WhatsApp PersonalBusiness
Recruitly LogoRecruitly
Engineering

Can AI clean a recruitment database?

Yes, and it is better at most of the job than people are, as long as it only reads and proposes while ordinary code decides what is written. What AI is good and bad at in CRM clean-up, and how the Data Agent keeps it honest.

Ask AI about this

ChatGPT
Perplexity
Grok
Claude
Google AI

Yes, AI can clean a recruitment database, and it is better at most of the job than people are, provided it is only allowed to read and propose while ordinary code decides what is written. AI reads CVs and notes into fields, tidies titles and names, classifies people into your own industries and sectors, and spots addresses that contradict themselves, across thousands of records. It must not be trusted to invent facts it was not shown, to overwrite what your team typed, or to merge people. Recruitly's Data Agent is built on exactly that split.

What can AI do well when cleaning CRM data?

AI is good at the jobs that need reading and judgement but no new facts. A recruitment database is mostly unstructured text: CVs in every layout, notes typed in a hurry, job titles copied from LinkedIn with emojis and slogans attached. Turning that into clean fields was always a job for a person with a lot of time. It is now a job for a model with a person checking a sample.

JobAI isWhy
Reading a CV into fieldsGoodIt reads unstructured text the way a person does, in any layout
Tidying titles, names, placeholders, datesGoodPattern work across thousands of records without getting bored
Mapping industries and sectors to your listsGoodIt matches meaning, not spelling, against a fixed list you own
Spotting an address that contradicts itselfGoodIt notices a city in one country and a country that is another
Knowing where someone works todayBad on its ownIt only knows what it is shown; the answer has to come from a live source such as LinkedIn
Finding a phone numberBadA model that does not know will make one up; that is a data lookup, not AI
Deciding two records are one personRiskyA wrong merge joins two people's histories; it should suggest, a person decides

The pattern in that table is the whole answer. AI is strong where the facts are already on the record and need reading, and weak where the facts have to come from somewhere else. A cleaning system that respects the difference works. One that does not produces confident nonsense at scale.

How does the Data Agent stop AI making things up?

By never letting the model write anything. Each record goes through four steps, and the model is only in one of them.

CODEAI1. Gatherrecord, CV, notes, LinkedIn2. Read and proposevalue, source, confidence3. Enforce the rulesblanks only, no low confidence4. Preview or writeevery change keptFilling writes to empty fields onlyNever blank an email, phone or LinkedIn linkDrop anything the AI is not sure ofKeep the tags the team addedPostcode only when it is written down
The AI never writes to a record. It reads and proposes, and code it cannot argue with decides what is applied. A rule in code holds on the ten-thousandth record; a rule in a prompt holds most of the time.
  1. Code gathers. The record, its CV and notes, and, where ticked, the live LinkedIn profile and the company's own record.
  2. The AI reads and proposes. For each field it returns a value, the source it came from (the CV, a note, LinkedIn, the company) and how sure it is, with a one-line reason.
  3. Code enforces the rules. Filled values for fields that already have one are not applied, and the tidy-ups you tick only reformat what is there. Values the model is not sure of are recorded and left out. The address guard puts back any address field the model changed, after the merge, where no prompt can talk it round. A clean-up can never empty an email, phone or LinkedIn link the record had. Tags are read from the record and sent back with any new ones, so a run cannot wipe what your team added.
  4. Preview or write. In a preview nothing is written. In a real job every change is stored per record with its source and confidence, so anyone can see why a value is there.

Some rules are about what the model may assume. A postcode is filled only when it is written somewhere on the record, never guessed. A phone number's dial code can suggest a country, but only at low confidence, so it is recorded and not applied. A contact's location taken from a company with several offices is low confidence for the same reason. Nationality is filled only when the CV states it.

Why put the rules in code instead of the AI's instructions?

Because instructions are followed most of the time, and most of the time is not good enough for ten thousand records. A model told "never change a filled field" will still change a few, usually where the record looks wrong to it, which is exactly where a person typed something on purpose. There are quieter traps too. A model's answer can simply leave a field out, and read naively that looks exactly like the field was deleted. So the response is laid over the original record, an omitted field keeps its value, and contact fields cannot be emptied by a clean-up at all.

The general rule is that code gathers and the AI judges, and code has the last word. Is AI accurate enough to change your CRM? goes deeper into accuracy, and Can you trust AI to update your CRM records? into trust.

How do I check what the AI will change before it runs?

Press Preview. The Data Agent runs the whole pipeline on up to five records, writes nothing, charges nothing, and shows each proposed change card by card, with where the value came from and how sure it is. Paid lookups such as emails and phones are listed but not run. If a rule in your data trips it up, you see it on five records, not five thousand.

After a job, a super user can open it in AI Hub and see every record it worked on, and each change before and after.

Who should approve an AI clean-up?

A super user. A job started by anyone else waits until a super user starts it in AI Hub, and nothing is charged before then. The cost is estimated before the job starts and checked against your agency's monthly cap and balance; the estimate is reserved, and anything not used goes back when the job ends. If the same records already have a job running, Recruitly warns that you would pay for the same work twice.

Can AI keep the database clean, not just clean it once?

Yes, on a schedule. A super user can set a Data Agent job on a saved search to repeat every day, week or month, and each run cleans the records added since the last one. The same rules apply on every run. Alongside it, job change alerts read LinkedIn for the people who matter most and Email List Clean-up checks campaign lists, each with its own safeguards, such as holding back a batch of bad email results that looks like a fault at one provider instead of blacklisting it. Is there a tool that cleans my CRM automatically? sets out the whole system.

What should never be left to AI?

Three things: merging records, overwriting a value a person entered, and deleting people. In each case a wrong answer is expensive and hard to notice. Recruitly finds likely duplicates and leaves the merge to a person, fills blanks and leaves typed values alone, and leaves deletion to people with a 90-day recovery. The recruiter's judgement about who someone is and what they told you on a call is still the most valuable data in the building, and AI cleaning exists to protect it, not to replace it.

Can AI read scanned or image CVs?

Yes. Recruitly's scanned CV parser reads CVs that are images or scans, so their contents can fill the record like any other CV.

How much does AI database clean-up cost?

$0.005 a record for the clean-up, whatever you tick, plus $0.005 for each LinkedIn refresh. Cleaning 10,000 candidates costs $50. Every agency gets $10 of Data Agent credit on first use. See the pricing page.

AIdata cleaningData AgentguardrailsCRM
Recruitment software for agencies

The product this came out of

Nineteen modules on one record: sourcing, screening, campaigns, calls, e-signature and billing, without a second system to keep in step. Free to start, no card, no call.