2records
Duplicate contacts in a CRM: why a small business gets them and how to fix them
One invented heating and cooling company finds a customer in its CRM twice, three days after an email asked her whether she was still thinking of buying. Why duplicate contacts are never carelessness. Why the published method for matching records has three answers, not two. The four ways a sync quietly damages a record. And a twenty-minute check you can run on your own CRM today.
Duplicate contacts in a CRM come from typos, from repeated imports and, the kind careful typing never stops, from two systems that are each right about the same customer. This page gives you, free, the fix we build for small businesses: a rule about identity, a written rule for which side wins, and one person who settles the doubtful pairs before anything is merged. Below, one invented heating and cooling company finds a customer in its CRM twice, and you will see why the two obvious fixes, more careful typing and a tool that merges everything on its own, are the wrong first build. It is the method we use when we build a two-way CRM sync for a small business, from a heating company to a real estate team, and you can run the first half of it on your own. By the end you will know how to check your own CRM for duplicates, which questions decide whether a sync damages your records, and how to set up a review queue a person can keep.
Dana's heating and cooling company (invented). Dana owns a heating and cooling company with four people: herself, Luis in the office and two technicians on the road. Five things can create a contact in her CRM: the quote form on the website, the phone system, the booking app, the invoicing program and the sheet of names Luis types up after a home show. Dana, Luis, their customers and every figure about the company are made up; their records run through this page as one worked example of a real method.
Dana's heating and cooling company (invented). In March a woman asked the website for a quote on a new furnace. She typed her name the way she says it out loud, so the record says Kathy Brown, with a personal email and a mobile number typed with one thumb. In June she rang the office because her air conditioner had stopped. Luis did everything right. He took the details and typed them in properly, with the name spelled the way it appears on her bills: Katherine Brown. A work email this time, and the mobile she actually answers. In August she signed for a new heat pump. In September the technicians installed it. And on the Tuesday three days after the install, an automated email went out asking Kathy whether she was still thinking about replacing her furnace this year.
Nothing broke. Every part of that ran the way it was built to run. The website did its job, Luis did his, and the campaign was pointed at exactly the segment it was meant for: people who asked for a furnace quote and have not booked. The system was working. It did not know that the two women in it were one woman. A real estate team meets the same duplicate contacts with a deal attached, in an invented example from that trade: a seller fills in a valuation form, then rings the office, and her listing and contract go on one record while the nurture campaign keeps reading the other, which is how she gets asked whether she is still thinking of selling three days before her closing.
Everything on this page is free to use: the method, the twenty-minute check and the questions to ask whoever builds your sync. What we sell is building the sync with you and keeping it running.
In short
- Duplicate contacts in a CRM are not a discipline problem. Every route into a business creates a record. Each one is separately correct about the customer in front of it. And nothing in the building has the job of noticing that two of them are the same person.
- Deciding whether two records are one person is a studied problem with a published answer. The answer has three outcomes rather than two: a match, a nonmatch, and a middle band held for a person to look at. The thresholds either side of that band are set from the two error rates you are willing to accept. And you cannot set both of them to zero.
- Most of the damage a sync does is not a failure. It is a design decision nobody made out loud. Whether an update replaces the whole record or changes named fields. What happens when the same message arrives twice. And which side wins when both of them changed.
Why does a CRM end up with duplicate contacts?
A CRM ends up with duplicate contacts because whichever system meets a customer first creates a record, and none of those systems can see the others. Your website makes a contact when a form is submitted. Your phone system makes one when a number it does not recognise rings in. A booking app makes one when a new name is booked, and a list someone types up on Monday morning makes one for every line. Every one of those routes is doing the correct thing.
That is the whole mechanism, and it matters because the explanation people reach for first is carelessness. A typo can make a duplicate too, and training helps with typos. The duplicates that survive a careful office come from two systems being separately right about the same human being, at two different moments, using whatever she told each of them at the time. She was in a hurry in March and typed the short version of her name. In June she was talking to a person and gave the formal version, because that is what you give a person who is writing something down.
The reason this is worth understanding rather than just fixing is that it tells you what the fix has to be. If the problem were carelessness, the answer would be training. Where nobody typed anything wrong, training has nothing to correct. The answer is a rule about identity. It lives outside all of those systems, and it is applied every time any of them creates a record.
Dana's heating and cooling company (invented). Dana's first thought is the obvious fix: tell Luis to search the CRM before he creates anyone. Luis had searched. In June he looked up the mobile number Katherine gave him, and the March record held a different one; he looked up her work email, and the March record held her personal one. A careful person, doing the careful thing, found nothing, because nothing in the CRM matched what she told him that day.
The two records
Both of them are correct. That is the problem.
Created in March, by your website
Kathy Brown. A personal email address. A mobile number typed into a phone with one thumb, on a quote form, late in the evening. Source: furnace quote request. No owner, because nobody had picked it up yet. Everything on it is exactly what she typed, which is what a form is for.
Created in June, by whoever answered
Katherine Brown. A work email. A different mobile, the one she actually answers. Source: inbound call. Owner: the person who took it. Everything on it is exactly what she said out loud to someone who was writing it down. That is a different question, and it gets a different answer.
Neither of them is the wrong one
This is the part that changes what the fix has to be. There is no mistake in either record, and no training that would have prevented it. Both systems did the correct thing with the information they were given. The duplicate is not an error. It is the shape two correct systems make when neither can see the other.
And only one of them knows she bought
The August quote she signed and the September install were entered against the June record, because that is the one the office was working from. The March record still says she is thinking about it. The campaign that emailed her three days after the install was pointed at exactly the right segment. There is nothing in it that failed.
Your turn. List every route that can create a contact in your CRM: each form, each phone line, each app, each sheet somebody types up.
In a real estate office, specifically
A real estate team has the same problem with more doors: a seller's valuation form on the website, a buyer's enquiry from a listing portal, an open house sign-in sheet typed up on Monday, a call to the office. Each creates a contact, and the deal, first the listing and then the contract, is entered against whichever record the agent opened. The other record stays in the nurture campaign, which is how the seller in the invented example at the top of this page gets asked whether she is still thinking of selling three days before her closing. To dedupe contacts in a real estate CRM the method on this page is the same: one identity rule, a merge process with a person in the middle, and every deal checked against the record that survives. Whether that person is the agent, the office manager or a virtual assistant doing the duplicate lead cleanup matters less than that one named person owns it.
How does a computer decide two contacts are the same person?
A computer decides whether two contacts are the same person by comparing their fields and measuring how similar the text is: it has two strings and a way of asking how close they are, not the recognition you used the moment you read Kathy and Katherine.
This is the part of the subject that has been studied, and it has a name: record linkage. The clearest published overview is by William Winkler of the United States Census Bureau, written for the Bureau's own research report series. The Bureau matches a whole census against a follow-up survey, a very large matching job that Winkler describes below, and the ideas behind it go back to Howard Newcombe in the 1950s. One paragraph in it is worth reading now.
Winkler is describing what counts as a "typographical error" between two records that belong to the same person. He gives four examples of first name pairs: Bill and William, Mr and William, William and James, William and Willam. Look at the third one. William and James are not a misspelling of each other; they are different names. The paper counts the pair anyway, because its definition is any difference in the way corresponding fields are written between two records that are in fact a match. Middle names get used as first names. People go by their second name all their lives. The record is not wrong. It just does not look like the other one.
How often does that happen? Winkler cites his own earlier measurement: even high quality files might contain more than 20 percent error in first name pairs and more than 10 percent error in last name pairs among pairs that are true matches (Winkler, Census Bureau research report). Read that twice, because it is the number that decides everything downstream, and read the word "might" in it as well: it says what a careful file can look like, not what every file does. In a file someone had already taken care over, more than one pair in five that describe the same person can disagree about her first name.
Can a CRM match duplicate contacts by name alone?
A CRM cannot match duplicate contacts by name alone, because names disagree when they should agree and agree when they mean nothing: 1,437,026 people in the 2010 Census carried the surname Brown.
The Census Bureau publishes the count of every surname that occurred at least 100 times in the 2010 Census: 162,253 of them (Census Bureau, frequently occurring surnames). Brown was carried by 1,437,026 people. So a matching rule that treats a surname as evidence is not wrong, exactly. But a surname is worth almost nothing on its own, and it is worth a wildly different amount depending on which surname it is. Winkler makes the same point with two examples of his own: a relatively rarer last name string such as Zabrinsky has more distinguishing power than a string such as Smith.
The evidence
How many people in the United States carried each of these surnames
Counts of people by surname from the 2010 Census returns, as published by the Census Bureau. The five drawn here are the five commonest of the 162,253 surnames that occurred at least a hundred times. The fourth is the surname in the story at the top of this article. Source: United States Census Bureau, Frequently Occurring Surnames from the 2010 Census.
This is the whole country in 2010, and your database is not the whole country. So treat the bars as an order of magnitude rather than as a rate that applies to your file. Two limits matter more than the age. Surnames are spread very unevenly by place. A name that is unremarkable nationally can be the only one in a village, and truly useful evidence there. And a matching rule never uses a surname on its own, which is exactly the point. The chart is here to show what a surname is worth by itself, which is close to nothing. That is why everything else on the record has to do the work.
There is a second, duller constraint that shapes every real matching system. You cannot compare every record against every other record. Ten thousand contacts is just under fifty million pairs, and the overwhelming majority of them are two people who have nothing to do with each other. The standard answer, which Winkler credits to Newcombe in 1962, is called blocking: only compare pairs that already agree on something, such as a surname or a date of birth. It is a good answer with a cost built into it. Any true match whose blocking field is wrong on one side will never be looked at, because the two records were never in the same pile.
Dana's heating and cooling company (invented). Dana's CRM holds about 3,000 contacts, which makes 3,000 times 2,999 divided by two, about four and a half million possible pairs, and six of those contacts are named Brown. Blocking on the surname puts Kathy and Katherine in the same pile of six, so they get compared. Blocking on the mobile number would have kept them apart for good, because the two records hold different numbers.
Your turn. Search your CRM for your commonest surname and count how many of the records that come back are people you could tell apart without the first name.

What is the published method for matching duplicate records?
The published method for matching duplicate records is the model Fellegi and Sunter set out in 1969, and it gives three answers, not two: a match, a possible match held for a person, and a nonmatch. It formalised ideas Newcombe had introduced ten years earlier, and Winkler's overview restates it.
You take a pair of records and look at the pattern of what agrees and what does not. Same last name, different first name, same street number, no email on one side. Then you ask a ratio: how likely is that exact pattern among pairs that really are the same person, against how likely it is among pairs that really are not? A high ratio means the agreements are the kind only matches produce. A low one means they are the kind strangers produce by coincidence.
Then comes the decision rule, quoted here almost exactly as Winkler writes it. If the ratio is above an upper threshold, designate the pair a match. If it is between the two thresholds, designate it a possible match and hold it for clerical review. If it is below the lower threshold, designate it a nonmatch.
The decision rule
Three outcomes, and the middle one is a person.
Above the line: merge them
The pattern of agreements is the kind only the same person produces. The two records become one, and the second history is folded into the first. When this one is wrong you have fused two people. Then someone opens a contact expecting one conversation and finds a stranger's in it. That is the expensive mistake, and it is the one an over-confident setting produces.
In the middle: ask somebody
Enough agrees to be interesting, and not enough to be sure. The pair waits in a queue for a human being who looks at both and decides. A person with context that is in neither row can decide at a glance. This band is the price of the other two being right. And every product that advertises full automation has quietly abolished it.
Below the line: leave them alone
Two people who happen to share a surname, or the same person with nothing in common between the two rows. When this one is wrong she stays in your database twice, and every campaign counts her twice. It is the cheap mistake, and it is the one you already have. That is precisely why the pressure is always to move the line and merge more.
Dana's heating and cooling company (invented). Dana starts without scores, with a rule anyone in the office can apply. A pair that agrees on the email or the mobile, with nothing that contradicts it, is above the line. A pair that agrees on the surname and the street and nothing else is in the middle. A pair that agrees on the surname alone is below the line. Kathy and Katherine land in the middle. A surname and a street address that agree are the kind of agreement a true match produces; a first name in two forms, two emails and two mobiles pull the other way. The rule does not merge them and does not ignore them. It holds the pair for a person. A shared surname and street are not proof on their own, because a household can share both, so Luis, who took the June call, rings the number Katherine answers and asks. She sent the March form, and she tells him which mobile to keep as the main one.
Your turn. For one pair you found, write down which fields agree and which do not, and decide which of the three answers you would give it.
Why does matching duplicate contacts need a person in the middle?
Matching duplicate contacts needs a person in the middle because no setting of the two thresholds makes both kinds of error zero, and the pairs between them are the price of keeping both low. A product that either finds a duplicate or does not is describing two outcomes. The model that works has three, and the middle one is a queue of pairs that a human being looks at. Winkler's own name for that band is the no-decision region.
The reason it exists is the sentence underneath the rule, and it is the honest centre of this page. The two thresholds are set from error bounds you choose in advance, on false matches and on false nonmatches. You get to pick how often the system is allowed to merge two people who are not the same person. You get to pick how often it is allowed to leave one person sitting in the database twice. You do not get to pick zero for both. With the middle band held where it is, lowering one of those numbers raises the other. The only way to lower both is to widen the middle band, and the middle band is a person's afternoon. Better evidence on each record, a second field that agrees, moves pairs out of the band; it does not remove the trade at the line.
That is not a limitation of the software you were quoted. It is a property of the problem, published in 1969. A single percentage, given in answer to "how accurate is your deduplication", leaves out the other half of the trade.
Winkler records what the computerised procedures did in what he calls a very large 1990 Decennial Census application: they reduced the need for clerks and field follow-up from an estimated 3,000 individuals over 3 months to 200 individuals over 6 weeks (Winkler, Census Bureau research report).
The evidence
People needed to finish the matching in the 1990 Decennial Census
The figures Winkler records for the 1990 Decennial Census matching application: an estimated 3,000 individuals over three months before the computerised procedures, and 200 individuals over six weeks after them. The two bars cover different lengths of time, and a bar chart cannot show that. So both periods are written into the labels. Source: William E. Winkler, Overview of Record Linkage and Current Research Directions, U.S. Census Bureau Research Report Series (Statistics #2006-2), 2006.
This was a national census in 1990, not a small business database. It was matched against a survey designed to be matched against it. That is far better raw material than anything in a CRM. Read the second bar rather than the first. Two hundred people were still needed at the end. The reason the paper gives is not that the matching was not good enough. It is that first name and age were simply missing from some of the forms. There is no procedure for a blank field. The transferable finding is the shape rather than the ratio. The shape is that automation takes the volume out and leaves the ambiguity in.
And then he says why the 200 were still needed, which is the part worth quoting. Both first name and age were missing from a small proportion of the Census forms and the survey forms matched against them (the same source). Not wrong. Missing. An algorithm cannot compare a field that is blank; it can only lean harder on the fields that are left, and where those are not enough, a person decides. The matching job Winkler describes, with the Census Bureau's procedures behind it, still ended with two hundred people deciding.
Dana's heating and cooling company (invented). Dana's second thought is the other obvious fix: a tool that merges every duplicate on its own. A rule loose enough to merge Kathy and Katherine without asking (the same surname, the same address, first names that begin with the same letter) would also merge Kenneth and Kevin Brown, a father and son who share a duplex across town. Their two service histories would become one, and the next technician at the door would read the wrong man's furnace. Dana keeps the middle band, names Luis as the person who settles it, and gives it a slot: Friday after lunch.
Your turn. Name the person who would settle the doubtful pairs in your business, and the slot in their week when they would do it.
You can decide how often it merges two people who are not the same person, and you can decide how often it leaves one person in there twice. You cannot decide both, and the difference between them is somebody's afternoon.
What is two-way CRM sync made of?
Two-way CRM sync is made of three decisions: which record an update belongs to, which fields move between the systems, and which direction each field is written in. Which record is this. It is settled before anything else can happen. A piece of information cannot be written to a contact until someone has decided which contact it belongs to.
Which fields move. Your phone system knows a call happened, its length and its outcome. Your CRM has somewhere to put a call, and fields the phone system has never heard of. Someone has to sit down and say that this field over here becomes that field over there, and that the ones with no partner stay where they are.
Which direction. A field can be written one way, the other way, or both ways. The third is the only one that deserves the name two-way sync, and it creates every hard problem in the rest of this page.
The pieces are drawn on the home page. The other half of the subject is the systems doing the writing: an AI voice agent for small business that answers the phone when nobody can, a website chatbot that talks to a visitor at midnight, and workflow automation for small business, the wiring that moves the work between them. Each of them creates contacts, so each needs the identity rule on this page.
Dana's heating and cooling company (invented). Dana writes her map on one page, the whole list, each system with what it does to a contact:
- The website quote form. Creates a contact from what the customer types. One way, into the CRM.
- The phone system. Creates a contact when an unknown number rings, and logs every call. Both ways: the CRM sends back the name to show on the screen.
- The booking app. Creates a customer when Luis books someone new, and holds the stage of each job. Both ways: the stage is written back to the CRM.
- The invoicing program. Creates a customer when an invoice goes to a new name. One way, into the CRM, for the amount and the date it was paid.
- The home show sheet. Typed in by Luis on the Monday after a show. One way, into the CRM.
- The email tool. Creates nothing; it holds a copy of the list for campaigns. One way, out of the CRM, for the segment and the stage.
The campaign that wrote to Kathy read its segment from the March record. The heat pump job lived on the June one.
The system
From something happening to one record both systems agree on.
Six hops, and the two that decide whether this works are the second and the fifth. Everything between them is plumbing that either exists or does not. The identity decision and the conflict rule are judgements about your business. They are made once, and they are the two nobody asks about before signing.
Scroll to follow the chain
- The event: A call or a form
- The identity: Which contact is this
- The mapping: This field, that field
- The direction: One way, or both ways
- The conflict rule: Which side wins, agreed
- The record: One person, one history
Your turn. Draw your own map on one page: each system, an arrow for each field it moves, and one arrowhead or two.
Why does a CRM sync erase a field?
A CRM sync erases a field when an update replaces the whole contact and the sending system has an empty box where the receiving system held a value: the receiving system does what it was asked. It looks like data loss, and it is a design decision someone made without noticing.
The web has two ways of changing something that already exists, and they are not variations on a theme. The older one is a replace, which the specification calls PUT. The specification for it says that the sender is requesting that the stored version be replaced. The other one, PATCH, was defined in 2010, in a document whose introduction says in its first paragraph that a new method was necessary to improve interoperability and prevent errors. It sends a set of instructions describing how the thing currently stored should be modified to produce a new version.
In plain terms: one of them says "here is the contact, make it look like this", and the other says "change these two fields and leave everything else alone". Which one your sync uses, and what it sends for a field it has no value for, is set by each system's own documentation, and a partial update that names a field and sends it empty clears that field too.
Take an invented case. Your sync uses the first one, and the sending system has an empty box where the receiving system has a mobile number someone typed in by hand two years ago. The mobile number is gone. Nobody deleted it. The sending system described the whole contact, in its description that field was empty, and the receiving system did what it was asked. That is how a sync destroys information without raising an error, and the tell is always the same: the missing data is missing from exactly the fields the other system does not have.
The specification for the second method has a rule worth holding a builder to. The server must apply the entire set of changes atomically and must never provide a partially modified representation, and if the whole patch cannot be applied, it must apply none of it. All of it or none of it. A half-updated contact is not a smaller version of a successful update. It is a record in a state nobody intended, and the standard says it must not be allowed to exist.
Dana's heating and cooling company (invented). After Katherine's install, a technician typed her side gate code into the CRM. The booking app has no field for a gate code. The first time the booking app sent her whole contact back with a changed appointment, the code was gone, and the next technician stood at a locked gate on a Tuesday morning.
Your turn. Ask whoever built your integration one question: does an update replace the whole contact, or change only the fields named in it?
Why does the same update arrive twice?
The same update arrives twice when a network fails in the middle of a request and the sending system tries again without knowing whether the first attempt landed. Every sync you will ever own has to decide what to do about it.
The current HTTP specification has a precise word for the property that decides the answer. A request method is idempotent if the intended effect on the server of multiple identical requests is the same as the effect of a single one. The specification is explicit about why it bothers to define this: idempotent requests can be repeated automatically if a communication failure happens before the sender learns whether the first one worked.
Apply that to a contact record, in an invented case, and it stops being computer science. "Set this contact's stage to Booked" is idempotent: send it five times and the stage is Booked. "Add a note to this contact" is not: send it five times and there are five notes. The person who opens that record on Thursday sees the same message from you five times, and draws a conclusion about you.
The specification also says what a careful system does about it. A client should not automatically retry a request that is not idempotent, unless it has some way of knowing that the request is safe to repeat, or some way of detecting that the original never landed. It then says, drily, that some clients take a riskier approach and attempt to guess when an automatic retry is possible. Ask whether that sentence describes the integration running in your business. The symptom is duplicate activity on a contact record, not an error anyone sees.
Your turn. Open the record of a customer who had a busy week with you and look for the same note, text or call logged more than once at the same time.
What happens when both systems change a contact at once?
When both systems change a contact at once, the change written last wins and the other one disappears without an error, unless the sync checks what was there before it writes. In an invented case, you open the contact in the CRM and change the phone number. At the same moment an automation, acting on what it read a second earlier, writes the whole contact back. Your change is gone, and nothing anywhere records that there was ever a disagreement.
The specification calls this the lost update problem, and names it in exactly those words. It describes the mechanism that exists to prevent it: a conditional request. The sender includes a marker for the version it last saw. The receiving system is required not to perform the change if that marker no longer matches, and to refuse with a 412 (Precondition Failed) answer instead. The specification for partial updates goes further. It warns that collisions between two of them can be more dangerous than collisions between two replaces, because some kinds of change need to start from a known base point or they will corrupt what they are changing.
All of it is written down and free to read, and none of it is guaranteed in an integration unless someone asks for it. The cheap way to build a sync is to write the newest thing you have and not ask what was there before. That choice has a name: last write wins. It is a respectable choice for some fields and a disaster for others. The point is that someone has to make it deliberately, field by field, and be able to tell you what they picked.
Your turn. Ask whoever built your sync what it does when the version it read is no longer the version stored: does it check first, or write anyway?
Which system wins when two contact records disagree?
When two systems disagree about a contact, the system that a written conflict rule names wins; there is no clever way out and no default that is right for every field. Every two-way sync eventually receives two different answers to the same question and has to pick one, so the build either contains a decision or it contains an accident.
One rule has two halves. The most recently changed value wins for anything factual, such as a phone number or an address, and the CRM wins for anything about the relationship, such as the stage of the job or who looks after the customer. Another rule is that anything a person typed beats anything a machine wrote, always, on the grounds that the person was talking to the customer while they typed it.
Neither is right for every business. Three things matter. That your rule is written down somewhere you can read it. That it was chosen by someone who understood what each field is for. And that you can find out what it is without opening a support ticket. Our own service says plainly that the sync does not decide which side is right, and that the conflict rules are agreed when it is built: the most important conversation of the project.
Dana's heating and cooling company (invented). Dana writes her rule for the fields that both sides can change:
- Mobile and email: the most recently changed value wins, and the old one is kept as a second number or a second email, never thrown away.
- Street address: the most recently changed value wins.
- Stage of a job: the booking app wins, because a job is booked and finished there first.
- Who looks after the customer: the CRM wins.
- Gate code and access notes: whatever a person typed wins over anything a machine wrote, and no system may send the field empty.
Your turn. Write the rule for five fields in your own CRM, the ones that both sides can change.
Which field decides whether a contact is new or a duplicate?
The field that decides whether a contact is new or a duplicate is the identity field your integration was told to match on, such as the email address or an ID of your own, and you can find out which one it is today, for free.
The developer documentation of one CRM, HubSpot's, is quoted here as the source for how one system handles this, not as a recommendation, because it is unusually direct about it. Its combined create-and-update endpoint asks you to name the property you are identifying people by. It says that you can use email or a custom unique identifier property, and that following the request, if the contacts already exist they will be updated, and if the contacts do not exist they will be created. One field, nominated by whoever built your integration. Every duplicate that integration creates is downstream of it.
The same page carries a warning that is easier to read now than to find by hitting it. Partial upserts are not supported when using email as the identifying property for contacts. The documented remedy is to identify people by a custom property of your own instead. That is worth knowing before the build rather than after: where a partial update is not on offer, a builder may end up sending the whole contact, which is the road to the erasing behaviour described above.
And it documents what survives a merge, a detail worth knowing before your first one. When two contacts are combined, the losing record's email address does not evaporate. It is kept as an additional email on the surviving record. Those additional addresses are still unique identifiers, so no other contact can take them. The database remembers that Kathy existed. It has to, because the next time a message comes in from that address, something has to know where to put it.
Dana's heating and cooling company (invented). Dana's CRM matches web forms on the email address. Kathy used her personal email in March and Katherine her work email in June, so the identity field did exactly what it was told, twice. Dana's rule from now on: the website and the phone system look for an existing contact by email, then by mobile, before they create one, and anything that agrees on the surname and the street but nothing else goes to Luis's Friday list. So does any match on an email two people share, such as a household address, and any case where the email points at one contact and the mobile at another: both are doubtful pairs, never obvious ones.
Your turn. Find out which field each of your systems matches on before it creates a contact. The answer is in each system's settings or in the notes of whoever built it.
How do you check a CRM for duplicate contacts in twenty minutes?
To check a CRM for duplicate contacts in twenty minutes, take your last ten customers, search for each by surname, first name, email and mobile, and count the ones who come back more than once. Nobody needs a consultant for the first pass. Do it before anyone quotes you, because the quote is worth more when you have the answer.
- Your last ten customers. For each of those people, search your CRM for the surname on its own, then the first name on its own, then the email address and the mobile number. Count the records that come back and are plainly the same human being. Ten is a small sample, and it is not meant to be a statistic. It gives you a first sense of whether the answer is close to zero or common in your file, which is enough to decide whether to look further.
- Yourself. Put your own name into your CRM, your marketing tool and your phone system in turn, and see what each of them thinks it knows about you. Your own record is the one you can check instantly and argue with nobody about.
- Your automated messages. Open the last one that went out and look at whether the name in the greeting comes from a field a person typed or a field a form captured. Kathy is in that email because of which field someone chose, once, quickly, when it was set up.
- Your builder. Ask whoever built your current integration what happens if the same webhook arrives twice. If the answer is a shrug, you have learned something more useful than any audit, and it cost you a sentence.
Dana's heating and cooling company (invented). Dana and Luis run the check on a Thursday afternoon. Of their last ten customers, four are in the CRM twice and one three times. Dana is in it twice herself, once from the day she tested the website form. The last campaign email greeted people by the name typed on the form, which is why it said Kathy.
How do you fix duplicate contacts already in a CRM?
To fix duplicate contacts already in a CRM, reconcile the records you have before you switch on any sync, in six steps, with a copy of the CRM kept first and a person deciding every pair that is in doubt. Reconciling an existing database and syncing it forward are two separate jobs, and doing them in the wrong order turns one project into two.
- Measure what is in there. The twenty-minute check above, then the same count on a larger sample.
- Export a full copy of the CRM, and try a restore. Every contact with its notes, jobs and history, and the links between them. A copy you have never restored from is a hope, so restore one contact from it before the first merge.
- Agree the identity rule. Which field, or which order of fields, says that a contact already exists.
- Reconcile the obvious matches. Pairs that agree on the identity field, with nothing that contradicts them, merged on your written say-so. For each, decide which record survives, then check that the jobs, notes, invoices and the customer's answer on marketing messages came across to it. Where the two records disagree on that answer, keep the stricter one until the customer tells you otherwise.
- Put the doubtful pairs in front of a person. Every pair in the middle band, settled by the person you named, in the slot you gave it.
- Only then switch the ongoing sync on, and test it. With the conflict rule written down, a named person told when a sync fails, and the test call from the questions below made once it is live.
Dana's heating and cooling company (invented). Dana's numbers are the starting values of the calculator below, and two of them are her own guesses, not measurements. She guesses that a quarter of her 3,000 contacts belong to someone who reached the company more than one way: 750 contacts that may have a second record somewhere. She guesses that for a fifth of those the details would not match exactly: about 150 pairs that need a person to decide. At about two minutes a pair, that is 300 minutes, five hours of Luis's attention, which he spreads over the first three Friday slots. The obvious matches, the ones that agree on email or mobile with nothing that contradicts them, are merged on Dana's say-so before he starts.
In your numbers
How much of this is a computer's job and how much is somebody's afternoon?
The number on the dashboard, including the years you have stopped thinking about.
Anyone who might have filled in a form and also rung, or booked online and also been typed in by hand at the office. Past customers who came back for a second job are the place to start.
A short name against a formal one, a work email against a personal one, a moved house. The Census research on high quality files found more than one pair in five disagreed on the first name alone.
Seconds when you know the person, and several minutes when you do not.
Pairs a computer should not settle on its own
150pairs to settle
- Contacts in the databaseyour 3,000
- 3,000contacts
- Who could have arrived more than onceyour 25%
- 750possible pairs
- Where the details would not match exactlyyour 20%
- 150pairs to settle
- At your deciding timeyour 2
- 300minutes
- In hours60 minutes in an hour
- 5hours of somebody's attention
The headline is the middle row rather than the hours. The hours are the easy half. The number that decides whether this project is an afternoon or a fortnight is how many pairs end up in front of a person. That is the number nobody quotes on. Sliders that multiply shares will produce a fraction of a pair, which is not a thing. Read two and a half as between two and three. What this refuses to do is tell you how many duplicates you actually have. Figures for that circulate widely, and none of those we found states a sample or a method. So both of the shares above are yours to supply. The second of them carries a hint pointing at the one measurement in this subject that does state a method. There is also no row for what a stale record costs in lost work. There was never going to be one. The email that went to the wrong version of someone is real. Nobody can show you the sums that turn it into money.
Dana's heating and cooling company (invented). Six weeks later Kathy and Katherine are one record, under the name on her bills, with both emails and both mobiles kept. The website form, the phone system, the booking app and the invoicing program each look for an existing contact by email, then by mobile, before they create one, and the home show sheet goes through the same check when Luis types it in. The email tool reads its segment and the stage of each job from the one record, so a customer with a booked or finished job is out of the furnace campaign. No system may send the gate code empty. And the Friday list, which held 150 pairs in its first three weeks, now holds four or five new ones a week: about ten minutes of Luis's Friday.
Your turn. Before anything is connected, write down who settles the doubtful pairs and when, because the queue is the one part of this that never goes away.
The honest read
Send us a list of the systems that create a contact in your business, and the one field each of them uses to recognise someone. We will send back where the duplicates are coming from. We will say which of the joins is doing the damage, and which parts of the fix are a week and which are an afternoon.
It is a short reply from a person, and it costs nothing. We do not need access to your database to answer it.
What does it take to fix duplicate contacts in a CRM?
Fixing duplicate contacts in a CRM takes an amount of work set by four things, and none of them can be guessed from a distance: how many systems have to be joined, whether each of them exposes an interface a program can use, how many fields have to be mapped by hand rather than by name, and whether the records already in there have to be reconciled before anything is switched on. The fourth is the one a quote that only counts connections leaves out, and it is the one that needs a person, because every doubtful pair is a decision.
The shape of the work can be described without a number. A single pair of systems with a clear identity field and a handful of mapped fields is a smaller job than a business with an old CRM, a newer CRM nobody finished moving to, a phone system and a booking app. In either, the connecting is the part software does. The part that needs your time is someone sitting with you over a list of fields and asking what each one is for, the same kind of conversation that sits at the heart of an AI assessment. It cannot be hurried, because some fields were set up by someone who has since left, and nobody remembers what they were for.
Then there is the recurring cost, which is a person's attention. A sync is software, and it can break the week an app renames a field, so something has to notice. And the middle band from the matching model never goes away. If you want fewer wrong merges, more pairs land in front of a person. That is a standing item on someone's week, not a one-off.
Here is the number we cannot give you, and it is the one you asked for: how many duplicates are in your CRM right now. Figures for this circulate, and none of those we found states a sample or a method, so this page does not use them. The calculator above asks you for the inputs instead, and the honest first step of any real project is measuring your own file, not accepting someone's average.

What does a CRM sync not fix?
A CRM sync does not clean the records already there, remove the human decision, open a system that will not open, make anyone use the CRM, or tell you which of two records is the real one.
A sync does not clean what is already in there. Keeping two systems in step from today onward and reconciling years of accumulated records are two different jobs of two different sizes. A sync switched on over an unreconciled database faithfully copies every mess in it to a second system.
A sync does not remove the human decision. That is the point of the three-answer rule. Any build that reports a hundred percent automatic resolution has emptied the middle band, so it settles the doubtful pairs on a guess, and some of those guesses can merge two people. The failure of an over-eager deduplicator is worse than the one it fixes. Two separated records are an embarrassment. Two people fused into one is a stranger reading someone else's history.
A sync does not open a system that will not open. A CRM that publishes an interface for other software can be connected; a platform that publishes none is not going to be talked round. Screen scraping something with no interface is not a sync. It is a liability with a schedule.
A sync does not make anyone use the CRM. A record that is finally right is worth nothing at all if the appointment still lives on a sticky note. No integration reaches into a notebook.
And a sync will not tell you which of the two records is the real one, in the cases that matter. It will tell you they are probably the same person. Which email she reads, which number she answers, and which of the two histories you should have in front of you before you ring her: that is judgement, and it belongs to whoever knows her.
Three ways a good build produces nothing
None of them are the software.
The review queue nobody opens
The middle band only works if a person actually goes there. A queue nobody owns is not a safety mechanism. It is a filing cabinet, and the pairs sitting in it are the exact records most likely to be the ones you care about. A short slot on a Friday is the whole job. Nobody schedules it, because on the day it is set up it is empty.
Fields that were never agreed, only defaulted
Someone has to say which side wins for each field that both systems can change. It is a slow conversation, because half the answers turn out to be that nobody remembers what the field was for. Skipping it does not remove the decision. It hands the decision to whichever system happens to write last. That is a rule you will discover months later, by losing something.
A record that is finally true and nobody reads
The whole point of this is that the thing in front of you when you pick up the phone is current. If the team still works from a notebook, a group chat and their own memory, the CRM becomes a spotless archive that describes a business happening somewhere else. No integration has ever reached into a notebook. And no amount of accuracy makes up for someone looking.
Common questions about duplicate contacts and CRM sync
What is two-way CRM sync?
Two-way CRM sync means information travels in both directions between your CRM and the other systems you use: calls, texts and bookings write into the CRM, and changes made in the CRM flow back out to the systems that act on them. With one-way sync, a change made on the receiving side never travels back, and when the side that missed it is the one sending the automated messages, the customer sees it.
How is CRM sync different from workflow automation?
CRM sync is about one fact being true in more than one place at the same time, while workflow automation is about work moving between systems: a form is submitted, so a task is created and a reply goes out. A business with fine automation and no sync sends perfectly timed messages to the wrong version of a person.
Will a CRM sync create duplicate contacts?
A correctly built CRM sync exists to prevent duplicate contacts: an inbound record that matches an existing contact on the agreed identity rule updates that contact instead of creating a second one. No honest build promises that the matching is never wrong, because the published model says you choose between two kinds of error and cannot have zero of both.
Does a CRM sync work with the CRM I already use?
A CRM sync works with a CRM that publishes a documented interface for other software, and the connection is built to your systems rather than limited to a list of supported ones. The better question about any platform is the three on this page: which field identifies a person, whether an update changes named fields or replaces the whole record, and whether the same message arriving twice creates one thing or two.
What happens to the duplicate contacts already in the CRM?
Nothing happens to the duplicate contacts already in a CRM until someone decides what should happen: reconciling them is a separate job from syncing forward, done first, in the six steps above.
Can a CRM merge duplicate contacts automatically?
A CRM can merge some duplicate contacts automatically, and the share depends on a threshold someone sets: the dial between a system that leaves work for you and one that occasionally fuses two people together. On a first build, merge automatically only the pairs with no doubt in them, send the rest to a review queue, and widen it once you have watched what the queue holds.
Is fixing duplicate contacts worth it for a one-person business?
Fixing duplicate contacts can be worth more to a one-person business than to a large one: in a big office several people touch a record, so a duplicate has more chances of being caught, and on your own you are the only person who would notice. The message that goes to the wrong version of someone goes out under your own name.
How do you check that call records sync correctly with your CRM?
To check that call records sync correctly with your CRM, make one test call from a number the CRM already holds and one from a number it does not, then look at the CRM: the first call should appear once, on the existing contact, with its time, length and outcome, and the second should be handled once, the way your rule for a new caller says. Whether a retried write lands twice is a question for the builder, answered from the sync's log, not from a test call.
How do you know a CRM sync is working after it goes live?
Two signs say a CRM sync is working after it goes live: a review queue that someone opens and empties, and a monthly test in which you change one field in each system in turn and see the other side follow. An empty queue in the first week is a warning: either the records were clean or the matching is too confident, and the second is the one to rule out first.
Do you work with businesses outside New York?
We are based in New York, and the call works for an owner anywhere. The sync is built to the systems your business already uses, wherever the business is.
What to do first about duplicate contacts in a CRM
The quickest check on this page is your own name in your own CRM. It costs nothing, and it settles in one screen an argument that people have in meetings for months.
If more than one of you comes back, you already know what the automated message going out on Tuesday is addressed to. And if your duplicates are like Kathy's, you know something else: it was never a discipline problem, because nobody typed anything wrong. Two systems were separately right about the same person, and nothing in the building had the job of noticing.
Katherine is one person. Your database is entitled to know that.
Type your own surname into your own CRM and see how many of you there are. It takes one screen, and it settles an argument people have in meetings for months. If more than one of you comes back, you already know which version of everybody else your Tuesday email is talking to.
There is no price here, because the cost tracks four things nobody can guess from an article. How many systems there are to join. Whether each of them has an interface a program can use. How many fields need mapping by hand. And whether the records already in there have to be reconciled first.
Start with the page on CRM integration for small business: what the sync connects, what you approve, and what it will not do.
Sources
- William E. Winkler, "Overview of Record Linkage and Current Research Directions", 2006, report 2006-2 of the statistical research division of the US Census Bureau, census.gov. An older source, kept for the argument: the matching model it restates is still the published one.
- U.S. Census Bureau, "Frequently Occurring Surnames from the 2010 Census", census.gov. An older source, kept for the argument: the counts are those of the 2010 Census.
- Ivan P. Fellegi and Alan B. Sunter, "A Theory for Record Linkage", Journal of the American Statistical Association, 1969. Not read here: cited through Winkler's restatement above. An older source, kept for the argument.
- RFC 9110, HTTP Semantics, 2022: the replace method, idempotent methods and the lost update problem, rfc-editor.org. An older source, kept for the argument: it is the current HTTP specification.
- RFC 5789, PATCH Method for HTTP, 2010, rfc-editor.org. An older source, kept for the argument: it defines the partial update.
- HubSpot, developer documentation for the contacts endpoints, developers.hubspot.com: the identity property, partial updates with email, and what survives a merge. Live documentation.
Keep reading
CRM data cleansing: stale contact records, data enrichment and the field it overwrites 
August 25, 2026
CRM data cleansing: stale contact records, data enrichment and the field it overwrites
Why emails go to spam: email deliverability for a small business 
August 25, 2026
Why emails go to spam: email deliverability for a small business
Lead follow-up for old leads: how to win back past customers 
July 30, 2026
Lead follow-up for old leads: how to win back past customers
