Step by step: which parts of lead enrichment are lookups and rules, where AI reads company websites and matches duplicate records, and which decisions about consent and key accounts stay with a person.
Partly, and most of it is not AI. Lead enrichment is mostly a lookup: a rule takes the email domain, asks a data provider for company size, industry and location, and writes the result into your CRM. AI earns its place where the data is messy, by reading a company website to see what the business really does and by matching records that look different but describe the same company. A person stays responsible for the legal basis for using the data and for the key accounts where a wrong guess costs a deal. Below is the split per step, what you need in place first and a small pilot you can run on one lead source.
What each step needs
01● Standard automation
Capture the lead and clean the basics
Enrichment starts the moment a lead arrives from a form, a chat, an event list or an import. A rule trims whitespace, lowercases the email, splits the domain from the address, flags free mailbox domains such as gmail.com and normalises country and phone formats. This is plain logic and it should be, because everything later depends on a clean domain. Do not ask a language model to guess a company from a personal email address; if the domain is a free mailbox, the rule marks the lead as "company unknown" and moves on to the next steps that need a human or a form question.
⛨Works when every source writes to the same lead fields. Keep a raw copy of what the lead typed so you can always see what was changed.
02● Standard automation
Look up company and contact data with a provider
The reliable core of enrichment is an API call to a data provider or to your own customer database. Given the domain, a rule requests company name, size band, industry code, headquarters and website, and for contacts a job title and seniority where the provider has them. This is automation, not AI: the same input should give the same output, and every field should carry its source and the date it was fetched. If the provider returns nothing or a low-confidence match, the lead is not guessed at; it simply continues with fewer fields.
⛨Works when you have a data provider or internal source you trust for your market, and when fields are stored with source and date.
03● AI candidate
Read the company website for context
Provider data tells you the size and a code, not what the company sells or whether it fits you. A language model can read the homepage and a product page and answer specific, checkable questions: what does this company sell, to whom, in which countries, and does it mention the problem you solve. The output should be a short structured note with a quote or page reference for each claim, plus a confidence level. This is where AI adds something a lookup cannot, but only if you ask narrow questions and keep the source text next to the answer.
⛨Works when the questions are fixed and tied to your ideal customer profile, and when the model may only use the pages it was given.
04● AI candidate
Match duplicates and merge conflicting records
The same company shows up as "Acme Ltd", "ACME Limited" and "acme.com" across your CRM, your billing tool and an event list. Exact rules catch the easy cases through the domain; AI helps with the near matches by comparing names, addresses, websites and contacts and proposing which records belong together, with a score. Merging is not done on a score alone. High-confidence matches can be merged with an undo log, while medium-confidence ones go to a short review queue so one wrong merge does not wipe out an account history.
⛨Works when you keep a merge log, can undo a merge, and review a sample of automatic merges every week.
05● Standard automation
Decide which fields are written to the CRM
Not every enriched value should overwrite what is already there. A rule decides per field: never replace a value a person typed, only fill empty fields, replace provider data only when the new value is newer and from a better source, and never write a value below your confidence threshold. The same rule set writes the enrichment date and source to a field history. This step is deliberately dull, and that is what keeps the CRM trustworthy. Salespeople stop trusting enriched fields the first time one overwrites something they knew was right.
⛨Works when field ownership is written down: which fields belong to people, which to the provider and which to the model.
06● Keep human
Check the legal basis and the data you may use
Whether you may collect, store and use this data for outreach is a legal and policy question, not a data question. In Ireland the GDPR and the ePrivacy Regulations apply, and the Data Protection Commission explains what you need before you store data about people and contact them for marketing. The person responsible decides which sources are allowed, how long data is kept, how people are told about it and how opt-outs are handled, and writes that into rules the automation follows. AI can help draft the notice or summarise a policy, but it does not decide what is lawful, and an enrichment tool should never widen its sources without that sign-off.
⛨Applies to every market you contact. Review the rules when you add a new data source, a new country or a new use of the data.
07● Human review
Review key accounts and unclear leads
A small share of leads deserves a person: large target accounts, leads where the provider and the website disagree, and anything the automation marked as low confidence. A salesperson or an operations owner looks at the enriched record, fixes what is wrong and, ideally, leaves a short reason. Those corrections are valuable; they show you where the sources fail and which questions to ask the model differently. Keep this queue small and fast, and review it daily, or the enrichment will look finished while the bad records pile up.
⛨Works when the review queue has an owner, a daily time and a clear rule for what lands in it.
A sensible first experiment
Pick one lead source, such as your website demo form, and last month's leads from it. Run the cleaning, the lookup and the website reading on those leads without writing anything to the CRM, then compare the output with what your team already knows about a sample of fifty records. Measure three numbers: the share of leads where each field was filled, the share of filled fields that were correct, and the minutes a salesperson needs to prepare a first contact before and after. If correctness is high, switch on writing for empty fields only, keep the review queue for low confidence, and extend to the next source a month later.
The trap to avoid
The classic mistake is treating enrichment as a data purchase instead of a process. Teams buy a large data set, let a model fill every field, and overwrite what salespeople typed, until nobody trusts the CRM. A second mistake is asking a model for facts it can only guess, such as a revenue figure or a personal phone number, and storing the answer as if it were a record. Let rules fetch facts from sources, let AI read and classify what is in front of it, keep a source behind every field, and never enrich beyond what your privacy rules allow.
Questions teams ask
Is AI better than a data provider for lead enrichment?
They do different jobs. A data provider gives structured facts such as size, industry and location, with a source and a date. AI is useful for what a provider does not hold, like reading a website to see what a company sells, or deciding whether two records describe the same company. Used for facts it cannot know, a model will guess, so use the provider for facts and the model for reading and matching, with the source text kept next to every answer.
Can AI find a prospect's email address or phone number?
A model cannot look up a private contact detail; at best it guesses a pattern, and guessed addresses bounce and can harm your sender reputation. Use contact data from a source with a clear legal basis, such as a provider you have vetted, the lead's own form entry or your existing customers, and let rules validate the address format. Whether you may contact the person at all depends on the rules in your market, so settle that before enrichment, not after.
Which fields should enrichment never overwrite?
Anything a person entered or confirmed: owner, deal stage, notes, a contact's own stated job title and any opt-out or consent status. Write enriched values only into empty fields or into separate enrichment fields, and keep a history with source and date. If a model suggests a change to a protected field, it goes to the review queue instead of the record.
How do I start lead enrichment in a small organisation?
Start with the one decision it supports, for example which new leads a salesperson should call first. Choose the three or four fields that decision needs, such as company size, industry and country, fill them with rules from one source, and add website reading only for the leads that remain unclear. A spreadsheet export and a single automation are enough for the first month; the aim is to learn which fields change what your team does.
Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.