | A dated detection; the date travels with the value |
| Funding and news | A primary announcement or filing | Event-driven | A dated event with its source |
| People (title, seniority, department) | A current role from a profile or an announcement | Months | A role confirmed at the resolved company |
| Contact (email, phone) | Verified deliverable or connected | Months | Verification, never a pattern guess |
The run has an order of its own. Identity first, because every other family keys on the resolved domain and a wrong domain poisons the row. Firmographics second, because they decide whether the row is worth spending more on: a company that turns out to be outside the size band exits the waterfall here, and the credits that would have gone into finding its people stay unspent. Then technographics, funding and news in whichever order the objective needs. People and contacts last, because they are the most expensive fields and the ones that need verification, and there is no point verifying an email at a company you have just discovered is a customer.
Ordering the sources: accuracy, cost, freshness
Three properties decide the order, and the weights are yours. Accuracy first: a source that is right 95% of the time in your segment goes ahead of one that is right 80% of the time, regardless of price, because a wrong value costs more downstream than any lookup. Freshness second, and it matters more for some families than others: a technographic observation from last week beats one from last year even at higher cost; an HQ address from last year is fine. Cost third, as the tie-breaker among sources that clear the accuracy and freshness bars - which, on a list of thousands of rows, still moves the bill materially.
The order should differ per segment, because accuracy does. The practical method is to run a sample of a few hundred rows through every source, compare against ground truth you trust (closed-won accounts, verified contacts), and let the measured accuracy per segment set the order - then re-measure quarterly, because sources improve and decay. The company and person data guide covers what to demand from each source's data contract so the comparison is fair.
Provenance and the conflict rules
A waterfall without provenance is a guessing machine with extra steps. Every value written should carry three things: the source that produced it, the date it was observed, and the confidence the run assigned. That is what lets a rep reading "employee count: 240" see that it came from an observed source in August rather than an estimate in 2023, and what lets the next run decide whether the value is fresh enough to keep.
The same rules are what make CRM hygiene safe to run as a standing mission: an enrichment pass that respects them can run weekly against a live CRM without a rep ever discovering that a value they typed was replaced by a vendor's guess.
Verification before delivery
A filled field is a candidate, and a finished row has passed four checks. The contact fields are verified - email deliverability at the mailbox level and phone connectivity where the source supports it - because an unverified address in a sequence is a reputation cost paid by every future send, the mechanics of which are in email verification and deliverability for outbound. The row is deduplicated against the list, the CRM and the history of accounts already worked, on confidence tiers. Customers and competitors are suppressed. And the enriched row is re-tested against the ICP, because the whole point of enrichment was to learn things you did not know, and some of what you learn is that the company never belonged on the list.
- Identity resolved and confirmed by firmographics
- Every requested field filled or explicitly marked unfillable, with provenance
- Emails verified deliverable; phones verified where possible
- Deduplicated against the list, the CRM and worked-account history
- Customers and competitors suppressed
- Still inside the ICP after enrichment
What an agent changes
A waterfall wired by an operator is a fixed pipe: the same source order for every row, tuned once, drifting as sources change. An agent plans the waterfall per row from what it already knows. It has resolved the domain and read the firmographics before it chooses the contact sources, so it can pick the order that measured best for this segment and skip the sources that never answer in it. It stops the moment the confidence bar is met instead of running the whole chain. When a source returns a value that contradicts an observation, it lowers confidence and tries another rather than writing the contradiction. And it holds the budget: a row whose fields are not filling exits the waterfall before it consumes the credits meant for the rows that are.
That is the practical meaning of the agentic workflow anatomy applied to data: objective in, plan per row, sources chosen at runtime, verification before delivery. The same agent, given a list of ten rows or ten thousand, runs the same method at the same standard, which is what makes enrichment schedulable as a standing mission rather than a quarterly project.
The cost model: pay when a source answers
Waterfalls have a cost failure mode of their own: paying for every attempt, including the misses. A row that runs through five sources and fills on the fifth should cost one answer, not five lookups, and a row that never fills should cost close to nothing. The pricing question to put to any enrichment platform is therefore simple - is a lookup that returns nothing charged? - and the operating question that follows is whether the ceiling is enforced before the spend. A waterfall that can run away on a bad list is a waterfall nobody schedules, which is why the budget mechanics belong in the architecture rather than in an alert.
How AstroFabric does it
The enrichment agent takes any list, CSV or CRM segment and an objective - which fields, which confidence bar, where to deliver - and runs the waterfall field by field: names resolved to domains, firmographics, technographics, funding, news and the buying committee filled source by source until the field is filled, with the source, date and confidence recorded on every value. Observed values are never overwritten with inferred ones; conflicts are flagged for a person. The verification agent runs the delivery bar - bulk email and phone verification, dedupe, suppression of customers and competitors, the ICP re-check - before anything lands.
Enrichment is only charged when a source answers, every run is bounded by a credit ceiling enforced before the spend, and every write into a CRM, sheet or outreach tool passes an approval gate. The finished list is persistent and refreshable, so the same waterfall re-runs on a schedule and reports what changed. It runs from the console, the REST API, MCP, the CLI, Slack, Telegram or email, and the AI agents for waterfall data enrichment page maps the full motion.
Go deeper in this cluster
- Company and person data for AI agents: the complete guide - What an agent needs from company and person data that a human operator never asked for: typed fields, provenance, freshness and cost per call. The company families, the person families, identity resolution, the data contract, delivery and the compliance rails.
- What is technographic data? - Technographic data is the record of which technologies a company uses - detected, dated and tracked over time. What it observes, how it is collected, why the changes matter more than the snapshot, and how it is used in prospecting and enrichment.
- What is firmographic data? - Firmographic data describes what a company is - size, industry, revenue, location, ownership, age. The fields, where they come from, why observed values beat estimates, and how firmographics anchor ICP filters, fit scores and every waterfall that follows.
- What is waterfall enrichment? - Waterfall enrichment fills an empty field by trying data sources in order and stopping at the first trustworthy answer. The definition, a worked example, why it beats a single source, and the two rules - provenance and pay-on-answer - that keep it honest.
- AstroFabric vs Ocean.io: lookalike discovery or the whole chain after it - Ocean.io is the specialist for lookalike account discovery: paste your best customers, get the companies that resemble them. AstroFabric runs the lookalike step and everything after it - people, verification, enrichment, signals, audiences - as agent missions. The honest map, with September 2026 pricing.
Frequently asked questions
What is waterfall data enrichment?
A method that fills an empty field by trying data sources in a ranked order and stopping at the first value that clears a confidence bar. Each field runs its own waterfall, so a row is enriched field by field with the source, date and confidence recorded on every value.
How much does a waterfall improve fill rates?
Roughly double the best single source on a mixed list, with the gain concentrated in the segments a single source covers worst. The exact number depends on your market and on how well the source order was measured against ground truth.
How should sources be ordered?
Accuracy first, freshness second, cost third - measured per segment against ground truth you trust, and re-measured quarterly. The best source for one market is often the third-best for another, so one global order leaves value on the table.
What is provenance and why does it matter?
The source, date and confidence attached to every value. It lets a person judge a field at a glance, lets the next run decide whether a value is fresh enough to keep, and makes the conflict rules enforceable: observed beats inferred, agreement raises confidence, freshness wins ties.
Does enrichment include verification?
It should end with it. A filled email is a candidate until it is verified deliverable, the row is deduplicated against what exists, customers and competitors are suppressed, and the enriched row is re-checked against the ICP. Delivering before those checks turns a list into a bounce report.
What does an AI agent add to a waterfall?
A plan per row instead of a fixed pipe: it chooses the next source from what it already knows about the company, stops when the bar is met, flags contradictions instead of writing them, and exits rows that are not filling before they consume the budget meant for rows that are.
Sources
- Gartner - data quality research (the cost of wrong values downstream)
- Google - Email sender guidelines (why verification gates delivery)
- European Commission - data protection rules that govern person data processing
Every playbook on this blog ships as a runnable mission.
Open a workspace and the playbook library is waiting - describe the outcome and the agents carry it end to end, on your plan's monthly credits.