A records request is a deadline bolted to a filing cabinet. The documents already exist; the expense is finding every one that answers the question and signing the claim that the set is complete. Buyers described the same scene repeatedly: a spreadsheet produced by hand every time a regulator shows up, and the clock starting the minute they ask.
Records, Discovery & Evidence
How Do You Respond to Records, Discovery
and Due-Diligence Requests With AI?
Where AI turns a manual dig into search-and-verify, what it never certifies for you, and the two archive problems to test before a deadline depends on them.
Built from real buyer questions in our sales meetings
Direct Answer
Answer the request as a retrieval problem, not a drafting problem. The responsive material already exists, so the cost sits in the finding. A governed corpus that cites a source on every answer turns the dig into search-and-verify. Iternal builds AirgapAI to retrieve over a selected data set and reveal the source behind each answer, and Blockify to prepare that corpus so every block traces back to its own file.
The limit: retrieval produces candidates, and a records request is judged on completeness. Those are two different standards, and nothing here certifies that every responsive item was found. The human sign-off is the deliverable rather than a formality, which is why Iternal advocates a human in the loop on AI output.
Two archive problems decide whether any of it works on your material. AirgapAI answers over documents already prepared into a data set, and scanned PDFs of old paper need a separate extraction step first. An agent turned loose on a machine can also surface an old, outdated version of a file as its answer. Ask what the tool does when the responsive document is the one it cannot open.
Answering a request and becoming the subject of one are separate questions. For more information on whether AI prompts and answers enter the legal record, visit the AI records and privilege page. For more information on proving an answer against its source, visit the answer accuracy page.
What the Request Actually Costs
The bill is paid in staff hours, and it lands on the people you can least spare. Quality teams described audit preparation eating a very significant share of the week; city staff get pulled in because residents cannot find public records on their own. Three characteristics separate the work from ordinary search:
- The clock is external. You choose neither when the request arrives nor how long you have.
- The evidence standard is tangible. Compliance answers must be backed by evidence rather than theoretical and by the book, as one buyer put it.
- The downside is asymmetric. A fine for failing to produce the information runs higher than any quote for the work.
Search and Verify: What a Governed Corpus Changes
Retrieval helps only if the person signing the response can open every candidate at its source. A research tool suggests; an evidence tool shows its work.
Where the citation lives. AirgapAI retrieves over the data set you enable, and each answer carries a block icon; hovering it reveals the source citation. Full file traceability loads the source file and its raw text, and the PDF view scrolls to the highlighted passage. Blockify stores the citation with the file location and hierarchy.
Why the standard is higher here. A records or due-diligence response is a legal artifact: an entity formally commits to the answers it supplies in a compliance questionnaire, and a production is read by a party whose job is to find what you left out. Iternal ran the same play on its own paperwork, preparing its procurement forms, questionnaires and due-diligence documents as a governed data set. For more information, visit the answer accuracy page.
The Two Failure Modes to Test Early
An archive punishes optimism. Both behaviors below are cheap to test long before a filing date rides on them.
Scanned paper is a separate problem from search. AirgapAI answers over documents already prepared into a data set, and scanned PDFs of old documents sit outside what it reads. The resolution lives in the pipeline: Iternal extracts from complex PDFs with multimodal models where ordinary OCR services fail, and Blockify turns an OCR-processed scan into structured content the assistant can query, with variants tuned for legal documents and court records.
An agent left to search alone can hand you the wrong version. An agent searching a laptop may pick an old, outdated version of a file as the answer, and where two source documents say different things about the same fact, Blockify needs a person to choose which one governs. A superseded document is worse than no answer, because it leaves with a signature on it.
Scope matters as much as capability: Iternal has done this record-set document work in veteran service records, and law-enforcement files would be new ground.
Pin it down: questions for your evaluation
-
Which scanned archive formats can Iternal extract today, and what is the acceptance test on a sample set we choose?Whether the responsive documents can be read at all, long before a deadline depends on it.
-
Where two documents say different things about the same fact, what does the system return and who decides?The version rule for your response, written down rather than assumed.
-
Which record types has Iternal delivered this document work in, and which would be a first engagement?The difference between a proven workflow and a pilot, on the record before anyone commits to a date.
Meeting Records, Minutes and the Answer Nobody Can Find
The same corpus solves the problem sitting beside it. Finding an answer in meeting minutes means reading many pages of PDFs or listening to a recording end to end; commission agendas carry a very large page count; residents search years of PDFs to learn when a decision was made. Loaded as a data set, that material answers with the meeting and date attached — a board meeting on a given date, returned as the citation itself. Agendas from other boards can join the same pool.
One design decision determines how far the result can be trusted. A set of minutes is one person's summary, and the verbatim transcript of the same meeting can record the moment differently. Loading both moves the burden onto vetting before upload. Public deployments work that way: staff vet source files before upload, and the site states plainly that answers may not reflect the most current information.
Accessibility Remediation: The Same Work, With a Statutory Deadline
Higher education runs the same workload with a date attached. University teams described a federal requirement that leaves every student-facing PDF a liability until it is remediated, and the work is done today by clicking around buttons in a SaaS tool. At smaller institutions no dedicated remediation group exists, so the faculty member does it. Off-the-shelf products have produced weak results, and complex tables and highly visual student media are hardest of all to sign off on.
Iternal has scoped the work as a pipeline — an ingestion layer, a validation engine and a remediation audit trail — that can run as a file share where documents go in and come back remediated, as PDFs or as an HTML conversion. The limit here is scope rather than ambition: Iternal has designed the use case and has not yet delivered it. University buyers said their blocker is verification of output quality, so pilot your hardest material and set the pass mark first.
Answered elsewhere
- What your executed agreements actually commit you to — see the contract portfolio page.
- Answering a solicitation from approved reusable content — see the proposals and solicitations page.
- Recurring reports, the close and quoting done by hand — see the reporting and back-office page.
- Inventorying undocumented legacy systems — see the legacy systems and IT operations page.
- Whether AI prompts and answers enter the legal record — see the AI records and privilege page.
- Grounding an answer and proving it against the source — see the answer accuracy page.
- Capturing speech on the device in the first place — see the languages and voice page.
FAQ
FAQ: Records, Discovery and Evidence Production
It converts a manual dig into search-and-verify. Load the material into a governed data set, ask what the requester asked, and candidates come back carrying a citation showing where each came from. Staff rule on responsiveness instead of hunting.
No, and no tool should claim otherwise. Retrieval returns candidates ranked by relevance; a records response is judged on completeness, a different standard. Iternal advocates a human in the loop, and a production headed for a regulator is the clearest case.
They need an extraction step first. AirgapAI answers over documents already prepared into a data set, so scanned paper sits outside what it reads. Iternal handles that upstream with multimodal extraction where ordinary OCR services fail, and Blockify structures the result.
Open it at the source. Each AirgapAI answer carries a block icon revealing the source citation on hover; full file traceability loads the original file and its raw text, and the PDF view scrolls to the highlighted passage.
Yes, and the citation makes it usable: answers name the meeting and the date, such as a board meeting on a given date, rather than a page somebody must then locate. Decide which artifact governs when a summary and a verbatim transcript describe one meeting differently.
Iternal has designed the pipeline — ingestion layer, validation engine and remediation audit trail, delivered as PDFs or HTML — and has not yet delivered the use case. University buyers said their blocker is verification of output quality, so pilot your hardest tables and visual media.
Test It on the Request You Answered Last Month
The cheapest evaluation available is a request your team has already satisfied. Load the same material, ask what the requester asked, and compare the result against what your people found by hand. Two things matter: what it missed, and whether every item can be opened at its source.