Why the case still belongs to a detective
Investigations turn on two things software handles badly: getting a person to talk, and standing behind a decision in court. A detective sits with a frightened witness, hears what the witness will not say, and comes back a week later when the story shifts. That same detective has to defend every step under oath while a defense attorney picks at it.
Evidence work also carries legal weight that attaches to a named officer. Warrant applications, interview procedure, chain of custody for seized items, the decision to arrest or to keep watching: each one is a judgment someone signs. A model can draft a warrant affidavit. It cannot be the affiant. It can flag an inconsistency in a statement, but it cannot decide that the inconsistency is grief rather than deception.
That is why roughly 61% of task time in this job sits in work that needs a person. The pay and the headcount reflect how much of the role is judgment under pressure: the BLS counted 114,430 detectives and criminal investigators in the United States at a median wage of $93,790 (BLS, 2025).
What software handles, what it assists, and what it leaves alone
Document-heavy steps have moved fastest. Pulling prior contacts out of records systems, transcribing hours of interview audio, matching names across jail bookings and license databases, summarizing a long file so a new detective can pick it up: tools do this work end to end on about 6% of task time. The cost gap shown above is the reason agencies try it. Reading a phone dump by hand takes days.
The assisted middle is larger in practice. Around 33% of task time is work a detective still owns while software speeds it up: sorting patterns in call logs and bank records, scanning surveillance video for a vehicle, drafting a report that the detective then corrects, building a timeline from mixed digital sources. The lead comes from the tool. The decision about whether the lead means anything does not.
What is left is the core of the job. Interviewing witnesses, victims and suspects. Working a scene and protecting the integrity of what is collected. Handling informants. Testifying. Coordinating with prosecutors on what a jury will accept. None of that is a document problem, and the share of task time that stays with people here has barely budged.
What the evidence shows, and what it does not
Evidence quality on the parity question grades D for this occupation. That is the bottom of our scale, and it means something narrow: no published study has yet tested an AI system against working detectives on detective tasks. So this page gives no parity number. Claims that a model matches an investigator are, at the last check, untested.
Coverage is the measurable part. How much of the task time AI can handle today scores 20 out of 100, and the method behind that figure is set out on the coverage scoring page. The headline Still needs a human figure for this job is 76 out of 100 (higher is safer).
Three kinds of study would settle the parity question. A blind trial where detectives and a model each work the same closed case files and are judged on the leads they produce. Field data from agencies using report-drafting and lead-generation tools, measured against clearance rates rather than hours saved. And a record of how courts treat AI-derived evidence when it is challenged. Until some of that exists, the honest answer is that nobody has checked.
When this job could change
Most likely after 2042 (8 in 10 of our scenarios). The replacement-year method explains what that window is actually measuring.
Two things could pull it earlier. Digital evidence keeps growing faster than caseloads, so phone extractions, cloud records and video now dominate the hours in many investigations, and that is the work tools do best. Flat hiring is the second pressure: the BLS projects employment for this job to grow about 0.2% from 2025 to 2035 (BLS, 2025), so agencies carrying the same caseload with the same headcount have a strong reason to buy software instead.
Two things hold it back. Admissibility and disclosure rules mean a prosecutor has to explain how a conclusion was reached, and an unexplainable output is a liability in a courtroom. And a meaningful slice of the role is physical, carried out at scenes, in interview rooms and at doors, which would need dexterous hardware that is not deployed anywhere near that standard.
What to do: get fluent with the digital-evidence tools your unit already buys, so you are the person who can explain their output to a prosecutor.
How detectives stay needed
Lean into the three things the task list puts firmly on the human side. Interviewing, including the slow work of building trust with reluctant witnesses and handling informants. Scene and evidence judgment, where the decision about what to collect and how to preserve it shapes everything downstream. And case presentation: testimony, and the back-and-forth with prosecutors about what will hold up.
Two skills compound on top of that. The first is digital-evidence literacy: knowing what a phone extraction, a video analysis or a financial trace can and cannot support, and where a tool’s confidence is misleading. The second is plain written reasoning, because the detective who can show how a conclusion was reached is the one whose cases survive challenge.
Nearby work is scored the same way. The closest comparison is police identification and records officers, whose evidence-handling tasks look different from investigative judgment. Also worth reading are police and sheriff’s patrol officers and private detectives and investigators, which carry far more document and surveillance work per case.
For the wider picture, see the law enforcement workers family, the government sector page, and the roles listed among the jobs that mostly need a person. You can put this job beside another on our compare tool, and the full scoring approach is on the methodology page.