Can You Ask AI Questions About Public Records? How AI Chatbots Work for Government Documents

Public-records investigations often begin with a simple question and a very large pile of documents.

A city may produce thousands of emails. A county may release contracts, invoices, spreadsheets, meeting minutes, and scanned PDFs. A state agency may provide records in multiple batches with inconsistent filenames, duplicate documents, and attachments scattered across folders.

The natural question is:

Can I just ask an AI chatbot questions about all of this?

Increasingly, the answer is yes.

But there is an important difference between an AI chatbot that can produce an answer and a public-records research system that can show why that answer is supported by the records.

For journalists, attorneys, auditors, researchers, watchdogs, and other public-interest investigators, that distinction matters.

What is an AI chatbot for public records?

An AI chatbot for public records allows a researcher to ask natural-language questions about a collection of government documents.

Instead of searching only for an exact keyword, a user might ask:

• Who communicated with this vendor before the contract was awarded?
• When did officials first discuss this property?
• Which records mention a particular address?
• What payments were made under this agreement?
• Did the explanation for a project change over time?
• Which documents refer to both a public official and a contractor?
• What happened before and after a particular meeting?
• Are there records that appear to contradict one another?

A useful AI public records search system can help locate relevant documents and passages even when the wording in the records is different from the wording in the question.

That can make large document collections much easier to explore.

But finding a relevant passage is only the beginning.

How AI chatbots search government documents

A typical AI document-analysis workflow has several stages.

1. Documents are collected and processed

The records may include:

• PDFs
• scanned images
• emails
• Word documents
• spreadsheets
• contracts
• agendas
• meeting minutes
• invoices
• procurement records
• property records

If a document is scanned rather than digitally searchable, optical character recognition, or OCR, may be used to convert the page into searchable text.

2. The text is indexed

Once text is extracted, the system can create a searchable index.

Traditional search can find exact words and phrases.

Semantic search can go further by identifying passages that are conceptually related to a question even when they do not use the same wording.

For example, a search for “contract extension” might surface records referring to an “amendment,” “renewal,” or “term modification.”

3. The question is matched to relevant records

When a user asks a question, the system searches the collection for documents and passages that appear relevant.

Those passages may then be provided to an AI model as context.

4. The AI generates a response

The chatbot uses the retrieved material to produce a response.

This is where the workflow can become either extremely useful or dangerously misleading.

The critical question is:

Can the user inspect the evidence behind the answer?

A chatbot answer is not automatically a fact

AI can summarize text, recognize patterns, and connect related passages quickly.

But an AI-generated statement should not automatically be treated as a verified fact.

Consider a records collection containing:

• an email mentioning a contractor;
• an invoice from the same contractor;
• meeting minutes discussing a related project; and
• a later contract amendment.

An AI system might identify those records as potentially related.

That is useful.

But the fact that the same entities appear across several documents does not automatically prove a particular relationship, motive, or sequence of events.

The system may have identified a lead.

A human investigator still needs to examine the underlying records and determine what they actually establish.

A defensible research workflow should therefore distinguish among:

• Verified fact: directly supported by the records.
• AI summary: a machine-generated description of the material.
• Inference: a conclusion drawn from multiple pieces of evidence.
• Investigative lead: something worth examining further.
• Conflicting evidence: records that appear inconsistent and require review.

That distinction is especially important in investigative journalism and public-interest research.

Why citations matter in AI public records search

Imagine an AI chatbot tells you:

> The agency began discussing the vendor several months before the procurement process started.

That could be important.

But before using it, a reporter needs to know:

• Which document supports that statement?
• Who sent the communication?
• When was it sent?
• What exactly did it say?
• Was the vendor actually being discussed in connection with the later contract?
• Are there other records that provide different context?

Without a citation or source link, the answer is little more than a lead.

With citations, the researcher can move from the AI-generated response back to the underlying evidence.

A strong AI chatbot for public records questions should therefore make it easy to inspect the source documents and relevant passages behind important answers.

Public-records questions are often timeline questions

Many significant findings do not appear in a single document.

They emerge from sequence.

For example:

1. An official exchanges emails with a vendor. 2. A project appears on a meeting agenda. 3. A procurement process begins. 4. A contract is awarded. 5. Payments begin. 6. The contract is amended. 7. Officials later describe the history of the project differently.

Each individual record may appear routine.

The investigative value comes from seeing the events together.

An AI system can help surface dates and related documents, but the timeline should remain traceable to the records supporting each event.

This is one reason public-records investigations require more than simple question answering.

Entity relationships require evidence

AI systems are also useful for identifying recurring entities such as:

• people;
• companies;
• agencies;
• properties;
• addresses;
• projects;
• contracts;
• account numbers; and
• payments.

But entity co-occurrence is not the same as a relationship.

If Person A and Company B appear in the same email, that does not automatically establish a business relationship.

If two companies share an address, that may be significant, or it may have an innocent explanation.

If a person appears repeatedly in procurement documents, that may justify further research, but it does not establish misconduct.

A good investigative system should help researchers identify possible connections while preserving the distinction between documented relationships and potential leads.

AI can help with FOIA document analysis

The same principles apply to federal Freedom of Information Act productions and state or local public-records responses.

A large FOIA release may contain thousands of pages spread across:

• correspondence;
• attachments;
• reports;
• spreadsheets;
• presentations;
• policy documents; and
• scanned records.

Traditional manual review can be slow.

An AI-assisted FOIA document analysis workflow can help researchers:

• search across the entire production;
• identify recurring names and organizations;
• locate dates and events;
• compare similar documents;
• identify potentially relevant passages;
• summarize document groups; and
• generate follow-up questions.

The advantage is speed.

The requirement is verification.

AI should help investigators decide where to look next, not eliminate the need to review the original record.

Questions AI is especially good at helping investigate

AI can be particularly useful for exploratory questions such as:

Who appears repeatedly?

This can surface frequently mentioned officials, vendors, attorneys, consultants, organizations, or property owners.

When did something begin?

Search across emails, agendas, contracts, and other records can help identify early references to a project, vendor, decision, or controversy.

What changed?

Comparing contracts, policies, statements, budgets, or drafts may reveal changes in language, amounts, dates, or responsibilities.

Which documents discuss the same subject?

Semantic search can locate related passages even when terminology differs across departments or authors.

What should I investigate next?

Patterns across records can suggest additional questions, agencies, datasets, or document categories worth requesting.

These uses make AI valuable as a research accelerator.

They do not turn the chatbot into the final authority.

Questions that require particular caution

Some questions invite conclusions the records may not actually support.

For example:

• Did this official favor this company?
• Was this contract corrupt?
• Did these people coordinate secretly?
• Was this payment improper?
• Did someone intentionally mislead the public?

An AI system may find records relevant to those questions.

It should not transform incomplete evidence into a factual accusation.

Instead, it should help identify:

• the relevant communications;
• contracts;
• payments;
• dates;
• decision points;
• conflicting statements; and
• missing records.

The human investigator remains responsible for determining what can actually be established.

What to look for in an AI chatbot for government documents

If you are evaluating an AI tool for public-records research, ask several questions.

Can it work with the formats agencies actually produce?

Government records rarely arrive as a clean set of searchable PDFs.

The system should be able to handle scanned material, spreadsheets, email exports, attachments, and other common formats.

Can it search across the entire collection?

Investigations frequently require following a name, address, vendor, project, or event across many documents.

Does it provide citations?

Important findings should point back to the original record.

Can you inspect the underlying passage?

A citation is more useful when the investigator can immediately see the relevant text in context.

Does it preserve uncertainty?

A lead should remain a lead until the evidence supports a stronger conclusion.

Can it support timelines and relationships?

Document collections often become more understandable when communications, decisions, contracts, payments, and other events are viewed chronologically.

Does it preserve the investigative record?

Editors, attorneys, collaborators, and other reviewers may need to understand how a conclusion was reached.

The strongest systems support an auditable trail from question to evidence.

Where DoTheTell fits

DoTheTell is being built specifically around public-records investigations where source traceability matters.

The goal is not simply to place a generic chatbot on top of a folder of PDFs.

The intended workflow connects:

• document ingestion and OCR;
• keyword and semantic search;
• question answering;
• source citations;
• entities and aliases;
• timelines;
• document comparisons;
• investigative leads;
• reviewable findings; and
• evidence-backed relationships.

The system is designed around a simple principle:

The finding should remain connected to the record that supports it.

That means an AI-generated observation should not silently become a verified fact.

A recurring name should not automatically become a relationship.

A suspicious pattern should remain an investigative lead until additional evidence supports it.

And conflicting evidence should be surfaced rather than forced into a single narrative.

DoTheTell is still a developing platform and should not replace reporting judgment, legal review, fact-checking, or verification against original records.

The purpose of AI is to make a difficult body of records easier to investigate.

So, can you ask AI questions about public records?

Yes.

And for large document collections, asking questions in natural language can be enormously useful.

But the more important question is not:

Can the AI answer?

It is:

Can you verify the answer?

For public-records investigations, the best AI systems will not simply produce confident summaries.

They will help investigators move quickly from a question to the relevant documents, passages, entities, dates, and events — while preserving the evidence needed to determine what the records actually prove.

That is where AI becomes genuinely useful for investigative research.

Not by replacing the investigator.

By helping the investigator find the evidence faster.

Explore public-records analysis, FOIA document analysis, and DoTheTell for journalists.

For a broader comparison, see document analysis tools for investigative journalists.