AI that respects the privilege you swore to protect.
The duty of confidentiality is not a setting. It is the architecture. Aethos is built so that a contract draft, a client memo or a brief never reaches a model that any other firm — or anyone else — can access.
Per-matter walls. Conflicts-gated retrieval. A model the firm owns, configured under the engagement letter your client already signed.
You.
A portrait of the firmA practice bound by professional secrecy.
You are a commercial law firm, a barristers' chambers, an in-house legal department, a litigation boutique, or a niche regulatory practice. The work product is privileged. The duty of confidentiality is absolute and survives the engagement. Anwaltsgeheimnis, secret professionnel, attorney-client privilege — different vocabulary, same rule.
You read the Bar's guidance note on generative AI when it came out. The associates already use ChatGPT informally and the partners know it. Your IT counsel has flagged the exposure. Your managing partner wants a defensible answer for the next professional-indemnity renewal — and a knowledge engine that finally beats a thirty-year-old document management system.
What's on your desk today.
Three pressures · privileged matterEach of these is being raised at partners' meetings right now — and at most firms there is no clean answer yet.
Pasting client text into a public chatbot is, on most readings, disclosure.
Whether the duty is the SRA's Principle 7, the Anwaltsgeheimnis of §54 RAO, or the common-law privilege of an EU member state, the analysis converges: passing client material to a third-party processor without the client's informed consent is at minimum a confidentiality breach and at worst a waiver of privilege. Several Bars have already issued opinions saying so.
One model that has seen every firm's data is the wrong architecture.
Even where a vendor offers a "private" tenant, the model itself was trained on a corpus that includes opposing counsel, adverse parties and matters you are not supposed to see. A shared model is a structural conflict-of-interest your ethics committee will struggle to clear.
Thirty years of memos, briefs and judgements are unsearchable.
The firm has paid associates to write five thousand opinions over twenty years. The next associate cannot find the relevant one in under an hour. Knowledge management has been the unsolved problem of legal practice for a generation. Off-the-shelf legal-tech indexes the surface — Aethos retrieves the reasoning.
What Aethos changes.
Privilege intact · matter walls · institutional memoryThree answers, mapped one-to-one to the three pressures above.
The model is part of the firm.
Aethos runs on the firm's own hardware, under the firm's engagement-letter terms. The model is software the firm licensed — not a service the firm subscribes to. No inference call ever leaves the perimeter. From a privilege analysis it is no different from running a document management system.
The conflicts file gates every query.
Aethos enforces matter-level access control. A junior on Matter 2451 cannot retrieve from Matter 1892 — even if she asks the same question. The retrieval layer reads the firm's conflicts file and the matter team list before returning a single source. The ethical wall is part of the query, not a policy hoping to be obeyed.
Past matters become compounding capital.
Every memo, brief, draft, judgement and opinion the firm has ever produced becomes searchable by question, not by keyword. With every answer citing matter ID, document and paragraph. The associate finds the right precedent in three minutes; the partner sees the firm's collective reasoning on every novel question.
The conflict check that started the matter must hold for every query about it.
The modules that matter most for you.
Where law firms startTwo modules carry almost the entire value for a legal practice. The remaining Aethos modules are largely irrelevant to law firms, and the page does not pretend otherwise.
The institutional memory engine.
Past memos, draft pleadings, opinion letters, board minutes, deal documents, settlement papers — indexed and answered by the firm's own model, with the source matter, document and paragraph cited on every answer. Conflicts-gated retrieval enforces matter walls at query time. Document corpus: yours. Reasoning style: the firm's. Citations: precise.
RAG module Module 02 · Aethos VoiceDictation, depositions, client interviews.
Partner dictation in any language, deposition transcripts, client interview notes — recognised and rendered on the firm's own infrastructure. Multilingual. Speaker-separated. The transcript lives in the matter file from the second it is spoken. No third-party transcription service hears a single word of the privileged conversation.
Voice moduleThe rules that bind.
What the rule asks · what Aethos providesProfessional rules, statute and regulation that bear on AI in legal practice — and where Aethos plugs into each.
| Source of duty | What it requires of AI use | What Aethos provides |
|---|---|---|
| Attorney-client privilege / Anwaltsgeheimnis | Privileged communications must not be disclosed to third parties without informed client consent. | No third party. The model is licensed software the firm operates. No inference call leaves the firm. Privilege analysis identical to an internal DMS. |
| Bar / SRA / RAO / RStBerG | Confidentiality (SRA Principle 7), competence in technology, supervision of staff using AI. | Per-matter access control · audit log of every prompt and response · supervisor-visible activity per associate · documented model-governance procedures. |
| EU AI Act | "Administration of justice" use cases (Annex III) are high-risk. Logging, human oversight, transparency, technical documentation. | Append-only signed audit log · per-skill human-in-the-loop policy · technical documentation shipped per release · model card per deployed model. |
| GDPR Art. 6, 9, 22 | Lawful basis (legitimate interest under engagement), special categories (matters involving health, criminal data), automated individual decision-making restrictions. | No data egress. Per-matter purpose binding. Audit log produces the explanation trail for Art. 22 challenges. |
| Conflicts of interest rules | The firm must not act against the interests of a current or former client; ethical walls must be effective. | Conflicts file gates retrieval at query time. A matter wall enforced by the database is a wall by construction, not by policy. |
| Professional-indemnity expectations | PII insurers increasingly require evidence of AI governance, training records and incident response. | Governance documentation pack, signed change manifests, incident-response runbook, training-records integration with the LMS. |
The next step.
One day · on premises · written outcomeBook a one-day Architecture Workshop.
One day at your firm with Kristijan Stojanović — founder of STK Engineering — and the architect assigned to legal practice. We work through priority use cases, the conflicts model, the matter taxonomy and the document corpus, and produce a signed sizing & integration plan you can present to the management committee.