Ask any litigator what makes evidence admissible and they'll give you two words: chain of custody. It isn't enough for a thing to be true — you must be able to trace every hand it passed through, back to its origin, or it doesn't exist in the eyes of the court.
Now hold that standard up to the AI tools being sold into law firms, and you'll see the problem Asideus was born to solve: when a probabilistic answer arrives with no provenance, why would a profession built on chain of custody ever trust it? The honest answer is that it shouldn't — and its risk teams don't. Which is why generic AI assistants stall at the door of every serious firm.
Asideus is what happens when you accept the profession's standard instead of fighting it.
The Client
Asideus AI is building an AI legal operating system for UK mid-market law firms and in-house teams, on two commitments that define everything downstream. The first is Silent Capture: instead of another point tool lawyers must remember to open, the platform integrates invisibly into the workflows they already live in — calls, email, documents, file shares — and turns that everyday exhaust into structured, queryable, auditable case knowledge. Think of it as the associate who's always in the room, always taking perfect notes, and never bills a minute. The second is Explainable AI: every output must carry a verifiable reasoning trail and citations back to source documents. In this product, an uncited answer isn't a limitation — it's classified as a defect.
The Challenge: A Profession That Can't Accept "Trust Me"
UK legal practice sits in a specific squeeze. Matter knowledge lives in individuals' heads, inboxes, and call notes, so context bleeds away on every handover, holiday, and departure. Fee earners burn advisory hours on capture and admin, and recoverable time leaks into write-offs. SRA and GDPR expectations keep rising while manual audit trails stay fragile and expensive to reconstruct. And the existing vendor landscape solves slices — research here, review there, practice management elsewhere — with each slice demanding the one thing lawyers won't give: behaviour change.
When the buyer is a firm's risk lead, then the bar isn't "impressive" — it's "admissible." Under those circumstances, the platform had to be designed backwards from trust.
The Build: Trust as Architecture
Rocket Now delivered Asideus end to end — product design, frontend, database and security model, serverless AI backend, billing, and marketing site — as a multi-tenant platform measuring 54,878 lines of application TypeScript and 10,739 lines of serverless functions: 177 components across 51 pages, 42 edge functions, and a 43-table Postgres schema over 52 migrations. Ten product surfaces cover the full matter lifecycle: the Knowledge Hub document vault with AI metadata and semantic search, Cases with timeline reconstruction and GraphRAG multi-hop querying, Transcribe with speaker identification and automatic extraction of action items, deadlines, and billable time, Craft for AI-assisted drafting and redlining, the citation-mandatory Chat assistant, Scan intake OCR, automatic Billable Time capture, the Board with its approval queue, the proactive Inbox, and a self-built contracts and e-signature flow with a full audit log.
But the architecture is the story, because every major decision is the chain-of-custody principle expressed in code:
Tenancy lives in the database. Multi-tenant isolation is enforced by row-level security scoped by firm and role through security-definer functions — so a client-side bug can never become a cross-firm data leak. When confidentiality is the profession's first duty, then isolation must be a property of the schema, not a promise of the application.
Citations are mandatory, structurally. The RAG pipeline retrieves from embedded document chunks before generation and returns source references; every claim carries a "View Source" pathway to the underlying excerpt. When a lawyer can trace the answer, then the answer is usable; when they can't, the system treats it as inadmissible.
Humans approve every write. AI-suggested tasks enter an approval queue rather than the live board; drafted emails are composed, never sent. The AI proposes; the fee earner disposes.
Routing is deterministic. The agent orchestrator classifies intent through explicit keyword routing across five specialist agents run in parallel — reproducible and loggable, because in this profession even the routing decision may one day need a chain of custody.
Conflicts are surfaced, never silently corrected. When documents contradict each other on a date, an entity, an obligation, then the contradiction is raised as an alert with both excerpts shown — the system flags; it does not quietly pick a winner.
Stated Plainly: Where This Stands
Rocket Now's case studies claim what's verified and label what isn't, so here it is straight. Asideus is pre-launch, in its Pilot Partner Program — the engineering figures above are measured from the codebase, and there are no production outcomes yet because there are no production tenants yet. The rollout plan treats that honestly too: a bounded pilot with 3–5 firms, an evaluation harness with a golden-set question bank gating every prompt and model change, and instrumented metrics — time-to-context on handover, recovered billable time, citation trust rate — that will be measured, not assumed. And the AI layer currently runs on other model providers through a single gateway abstraction, with a registered, evaluation-gated plan to shadow-test Claude on the reasoning-critical workloads — long-context document analysis, multi-document conflict detection, case summaries — precisely because a product whose deliverable is the reasoning trail belongs on the model that reasons out loud. When the evaluation data exists, then the routing decisions get made — per workload, on evidence. Naturally.
The Results
- A complete, deployable multi-tenant legal platform — ten product surfaces spanning the full matter lifecycle — built and shipped by a small team, verified at 54,878 application lines, 42 serverless functions, and 43 RLS-protected tables
- Compliance-first architecture as delivered fact: UK data-residency posture, GDPR-by-design, database-enforced tenancy, and audit logging on signature and contract events
- Every AI surface constrained by design: cite sources or route through human approval — no autonomous writes to matter data, anywhere
- A measurement-first rollout in motion: pilot exit criteria include zero cross-tenant findings in penetration review and ≥95% of sampled AI claims traceable to a correct source excerpt
"[Quote from Asideus leadership — pending sign-off.]" — [Name, Title], Asideus AI
Why It Worked: The Right Step Was the Standard
The Rocket Now principle, proven again: a single step in the right direction is worth exponentially more than many steps in the wrong one. The wrong steps were available in bulk — ship a chatbot, demo the magic, fight the risk team later. The one right step was adopting the profession's own evidentiary standard as the architecture on day one: provenance mandatory, tenancy structural, humans on every gate. Every feature since has inherited that admissibility, which means Asideus never has to win the trust argument twice. When trust is the foundation, then everything built on it is load-bearing.
Selling into a profession where "trust me" is a defect? Build to their standard of evidence. We know how. → RocketNow.com
Editor's note — before publishing:
- Confirm authorization to publicly name Asideus AI and replace the placeholder quote
- Pre-launch status and "no production outcomes" framing are intentional and accurate — keep them; the modelled pilot projections stay out of public copy entirely
- Current model providers and the Claude evaluation plan are described per the deal-registration docs (no Claude in runtime today); don't let edits imply otherwise