You Built a Search Box. The Statute Calls It Investigating for Hire.
Private-investigator licensing predates the internet by most of a century, and its definitions are written broadly enough to cover a great deal of modern data tooling. Unlicensed practice is a crime in a substantial number of states. Almost no AI people-search team has read the statute in the state where their largest customer sits.
Two regimes, stacked — and teams usually only see one. Every conversation in this space starts and ends with data rules: where the data came from, what purpose it may be used for, what the terms of service said. Those rules are real and this article covers them. But they sit on top of an older question that the data rules never answer, which is whether you are permitted to be doing this work at all. A perfectly sourced dataset used to perform an unlicensed investigation is still an unlicensed investigation, and the licensure question is the one with criminal exposure attached.
The Split That Decides Everything
Nearly every hard case in this category resolves to a single question: are you selling a tool that a customer uses, or performing an investigation for a customer? The statutes turn on conduct, so the answer is determined by six operational facts, not by the words on your pricing page. Read both columns and mark honestly which side each row of your business is on.
Tool posture
- ▸The customer states the question and the customer reads the raw result.
- ▸Output is a record set with provenance, not a narrative conclusion about a person.
- ▸No human on your side reviews, curates or interprets any individual matter.
- ▸Pricing is per seat or per query, not per case or per successful locate.
- ▸Marketing sells access to data and search capability, not outcomes.
- ▸You do not accept case-specific instructions, deadlines or follow-up requests.
Service posture
- ▸You accept a question about a named person and return an answer.
- ▸Output is a report — a resolved identity, a current address, an assessment, a conclusion.
- ▸Anyone on your side triages ambiguous matches or handles escalations for a customer.
- ▸Pricing is per successful locate, per case, or contingent on a result.
- ▸Marketing promises to find people, verify claims or uncover activity.
- ▸You take intake, run the matter and deliver, which is the shape of an investigative engagement.
The rows do not have to agree, and when they disagree the service-shaped ones win. A company that looks like a tool on five rows and a service on one — usually because an account manager helpfully resolves difficult matches, or because a large customer negotiated per-locate pricing — has the exposure of a service and the compliance programme of a tool.
Seven Surfaces, Graded
You sell data; the licensee performs the investigation and carries the licence, the insurance and the client relationship. Contract for it explicitly and verify the licence at onboarding rather than assuming it.
The customer forms the question and reads the output. It stays a tool only while nobody on your side touches an individual matter — a support queue that resolves ambiguous matches is the thing that changes the character.
Entity resolution asserts that these records describe the same person. That is an investigative determination presented as a data feature, and it is also the highest-consequence error in the whole category.
Accepting a subject and returning a located address is the paradigm licensed act in most statutes, whatever the pricing page calls it and whoever — or whatever — does the work.
Continuous monitoring of a named person implicates surveillance and anti-stalking statutes independently of licensure, and several states treat sustained monitoring differently from a one-time lookup.
Insurance investigation frequently has its own licence class and its own unfair-practices overlay. Doing it with a model does not move it out of the class.
Assembling information for an eligibility decision about a person is consumer-reporting conduct with notice, dispute, accuracy and adverse-action duties. A people-search product used this way inherits all of them.
The Data Regimes That Apply On Top
Licensure governs whether you may do the work. These govern what you may do it with, and each one has a specific way that an AI pipeline breaks it that a manual process did not.
Motor-vehicle records
The rule: Access and downstream use are restricted to enumerated permissible purposes, and resale is restricted too.
What AI breaks: A model that enriches a profile does not know why the customer asked. Permissible purpose is a property of the request, so it has to be captured at the request and carried through every derived artefact — including caches, embeddings and training sets.
Consumer-report data
The rule: Furnishing information for employment, credit, insurance or housing decisions triggers accuracy, notice, dispute and adverse-action obligations.
What AI breaks: The trigger is the customer's use, not your product category. A disclaimer stating the data may not be used for eligibility decisions is worth little if your own marketing describes screening use cases or your usage data shows them.
Scraped social and web content
The rule: Terms of service, anti-circumvention and computer-access statutes govern collection independently of whether the content is visible.
What AI breaks: Publicly visible is not the same as freely collectable, and collection at automated scale is treated differently from a person reading a page. Aggregation also changes the character of the result — assembling scattered facts into a dossier is the thing the statutes are about.
Location and mobility data
The rule: Sale and use are increasingly restricted by state privacy laws, with sensitive-location carve-outs and consent requirements.
What AI breaks: Inference is the gap. A model that derives a likely home or workplace from non-location signals produces location data without ever purchasing any, and no contractual restriction on the input covers the inferred output.
Biometric identifiers from images
The rule: Face templates generated from photographs are separately regulated in several states, with consent and retention rules and private rights of action.
What AI breaks: Face matching to confirm an identity generates a template. That is a regulated act performed on someone who is not your customer and never consented to anything.
The Wrong-Person Problem Is the Real One
Everything above is a compliance question with a compliance answer. The failure that actually harms people is narrower and more mundane: the system resolves an identity incorrectly, and something happens to a person who has no relationship with you and no way to find out. A collector contacts the wrong household. A claim is denied on someone else's photographs. A landlord declines an applicant on a record belonging to a person with the same name and a different middle initial.
Two properties of AI pipelines make this worse than the manual equivalent. Match rate is the metric everyone optimises, and optimising it directly trades against precision at exactly the ambiguous cases where precision matters. And confidence is usually collapsed to a single best answer before it reaches the customer, so the downstream user never learns that the system was choosing between two plausible people.
Five controls that reduce this specific harm
- 01Surface ambiguity instead of resolving it. When two candidates are close, return both with their distinguishing evidence rather than the higher score.
- 02Retain the evidence chain per assertion — which records, which fields, which match logic, at what version — because a re-run against a changed model is not a reconstruction of what you told the customer.
- 03Capture the customer's stated purpose at the request, bind it to every derived artefact, and refuse to serve restricted sources when the purpose does not permit them.
- 04Give the subject a route in. A dispute and correction path is required outright in some uses and is the cheapest possible mitigation in all of them.
- 05Measure precision on hard cases specifically — common names, shared addresses, same-name relatives, stale records — rather than reporting accuracy on a sample dominated by easy matches.
Frequently Asked Questions
Does an AI people-search product need a private investigator licence?
It depends on whether you are selling a tool or performing an investigation, and statutes decide that on conduct rather than description. Most states define private investigation as engaging in the business of obtaining information about a person — identity, whereabouts, character, conduct, affiliations or transactions — for another person, for compensation. That covers much of what a locate product does, and the licence typically requires experience hours, an examination, bonding and insurance, with unlicensed practice carrying criminal penalties in many states. The practical dividing line: does the customer form the question and read the raw output, or do you accept a question about a named person and return an answer? The second shape is investigation regardless of whether a model or a person did the work.
Isn't public-records data exempt from investigator licensing?
Some statutes carve out persons who merely furnish information from public records. The exemption is real but far narrower than the industry treats it, and three things usually defeat it. Aggregation: assembling scattered public facts into a profile of a named individual is a different act from furnishing a record, and it is the act the statutes were written about. Non-public sources: almost every commercial locate product blends public records with credit-header data, marketing files, utility and telecom data or scraped content, and the exemption does not cover the blend. Conclusions: a resolved identity, a confidence-scored current address or an assessment of activity is an investigative output rather than a record. Read your own state's exemption text — the wording varies substantially and a summary of the general rule is not a defence.
We are a data provider, not an investigator. Does that hold?
It can, and it is the cleanest posture here, but it has to be true operationally rather than in marketing copy. Six facts determine it: who forms the question; whether the output is records with provenance or a narrative conclusion about a person; whether any human on your side touches individual matters, including a support queue resolving ambiguous matches; how you price, since per seat or per query is data-shaped while per successful locate or contingent pricing is service-shaped; what you promise, access and capability versus outcomes; and whether you accept case-specific instructions and follow-up. A company that satisfies the first five and fails the sixth through a helpful account manager has become an investigative service without anyone deciding to.
What is the biggest technical risk in AI-driven identity resolution?
The wrong person, delivered with high confidence and downstream consequences. Resolution across common names, shared addresses, same-name relatives, stale records and transliterated variants is genuinely hard, and a model trained to maximise match rate produces confident false positives at exactly the ambiguous moments. The consequence attaches to someone who is not your customer and has no way to find out: a collection contact directed at the wrong household, a claim denied on someone else's activity, a person named in a report they will never see. Two design responses matter more than model quality — surface uncertainty honestly instead of collapsing it to a single best match, and retain the evidence chain for every assertion so a dispute can be answered without re-running a model that has since changed.
Does automated monitoring of a person create separate liability?
Yes, and it is the most underestimated surface, because monitoring feels like a smaller version of searching when it is legally a different thing. Sustained observation of a named individual implicates anti-stalking and surveillance statutes that turn on a course of conduct rather than a single act, and automation makes that course of conduct trivially provable — you logged every check with a timestamp. Several states separately regulate specific techniques, including location tracking without consent and certain recording. Two features convert a research product into a monitoring product almost by accident: a saved search that runs on a schedule, and alerting on change. If you have either, a person is being watched continuously by a system, and neither the licensure nor the surveillance question is answered by each observation having looked at public data.
Our customers are law firms and insurers. Does their status cover us?
Not by itself. A privileged or regulated customer does not confer licensure on a vendor, and several states are explicit that work performed for an attorney still requires a licence unless the worker is an employee of the firm rather than an outside contractor. There is a real structure available — operating as a data provider to a licensed investigator, or as an in-house resource of the firm — but it must be documented in the contract and reflected in how the work actually runs. Insurance adds a layer: claim investigation often has its own licence class plus an unfair-claims-practices overlay governing how results may be used, so an automated activity check driving a denial touches both regimes at once. The failure pattern is a vendor assuming the customer's licence covers the vendor, and a customer assuming the vendor is licensed. Verify both in writing at onboarding.
How does this interact with our AI assistant that just searches the web?
It interacts more than teams expect, because a general assistant with browsing performs the same acts when the prompt is about a named person. Three distinctions decide it. Whether the subject is a specific identified individual rather than a topic — 'summarise this company's litigation history' is research, while 'find where this person lives now' is a locate. Whether you charge, since compensation is an element of most investigator definitions and a bundled subscription counts. And whether the assistant aggregates and concludes rather than retrieving and citing. An assistant that answers open questions about named private individuals, for paying users, at scale, is doing the regulated act with a general-purpose wrapper — which is why refusing person-locate prompts is a licensing control and not just a safety policy.
The Support-Ticket Test
Pull the last hundred support tickets. Count how many involve a specific named subject where someone on your team looked at the data and told the customer which result was the right person.
Every one of those is a matter your company investigated. The licence question is not about the product architecture — it is about that number, in that customer's state, and it has never once been answered by a term of service.
Related Reading
- AI background checks and screening discrimination — what happens when the same data reaches an eligibility decision.
- AI identity verification and biometric law — the face-matching layer, where a template is generated on a non-customer.
- AI claims adjusting and unfair claims practices — the insurer side of an automated activity check.