Agentic OSINT · Practitioner guide
AI OSINT: what AI does in an open-source investigation, and where it stops
Very different tools are sold as AI OSINT. This guide sets out what the AI in them does well, where it goes wrong and how to check an AI finding before you rely on it. It names no vendor.
Practitioner guide · Published · Updated
AI OSINT in short
AI OSINT is open-source intelligence in which artificial intelligence does part of the work: summarizing a lookup, extracting names from a result, drafting a report or planning and carrying out the research steps. The term covers very different tools, so ask what the AI in a product does and what you can check. Agentic OSINT is one kind of AI OSINT: an AI agent plans and carries out a multi-step investigation toward an objective you define, cited findings link to retained source records, and you review, direct and decide.
The kinds of AI in OSINT tools
Every tool in this market works with open-source intelligence: intelligence produced from publicly available information to address a specific requirement, as US Public Law 109-163 describes it. The guide to choosing an OSINT platform sorts the tools into the five kinds below. A tool sold as AI OSINT can be the second, third or fifth kind, so ask which.
Free utilities and link directories
Single-purpose sites and curated lists of them. Each answers one narrow question. Choosing the checks, ordering them and recording what you found is your work.
Point tools with AI summaries
A lookup or search tool with a language model on top that summarizes the result or extracts names from it. The unit of work is still one lookup.
General AI chat assistants
Useful for orientation before a case begins. Research and citation features vary from one assistant to the next, so check each important claim against a source you can open.
Link-analysis tools and enterprise suites
Desktop graph tools and large suites, often sold by annual license or after a demo. They suit teams that model large datasets of their own or watch sources continuously.
Agentic OSINT platforms
An AI agent plans and carries out multi-step research toward an objective you define, inside a case. Cited findings link to retained source records, and you review, direct and decide. Nexus is a platform of this kind.
The kinds differ in the unit of work: one lookup, one question with one generated answer, or one objective pursued through a planned multi-step workflow. The agentic OSINT guide explains the fifth kind in detail.
What AI does well in an investigation
The volume of public information about a person or an organization has outgrown what one investigator can work through by hand. AI helps with four parts of that work.
Planning multi-step research
Given an objective, an AI agent can propose the research steps and their order: which sources are relevant, what to retrieve first and what follows from each result. You accept the plan or adjust it.
Covering many sources
One subject can appear on dozens of platforms, in several company registries and in news archives in more than one language, each with its own search interface, format and gaps. An agent can cover more of them, faster, than a person can.
Cross-referencing identifiers
Names, usernames, email addresses, domains and wallet addresses recur across sources. AI can compare them and point to records that may concern the same person or organization. Each link it proposes is a possible match until independent evidence corroborates it.
Summarizing with citations
A language model can turn retrieved material into a readable summary or a draft brief. The summary is worth having when each statement carries a citation you can open.
None of this changes the standard of proof. An agent cannot turn a possible match into a verified association; only corroborating evidence does that.
Where AI goes wrong
A language model writes fluent text whether or not the evidence supports it. These are five errors to look for in an AI finding.
Unsupported or invented statements
Sources may be missing, misattributed or invented, and uncertainty is flattened into confident sentences. A statement without a source you can open is an opinion.
A shared name or handle taken for the same person
A matching name, a reused username or a similar photo is a possible match, not proof of identity, ownership or control. A tool that merges two people into one profile produces a confident error.
Stale or missing sources
Public sources hold outdated records, and records are removed, redacted or never indexed. An answer can be no better than the sources that were reached on the day of the search.
“Nothing found” read as clearance
An unsuccessful search shows that nothing was found in the sources checked at that time. It does not show that the information or the activity is absent.
Text on a web page that tries to instruct the model
A web page that an AI tool reads can contain text written to instruct the model, for example to leave something out or to report something false. Fetched content is material to check, not instructions to follow, and a finding that rests on one page needs a second source.
How to check an AI finding
Five steps, in this order, for any finding you intend to rely on.
Open the cited source
Read the record or the page the citation points to and confirm that it says what the finding says. A citation shows where a statement came from. It does not make the statement true.
Check its date and its publisher
Note when the source was published, when it was retrieved and who stands behind it. A record from an official registry carries more weight than user-generated content, and an old record can describe a situation that no longer holds.
Corroborate with a second source
Look for an independent source that supports the same fact. A page that copies the first one is not a second source.
Keep possible matches apart from supported findings
A shared name, username or photo stays a possible match until you corroborate it. Record it as a lead, not as a fact.
Record what was not found
Write down which questions the sources could not answer and where you looked. A gap is a finding in its own right and never proof of absence.
Law and proportion
AI changes the speed of open-source work, not the law that governs it. Automation creates no new permissions: the work stays limited to publicly available information, and unauthorized access, deception and harassment remain outside legitimate practice, whether a person or an AI tool performs them. Data-protection rules such as the EU General Data Protection Regulation apply to personal data even when it came from an open source, so investigators need a lawful basis, a defined purpose and proportionality in what they collect and keep. An AI tool can gather far more than a person would, so restraint matters more: collect what the objective requires and keep what the case justifies.
The Berkeley Protocol on Digital Open Source Investigations sets out professional standards for gathering, analyzing and preserving open-source information, and work done with AI should be held to the same standards as work done by hand. The legal and ethical framework in the OSINT guide covers the subject at length. This section is general information, not legal advice: what is permitted depends on your jurisdiction, your purpose and your professional context.
How Nexus does it
Nexus is an agentic OSINT platform for investigators in the public and private sectors. You describe your objective in your own words, pick one of 47 investigative Skills or run a direct lookup. Nexus plans the research steps, runs them against public sources and returns findings with citations.
A citation opens a retained evidence record with the source name, the source type and when it was retrieved. Each Skill states its inputs, what it returns and its limitations; the Skill Library lists them all.
Four kinds of finding, kept apart
Nexus keeps these apart so you can see what each statement rests on.
- Supported findings
- Statements backed by retrieved evidence you can open, with the source and its context preserved.
- Possible matches
- Candidates that resemble the subject, such as a shared name, username or image, and have not been corroborated.
- Contradictions
- Places where sources disagree, kept side by side so you can weigh them yourself.
- Information gaps
- Questions the available evidence could not answer.
What Nexus will not claim
A match is a lead.
A shared name, username or photo stays a possible match until you corroborate it.
Nothing found is not a clearance.
It means nothing was found in the sources searched, on that date.
A citation shows where a statement came from.
It does not make the statement true.
A custody log and a report hash record how case data was handled.
A custody log exists only if you enabled it when the case was created, and a court decides admissibility.
Each Skill run is a point-in-time check.
Nexus does not monitor a subject or send alerts.
The investigator decides
Nexus runs the research. You verify the findings and answer for the conclusions. Nexus does not decide identity, ownership or wrongdoing.
Before you subscribe
Nexus has no free plan and no trial. To judge it before you subscribe, watch a recorded investigation or read on each Skill page what the Skill returns and what it cannot establish. Pro costs €59 a month with 150 credits a month, and Agency costs €89 a month with 300 credits a month. You subscribe on the pricing page and then create your account with the same email address.
Questions people ask about AI OSINT
What is AI OSINT?
AI OSINT is open-source intelligence in which artificial intelligence does part of the work: summarizing a lookup, extracting names from a result, drafting a report or planning and carrying out the research steps. The term covers several kinds of tool, from a lookup with an AI summary to a general AI chat assistant to an agentic OSINT platform. The sources stay the same: publicly available information that anyone can lawfully access. When comparing AI OSINT tools, check whether you can inspect the retrieved evidence, review the gaps and direct the follow-up work.
Is AI OSINT the same as agentic OSINT?
No. AI OSINT is the wider term, and agentic OSINT is one kind of it. In agentic OSINT an AI agent plans and carries out a multi-step investigation against an objective the investigator defines, cited findings link to retained source records, and the investigator reviews, directs and decides. Other AI OSINT tools summarize one lookup at a time or answer one question in a chat. Nexus is an agentic OSINT platform.
Can AI identify a person from public data?
AI cannot prove a person’s identity from public data. It can surface profiles, usernames, images or domains that may be associated with a subject, and it can show the evidence for and against each association. A matching name, a reused username or a similar image is a possible match, not proof of identity, ownership or control. The investigator decides what the evidence supports after corroboration from independent sources. Nexus does not decide identity, ownership or wrongdoing.
Can I use a general AI chat assistant for OSINT?
A general AI chat assistant is useful for orientation before a case begins. Research and citation features vary from one assistant to the next, so check each important claim against a source you can open. For an investigation, check also whether the tool retains evidence you can inspect, preserves case context and makes its limitations visible. A summary that cannot be traced to evidence is an opinion, and an opinion does not belong in a report that someone else will review.
Are AI OSINT tools legal?
Using an AI OSINT tool is lawful when the work is limited to publicly available information and stays within applicable law, the same standard that governs any open-source intelligence. Automation creates no new permissions. Data-protection rules such as the GDPR still apply to personal data gathered from open sources, so the investigator needs a lawful basis, a defined purpose and proportionality. Unauthorized access, deception and harassment remain outside legitimate practice, whether a person or an AI tool performs them. This answer is general information, not legal advice.
Does Nexus have a free plan?
No. Nexus has no free plan and no trial. To judge it before you subscribe, watch the recorded investigation on the home page, read on each Skill page what the Skill returns and what it cannot establish, or request a demo. Nexus has two self-serve plans: Pro at €59 a month with 150 credits a month and Agency at €89 a month with 300 credits a month. Plans are billed monthly or yearly, and you can cancel at any time and keep access until the paid period ends.
Sources
The definition, the law and the protocol this guide refers to, each on its publisher’s own page.
- National Defense Authorization Act for Fiscal Year 2006, Public Law 109-163, section 931 (open-source intelligence)GovInfo, U.S. Government Publishing Office
- Regulation (EU) 2016/679, General Data Protection RegulationEUR-Lex, Publications Office of the European Union
- Berkeley Protocol on Digital Open Source Investigations (2020)Human Rights Center, UC Berkeley School of Law, with the UN Human Rights Office (OHCHR)
Links checked on 11 November 2026.
