Blog·September 1, 2026

AI Is Great. Please Don't Let It Decide Which Client Record to Delete.

AI is getting shoved into basically every part of running a law firm: intake, document review, email, scheduling, research, CRM workflows and practice management. Someone is probably working on an AI-powered office coffee pot as this is being written. A lot of it is genuinely useful. But there’s a pretty important difference between asking AI to help you find something and letting AI make an irreversible decision about your data.

Duplicate contacts are a good example. On the surface, deduplication sounds like exactly the kind of boring job AI should handle. Give it two contacts, ask whether they’re the same person and let it clean everything up. Except real law-firm data is messy. Robert J. Smith, Bob Smith, Robert Smith and Robert J. Smith Jr. might all be the same person. Three might be the same person. Two might be father and son. One might be opposing counsel and another a former client. Somebody may also have typed the wrong phone number into one of those records six years ago.

AI can look at all of that information and make a very convincing guess. The problem is that convincing and correct aren't the same thing.

AI doesn't know. It predicts.

Modern AI systems are extraordinarily good at finding patterns and making judgments from incomplete information. That's part of what makes them useful. But when an AI model decides two records probably represent the same person, it's still making an inference. Even a system that's right 99% of the time gets uncomfortable when the other 1% involves combining records belonging to different people.

A contact in a law firm's database isn't just a name and email address. It can be connected to matters, communications, billing relationships, documents, custom fields, notes and years of history. Merge the wrong two people and the problem isn't simply that the address book looks funny. You've changed the structure of the firm's data. If the system helpfully deletes the duplicate afterward, you've also removed one of the records that would have made figuring out what happened considerably easier.

That's not something an AI model should quietly decide at 2:13 in the morning because it felt 94% confident.

Automation may actually be creating more of the problem

There's an irony here. The same wave of automation that's supposed to make law firms more efficient can create more opportunities for fragmented contact data.

A prospective client fills out a form and the intake platform creates a contact. Another integration sends information into the practice-management system. The person calls the firm later and an employee creates another contact because they don't realize one already exists. Another system syncs the name or email a little differently. Eventually Jennifer Thompson exists four times, and none of the systems necessarily did anything wrong. Each one performed the task it was given.

The difficult part is identity. “Create a contact for this submission” is easy. “Determine whether this human being is already represented somewhere among thousands of imperfect records, and be certain enough to modify or delete those records” is a very different problem.

AI can absolutely help with that problem. It just shouldn't own the final decision.

Finding duplicates is the easy part

Good deduplication isn't just similarity detection. Finding two records that look alike is only the beginning. You still need to determine whether they're actually the same person, what evidence supports that conclusion, what information exists on each record, what should survive, what other records are connected to them and what happens to those relationships during the merge.

Then there's the fun stuff: the records disagree.

Maybe the two contacts have different values in the same custom field. Maybe there's enough evidence to believe they're the same person, but not enough evidence to decide which value is correct. Software has several options at that point. It can pick one. It can overwrite one. It can ask an AI model which one looks more plausible.

Or it can stop.

MERGEguard is intentionally conservative about that kind of thing. It looks for evidence that records represent the same person, presents potential duplicates for review and handles the mechanics of combining them while preserving the relationships that need to survive. When it encounters a conflict that shouldn't be resolved automatically, it doesn't invent an answer just to keep the automation moving. It leaves the data alone and asks for human input.

Sometimes “human needed” is the correct result.

Human review isn't a bug

There's a strange assumption developing around AI that if software still needs a person to review something, the software isn't advanced enough. For some jobs, that's backwards. The more consequential an automated action becomes, the more valuable a human checkpoint can be.

If AI screws up an internal summary, somebody corrects the summary. If AI suggests the wrong duplicate contact, somebody can reject the suggestion. If AI autonomously merges two different people, moves the relationships attached to those records and deletes one of them, that's a different category of mistake.

The goal shouldn't be removing humans from every workflow. It should be removing the tedious work so humans can spend their time on the decisions where judgment actually matters. Let software sift through 15,000 contacts. Let it surface records that look suspiciously similar and assemble the evidence. Let it handle the tedious mechanics after the decision has been made. But when the evidence isn't good enough, don't guess.

AI is a tool, not a permission slip

AI is going to become a bigger part of legal technology, and that's not inherently a bad thing. It can make intake faster, make information easier to find and eliminate an enormous amount of tedious work. But “AI-powered” shouldn't automatically mean “allowed to make unsupervised changes to the system of record a law firm depends on every day.”

That's the philosophy behind MERGEguard: automate the tedious work, make decisions based on evidence and keep a human in control when the answer isn't clear.

Because the coolest AI feature in the world isn't particularly useful if somebody has to spend Friday afternoon figuring out why it decided Robert Smith and Bob Smith were the same guy, and why Bob is now somehow opposing counsel in his own divorce.