AI disclosures & limitations

Last updated: September 7, 2026

Courts and bar regulators increasingly expect a lawyer using AI to understand what the tool actually does. This page exists so you can answer that — to a client, to a judge, or to your own professional-responsibility committee — without taking a marketing claim on trust.

1. Where AI is used, and where it is not

• Evidence extraction (SMS/MMS, iMessage, WhatsApp) uses no AI at all. It is a deterministic parser: the same backup always produces the same report. No model decides what your evidence says.
• Legal Matter AI uses large language models for every step except routing and billing, which are ordinary code.

2. Which models, and who operates them

Anthropic’s Claude models (Sonnet and Haiku tiers) via Amazon Bedrock, running in our AWS account in the United States. Your documents are not sent to a third-party AI vendor, and are not used to train any model. Document text is converted to numeric representations for search using Amazon Titan embeddings.

3. What the system is allowed to read

Each matter has its own separate index. The agents can read only the matter the run was started against. This is enforced outside the model — retrieval is bound to one authorised index and accepts no argument naming a firm or matter — so no instruction, prompt or document content can widen it. A document’s type has no say in what is read either: the analysis enumerates by document, so a misread filename cannot drop a document from an answer.

There is no external legal corpus. The system has no access to case law, statutes, or any source outside the documents you place in the matter. It cannot verify that an authority exists, because it cannot look one up.

4. Known limitations

• Language models can produce confident, fluent text that is wrong. Every factual statement in a report carries a citation to the passage it came from, so it can be checked — but the citation being present does not prove the passage supports the claim. Check the ones that matter.
• Every selected document is read in full — retrieval is not a search. The analysis enumerates every passage of every document you select, and the report states the count it read. That is what makes an absence meaningful: when the analysis says nothing in the record supports an element, it is reporting on the whole record it was given, not on a search that may have missed it. A selection too large to read in full is refused before you are charged, rather than silently sampled.
• What it reads is still only what you upload. Full enumeration says nothing about completeness of the matter. A document nobody added, an unproduced exhibit or a page missing from a scan cannot be read, and the analysis cannot know it is absent.
• OCR introduces errors. Scanned documents are read by optical character recognition, which misreads handwriting, poor scans, stamps and marginalia. Text the analysis never saw cannot be analysed.
• Dates and figures are reproduced as written, not normalised or inferred — which avoids invented precision but means an ambiguous date stays ambiguous.
• Analysis is bounded. A critic reviews the analysis before you see it, and work can be sent back for re-analysis a limited number of times. At that ceiling the run proceeds to attorney review carrying the unresolved critique, rather than continuing indefinitely. Re-analysis re-examines the same passages; it does not go looking for new sources, because there are none beyond the documents in the matter.
• The output is not a legal conclusion. Issues are surfaced as potential, with a confidence rating and the basis in your documents, for an attorney to weigh.

5. The controls that are always on

• A supervising attorney must approve before any report is written. The run halts and waits. There is no autonomous mode, no scheduled run that files itself, and no configuration that skips this gate.
• The approver is recorded on the report and in the audit trail, with the time of approval.
• The report carries its own disclaimer, in the document, so it travels with the file rather than living only in the interface.
• Contradictory evidence is never omitted. If none was found, the report says so explicitly.
• A separate critic reviews the analysis before an attorney sees it, and rejects work with uncited claims, unaddressed contradictory evidence, or conclusions the record only supports as possibilities.

6. Your professional obligations do not transfer

Using this service does not change your duties of competence, supervision, confidentiality or candour to the tribunal. In particular, you remain responsible for:

• verifying every fact, citation and conclusion before relying on or filing it;
• confirming that any authority referred to exists and says what is claimed;
• determining whether disclosure of AI use is required in your jurisdiction or by a standing order;
• ensuring you have the right to process the material you upload;
• supervising non-lawyer assistance, which includes this software.

7. No unauthorised practice of law

We are a technology provider, not a law firm. Nothing the Service produces is legal advice, no attorney–client relationship is formed with us, and the Service does not exercise professional judgement on your behalf. Output is intended for use by, or under the supervision of, a licensed attorney.

8. Questions

If you need something specific for a court, a client or an insurer — how a particular report was produced, which model versions ran, what the audit trail shows — contact us at ap@idtsol.com or +1 954-604-4098.