Complex legal work, run as steps you can check
Long documents, several jurisdictions, a sequence of tasks. Justinian, the engine behind HAQQ, breaks the matter into steps, sends each to the right tools, and returns one result with every source and action for your review.
Can legal AI handle complex, multi-step legal work?
HAQQ runs it as a controlled sequence. Justinian identifies the intent, jurisdiction and governing rules, applies permissions, brings in statutes, case law and your precedents, splits the task into retrieval, reasoning, drafting and verification steps, and compiles one result with sources attached. The judgment, the advice and the signature stay with you.
- Decisions Justinian resolves before any frontier model is called
- 10
- Planted issues a 3-agent swarm caught, against 3/5 for one prompt, in a mock-calibrated experiment
- 5/5
- Jurisdictions with jurisdiction-aware answers
- 200+
- Legal teams using HAQQ
- 17,000+
One request, a controlled sequence
Each stage shows on screen as it runs, and every action stays reviewable by a lawyer.
- 1
Justinian reads the matter first
Before any model writes, Justinian identifies the intent, jurisdictions, entities, language and matter, then applies permissions, confidentiality, matter isolation and residency rules.
- Intent, jurisdiction and entities identified
- Matter isolation applied before anything runs
- Ten decisions resolved before the model call
- 2
The task is split into steps
Diligence, a playbook review, a research memo: these are sequences. Justinian decomposes the work into retrieval, reasoning, drafting and verification steps, and sends each step to the model best suited to it.
- Find, extract, test, reconcile, draft, check
- Routed by capability, jurisdiction, security and economics
- Sub-agent orchestration with a human in the loop
- 3
Long documents, read across the whole file
Point HAQQ at a data room of hundreds of documents. It extracts the dates, parties, change-of-control triggers and obligations into one table, and tells you which contracts need a closer look before the deal closes.
- Hundreds of documents in one pass
- Answers point back to the document and section
- Contracts that need a closer look, flagged
- 4
Compiled, cited and handed back to you
Justinian compiles the steps into one result. Sources, citations, outputs and actions come back with traceability for your review, and where a point is unsettled, HAQQ says so instead of guessing.
- One result from many steps
- Every source and action traceable
- Unsettled points flagged openly
What a complex matter asks of a legal AI
Long documents
Multi-Review analyzes hundreds of documents at once, and Data Room answers point back to the source document and section.
Cross-jurisdiction reasoning
When a question crosses borders, HAQQ answers it under each law from that jurisdiction's own sources, across 200+ jurisdictions.
Multi-step orchestration
Tool routing, sub-agent orchestration, memory, guardrails, deep reasoning, lookahead and rollback, with a human in the loop.
Citation accuracy
Every answer shows the steps it took and the sources it used, so the check is quick to do and easy to show.
Projects that remember
Every chat inside a project shares the same memory and files. Drop a brief in once and every agent knows about it.
Runs that finish when you leave
Anything still running when you close the tab finishes and picks back up on its own.
Sequences teams hand to it
HAQQ does the associate work. The signature stays yours.
Justinian is built never to act autonomously and never to replace legal judgment. Every action stays reviewable by a human professional.
Read about security- Compliant with SOC 2, ISO 27001, ISO 42001 and GDPR
- Permissions, matter isolation and residency applied before anything runs
- Your data is never used to train HAQQ's models
- Private and sovereign deployment for enterprises
Complex legal work questions
Does HAQQ act on its own?
No. Justinian is built never to act autonomously and never to replace legal judgment, and every action stays reviewable by a human professional. HAQQ does the associate work. The judgment, the advice and the signature stay with the lawyer.
Which agents does HAQQ use?
The changelog names them: the Paralegal, Associate and Partner agents, the Legal AI Twin, and, since April 2026, a Justinian agent for complex legal reasoning on Business and Enterprise. In August 2026 the agent selector was removed, so starting a new chat is one less decision.
How large a file can it work through?
HAQQ works across data rooms of hundreds of documents, and the diligence demo on this site runs on 340. Inside a project every chat shares the same files and memory, so each step builds on the last.
How do you know it works on complex matters?
Our post on complex legal work sets out how HAQQ evaluates it: long documents, cross-jurisdiction reasoning, multi-step tasks, citation accuracy and calibrated refusal, each reported separately. Exact items are scored with no model as judge, open-ended answers are read by a human, and answer keys are checked by a lawyer. The post contains no scores.
What did the swarm experiment show?
On the same model and the same 30-document data room, a single prompt caught 3 of 5 planted issues and a 3-agent swarm caught 5 of 5. The write-up discloses that the numbers are mock-calibrated to demonstrate an open-source harness, which is why we cite it for how work is structured and never as a product score.