Bilbs AI
Private AI for Québec law firms

Your AI, running on your firm’s history.

A private operating system for institutional intelligence, racked in your law firm’s own server room and trained on your firm’s own files. Your lawyers query the firm’s memory instead of pasting client matters into ChatGPT. Nothing leaves the building. Ever.

See how it works
Never leaves the building Trained on your firm’s files We can’t see your data
[ The thesis ]

Your firm needs a private
AI. Not another chatbot.

Your lawyers already use AI. They’re typing client matters into ChatGPT. They’re asking Copilot about confidential files. None of it is logged. None of it is governed. None of it is yours.

We’re what gets built around it. A server in your office, trained on your files, governed by your IT director, with a complete record of every question every lawyer ever asks. Bilbs doesn’t compete with ChatGPT or Copilot — we’re the private layer your firm puts around them.

01

Private — by the wiring, not a policy.

All AI work happens on hardware in your office. The server can run with the internet cable unplugged. No outside company processes your data — not Microsoft, not OpenAI, not us.

02

Logged — every question, every answer.

Every question a lawyer asks is recorded — on a server you own, exportable to a spreadsheet. When a client asks “who looked at our matter and what did they ask?”, the answer is one query away.

03

Trained on your firm. Belongs to your firm.

The AI reads every contract, memo, and matter your firm has worked on — and learns your firm’s way of doing things. At the end of the lease, $1 buys the server outright, and the trained AI inside it.

B
Bilbs AI Groupe Bilbs inc. · Montréal
See how it works
[ Before / After ]

Two paths. One ends at La Presse.

Without Bilbs

A policy on the wall.

  • ✕A junior associate pastes a client matter into ChatGPT to summarize it. The text now lives on an OpenAI server in the United States. Nobody at the firm knows it happened.
  • ✕A senior partner spends 90 minutes searching the document system for a clause she remembers using three years ago. She gives up and rewrites it from scratch.
  • ✕A major client sends a vendor questionnaire asking where the firm’s AI processes their data. There is no truthful one-page answer.
  • ✕A 30-year partner retires. Her negotiation playbook, her precedent files, her client instincts — all of it walks out the door with her.
  • ✕The managing partner hears about a Loi 25 incident from a journalist at La Presse before he hears about it from the firm.
With Bilbs

A server in your building.

  • ✓The same junior asks your firm’s AI to summarize the same matter. The text never leaves your office. The question is logged.
  • ✓The senior partner asks “Find me that clause from 2022.” Two seconds. The exact paragraph, cited to the source file. She drafts in 15 minutes instead of 2 hours.
  • ✓The vendor questionnaire becomes a one-page answer: Our server. Our firewall. Audit log on request. No cross-border processing.
  • ✓When the 30-year partner retires, her precedents and her playbook stay in the firm’s AI — and her successors can ask the questions she would have answered.
  • ✓The managing partner answers any client, regulator, or partnership question without hesitation — because the audit log is on a server he owns and the data never left.
0 B
Bytes leaving the building
0%
Source-cited responses
1 page
To answer a vendor questionnaire
[ How it works ]
[ 01 - The flagship · a server in your server room ]

Meet the Bilbs Box.

A private-AI server, sized to the firm, trained on your files, racked in your building.

On-premise · 2026 Hardware · Firmware · Weights
BILBS BOX · 4U S/N · BB-26-042
from $29
/ lawyer / month
Audit-first
we scope what your firm needs
Hardware
yours, or we source it

An AI that actually
knows your firm.

Picture this. Every contract, memo, opinion, and brief your law firm has ever written — instantly searchable in plain English. A junior associate asks: “Find the limitation-of-liability clause we used for Acme in 2021.” Two seconds. The exact paragraph, cited to the file. A partner asks: “Draft a non-disclosure agreement using our standard terms.” Ninety seconds. Written in your firm’s actual voice.

None of this leaves your building. Microsoft can’t see it. OpenAI can’t see it. We can’t see it. The AI runs on a server we install in your law firm’s office, on a network your IT controls. The data on it is yours, period.

✓Knows your firm’s precedents, contracts, and matter history
✓Drafts in your firm’s own style — not generic AI English
✓Cites every answer to the exact source document
✓Logs every question — a complete record that stays on your server
✓Talks to your document system, your M365, your practice software
✓Only your lawyers can see it. Not Microsoft, not OpenAI, not us.
[ Sized to the firm ] 4 configurations

From a boutique practice of 10 lawyers to a full-service firm on an 8-GPU cluster - same software, same runbook. The firm outgrows a tier, we swap the chassis and re-train on the new hardware over a weekend.

Book a free call Download whitepaper
[ 01b - The moat · AI memory of the firm ]

What changes the day it goes live.

The defensible asset isn’t offline inference — open-source models will be everywhere. It’s the firm’s accumulated, indexed, searchable memory, and the platform that makes it usable. Six things change the morning the server is racked.

01

Find any document, in seconds.

A first-year asks, in plain English: “How did Me Tremblay usually structure acquisition clauses for construction disputes between 2018 and 2022?” The AI retrieves the prior drafts, summaries, and reasoning — with citations. Months of mentorship friction collapse into a single query.

02

Draft from your firm’s own work.

Ask for a first version of an NDA, a memo, a settlement letter. The AI writes it in your firm’s actual voice — because it learned from your firm’s actual work. Junior associates start at senior-quality first drafts.

03

Onboard new lawyers in days, not months.

A new associate joins. Instead of asking three senior partners how the firm does things, she asks the AI. It knows. It learned from every matter the firm has handled — and the partners get their time back.

04

Answer client questions on one page.

Big clients now ask: “Where is our data processed? Who has access? Is there an audit log?” With Bilbs, every answer is one sentence: our server, our firewall, yes — and we can export the log for you.

05

Keep the knowledge when partners retire.

When a senior partner retires, decades of judgment normally walks out the door. With the AI indexing their files, emails and meeting transcripts, the firm keeps the negotiation styles and tactical patterns, the preferred contract structures and clause libraries, the winning arguments by case type and jurisdiction, and the client history, preferences, and relationship context — including the wording, tone, and stylistic conventions that made the partner who they were. For firms carrying succession risk on their top rainmakers, this is worth multiples of the platform’s price.

06

Sleep at night.

Your client files aren’t sitting in Virginia, Dublin, or São Paulo, waiting for a foreign subpoena, a data breach, or a Loi 25 incident. They’re behind your firewall, on a server you control. The most stress-reducing thing a managing partner can buy.

[ How it works — in plain English ]
  1. 1. We install a server. A physical computer, racked in your office’s server room. Your IT team plugs in two cables. We come on-site and set it up.
  2. 2. It learns your firm. It reads every document, contract, and matter file your firm has worked on — through your existing document system. It never leaves the server.
  3. 3. Lawyers ask questions. They open it in their browser, the same way they open Outlook. They ask in plain English. Answers come back in seconds, cited to the source.
  4. 4. Nothing ever leaves. Not to Microsoft, not to Google, not to OpenAI, not to us. The server doesn’t phone home. It can run with the internet cable unplugged.
[ 01c - Use cases ]

Six things lawyers actually do with it.

Not theoretical workflows. The six concrete jobs the firm hires the Box for from week one — and stops paying overtime to do by hand.

Case transfers

Hand off a file in seconds.

A lawyer taking over a file asks for the full history, strategy, notes and key documents. The AI returns a structured briefing — cited — in the time it takes to pour coffee.

Legal research

Internal precedent in one query.

The AI retrieves relevant precedents, internal contracts, prior memos and indexed statutes. Every citation links back to the exact paragraph in the source file.

Meeting transcripts

On-prem transcription · Whisper.

Offline transcription and summarization of client calls, depositions and internal meetings on a self-hosted Whisper engine. The audio never leaves the building.

Institutional knowledge

Senior expertise, queryable.

Senior partner expertise becomes a queryable asset rather than a single point of failure. Negotiation patterns, client preferences, and clause libraries — all indexed, all searchable.

Contract drafting

First drafts in the firm’s voice.

AI-assisted drafting using the firm’s own templates, historical files and validated clause libraries. Junior associates start at senior-quality first drafts.

Paralegal leverage

Free the team for higher-value work.

Repetitive administrative and research tasks get absorbed by the AI. Paralegals stop running the same precedent searches a hundred times a year and move up the value curve.

[ A typical session — five steps ] From sign-in to source-cited answer
  1. 01 Authenticate Lawyer opens the application and authenticates against the firm’s directory.
  2. 02 Scope The AI loads the user’s permissions and scopes accessible files, emails, precedents, transcripts and contracts.
  3. 03 Ask “Find the previous supplier dispute we handled for Client X and summarize the strategy we used.”
  4. 04 Retrieve & cite The AI retrieves the relevant documents, cites the exact source files, summarizes the prior strategy, and proposes next actions.
  5. 05 Verify The lawyer clicks any citation to open the original source document. The human stays in the loop on the final work product.
[ 01c2 - The interface ]

What the lawyer sees.

A React web client. Opens in the browser, the same way Outlook opens. A chat panel on the left, the cited source on the right, every answer one click from the file it came from. No new tab to learn.

bilbs.firm.local · internal only
B
Bilbs · firm memory Scope: M&A · Construction · 2018–2024

How did Me Tremblay usually structure acquisition clauses for construction disputes between 2018 and 2022?

Across 14 construction-dispute acquisitions Me Tremblay handled in that window, three patterns recur: (1) a layered indemnity cap (8% / 15% / 30% by claim class), (2) a separate construction-defect basket with a 24-month survival period, and (3) a holdback escrow funded at closing rather than letter-of-credit.

Sources · click to open
Ask the firm’s memory…
SPA-Beauvais-2021-v3.docx DMS · iManage · Matter 21-0419
Open source
CLAUSE 7.4 — INDEMNIFICATION CAP (CONSTRUCTION DEFECTS)

7.4.1 — The Vendor’s aggregate liability under this Article 7 shall not exceed eight percent (8%) of the Purchase Price for claims arising from breaches of representations relating to general matters…

7.4.2 — Notwithstanding Section 7.4.1, the Vendor’s aggregate liability for claims arising under Section 4.18 (Construction Defects) shall not exceed fifteen percent (15%) of the Purchase Price, and shall be subject to a separate basket of $250,000 with a survival period of twenty-four (24) months from Closing.

7.4.3 — Claims arising under Section 4.21 (Environmental Liabilities) shall be capped at thirty percent (30%) of the Purchase Price, with the holdback established under Section 2.7 serving as the first source of recourse…

Logged · audit row #00471-A
On-prem only
Cited · no answer without source
SSO · RBAC at retrieval
Every query logged
[ 01c3 - Overlay, not replace ]

Not an ERP. The intelligence layer on top of one.

A deliberate product decision. Replacing your practice management is a long, capital-heavy fight against entrenched incumbents. We don’t pick that fight. We sit on top of what you already run.

Clio · Actionstep · ProLaw Practice management stays the source of truth. We read from it; we don’t replace it. Time entry, billing, and conflicts checks keep running where they already run.
iManage · NetDocuments · SharePoint · file servers Documents are indexed in place. No migration. The DMS your firm spent years tuning is still the system of record — Bilbs reads it, cites it, never copies it out.
Outlook · Gmail (local archives) Email archives are indexed for retrieval and search. The mailbox keeps living in M365 or Workspace — Bilbs adds queryability without touching the inbox itself.
PDFs · Word · scanned documents OCR (Tesseract / PaddleOCR) and the embedding pipeline handle every format. The scanned 1998 settlement letter becomes as searchable as last week’s memo.
Active Directory · Okta · Entra ID SSO is your existing directory. RBAC permissions follow the same matter-level access rules your DMS already enforces. No second password, no second permission model.

Why this matters commercially. Law firms strongly resist migrations, retraining, and system replacement. Selling “AI on top of what you already have” collapses the objection set, shortens the sales cycle, and delivers value within weeks instead of quarters. ERP-style modules — billing, conflicts, knowledge management — can be developed later, once the intelligence layer is entrenched.

[ 01d - Why now ]

Three forces, all converging at once.

Over the next 24 to 36 months, Québec firms will divide into two categories: those that operationalize AI internally, and those that don’t. The division is structural, not cyclical.

01

Loi 25 enforcement is intensifying.

Maximum fines run to $10M or 2% of worldwide revenue. Client questionnaires now ask where data is processed, who has access, and what the audit trail looks like. The cloud-AI answer doesn’t survive the questionnaire anymore.

02

Open-source models crossed the threshold.

Gemma 3 and Gemma 4 are now practically useful for legal drafting and retrieval-grounded research. The capability gap between cloud-only and on-premise has collapsed for the queries lawyers actually run.

03

On-prem GPU cost dropped sharply.

A firm of 30 lawyers can now run a serious private AI stack on a single chassis, financed at a smaller monthly than the photocopier. The capex objection that held the conversation back through 2024 doesn’t apply in 2026.

The window is narrowing. The firms that act now will define the standard their competitors are measured against. The ones that wait keep partner expertise locked in PDFs, watch Loi 25 exposure grow, and pay the research-hour tax every month.

[ 02 - Pick a size ]

Sized to the firm.

Boutique practice to national firm - one runbook, one upgrade path.

[ One line on the invoice ] From $29 / lawyer / month.

Starts at $29 per lawyer per month. One line on the partnership’s invoice. It goes up only if the firm chooses additional services — fine-tuning on a niche practice area, high-volume transcription, multi-office sync, expanded storage. The hardware is not on that line: we audit the firm and either use the server you already have, or we source and install one for you if you don’t have IT — your call after the audit.

Start with the audit
01

We audit the firm.

Document system, practice management, identity directory, network, and the GPU situation — if any. We tell you what your firm needs to run a private AI, in plain English. Free, on the first call.

02

Path A · you have a server.

If the firm already has GPU-capable hardware in the server room (or budget approved through your usual IT channel), we deploy onto it. The firm keeps the asset on its books. $29 / lawyer / month covers the rest.

03

Path B · you don’t.

No IT director? No server room yet? We source the hardware, install it on-site, configure the network, and own the runbook. Hardware sits on a separate, transparent quote — your call once you see the spec. Then $29 / lawyer / month.

Typical fit · 10–30 lawyers
Practice Desktop · water-cooled
10–30
Lawyers, comfortably

The hardware reference for most of the firms we work with. Sits in your own server room — yours if you have one, sourced and installed by us if you don’t. Two NVIDIA RTX 5090 graphics cards, a Threadripper PRO processor, 256 GB of error-correcting memory. Trained on every contract, memo, and matter your firm has handled.

On the partnership’s invoice
from $29 / lawyer / month
One line. Goes up only if you add services.
Audit, deployment, training, updates & support Included
Hardware (yours, or we source it) Separate quote
Below Microsoft Copilot, on your own server
See the Practice spec
Foundation 5–10 lawyers

The smaller hardware reference. Fits in a supply closet. For boutiques, notarial offices, and small in-house teams.

Hardware spec See details →
Firm 30–100 lawyers

The bigger reference. A proper rack-mount with two power supplies and hot-swap drives. For full-service firms.

Hardware spec See details →
National 100+ lawyers

A half-rack cluster of eight enterprise GPUs. For national Canadian firms and large in-house teams.

Hardware spec See details →

Not sure which? A 20-minute call with the managing partner and IT director is enough to scope it. The tier names map to hardware references, not invoice lines. Once the audit is done, you tell us whether you already have a server we can use, or whether we should source and install one for the firm — the per-lawyer line stays the same either way. If we source it, it’s a separate, transparent quote against the OEM’s catalogue: we don’t mark up the GPU. Hardware carries the OEM’s 3-year warranty; we shepherd any RMA.

[ What each tier actually handles ]
Capacity dimension Foundation Practice Firm National
Lawyers (comfortable) 5–10 10–30 30–100 100–1,500
Concurrent users (peak) 5–15 25–60 200–250 1,200–1,500+
Indexed corpus 5M+ docs 25M+ docs 100M+ docs Multi-office · cross-office sync
Live transcription streams 2 streams 6 streams 16+ streams Unlimited (per-node)
Redundancy posture RAID 1 · UPS RAID 10 · dual PSU Hot-swap · dual NIC Hot-spare server · multi-node failover
Deployment Single office Single office 1–2 offices Hybrid · per-office nodes + sync

Numbers above are comfortable envelopes, not theoretical maxima. We size against your firm’s actual corpus, peak-hour user count, and meeting load — not a glossy datasheet. If the firm outgrows the chassis, we swap to the bigger spec and re-train on the new hardware over a weekend; the per-lawyer line stays the same.

[ What’s in the $29 / lawyer / month ]
Day 0 — audit + deploy No upfront fee
Included

Firm audit, infrastructure review, on-site installation, indexing of your existing documents, identity directory wiring. Two on-site training sessions per practice group. Before $29 starts ticking, everything works.

Every month — run it $29 / lawyer / mo
Included

Updates, model refresh on new open-weight releases, security patches, monitoring, encrypted backups, audit logging, private Slack / Teams with IT, <4h response, 24/7 Sev-1 pager. The line goes up only if the firm adds services below.

Hardware path Separate quote, your call
Optional

We use the server the firm already owns if it’s GPU-capable. If the firm has no IT or no hardware yet, we source it on a separate, transparent quote — at OEM cost, with the OEM’s 3-year warranty, no GPU markup.

Everything that lets the firm answer a client AI questionnaire sits in the $29 line. The only thing it does not cover is the physical box itself — because half the firms we work with already have one in the rack, and the ones that don’t deserve to see the hardware quote on its own page.

Portable by design

Firm outgrows its tier? We swap the chassis for the difference in list price, not a full rebuy, and re-train on the new hardware over a weekend. The model, the data, the runbook all stay with the firm - no re-platforming fee, ever.

[ 03 - What the firm does with it on day one ]

Six things your associates do day one

Use case 01

Ask the firm’s files

Every matter, every pleading, every memo in the DMS - answerable in plain language, with the source file cited on every response.

Use case 02

First-draft the memo

Research memos, client letters, opinion drafts - in the firm’s own voice, using the firm’s own precedents. The associate edits; they don’t start from blank.

Use case 03

Review and redline

Contract review against the firm’s own playbook. Flags missing clauses, non-standard language, and deviations from precedent - senior partners verify, not hunt.

Use case 04

Dictation & meeting notes

Client calls, dictation, hearing prep - transcribed and summarized entirely on your hardware. No Zoom AI, no Otter, no third-party processor.

Use case 05

Discovery & scan review

Scanned productions, PDF evidence bundles, exhibit triage - extract, classify, redact privileged material. A week of junior work in an afternoon.

Use case 06

Conflicts & precedent search

Find the memo a partner wrote in 2017 on this exact question. Cross-check conflicts against ten years of matters. Institutional memory that doesn’t retire.

[ 04 - Who it’s for ]

Three firms on the call this quarter

A
Profile A

The managing partner

60–150 lawyers, multi-practice firm in Montréal or Québec. A major client just sent a vendor questionnaire asking about AI. Associates are already using ChatGPT off the corporate network.

→ Needs: an answer that holds under scrutiny
→ Wins: audit-ready in 9 weeks
B
Profile B

The COO / director of operations

Told by the managing partner to “go look at this.” Has been burned by vendors before. Needs specifics, a total-cost answer, and a deployment plan that doesn’t blow up billing for three weeks.

→ Needs: fixed price, clear scope
→ Wins: no surprise invoices, ever
C
Profile C

The IT director

15 years at the firm, runs IT with one junior. Needs to know exactly what this is, where it sits in the network, who owns the keys, and what happens when the GPU fails on a Friday afternoon.

→ Needs: an architecture diagram
→ Wins: something IT can actually support
[ Not a fit: ] Under 10 lawyers · “Can we just use Copilot” firms with no confidentiality stakes · No IT staff or MSP
[ 05 - What the firm receives ]

The deliverable, itemised

01
The server, racked and running
Physically installed in your server room, wired to your network, reachable from every desk through SSO.
Deployed
02
Model trained on your files
The firm’s documents, precedents, and playbook. Weights owned by the firm - cryptographically signed, versioned, yours to keep.
.safetensors
03
Source code
The web app, admin console, and training harness. MIT-licensed to the firm. If we disappear tomorrow, IT keeps operating it.
Owned
04
Audit log & admin console
Every prompt, every response, every source citation - logged. Exportable, reviewable, and ready for a vendor questionnaire.
SOC 2 ready
05
Runbook & architecture doc
30–50 pages, written for IT. How to deploy, patch, back up, and swap models. A diagram the managing partner can hand to a client.
PDF + MD
06
Associate training & 90-day handover
Two sessions with associates, one with partners, one with IT. Then private Slack / Teams, <4h response for 90 days. Afterwards the firm is self-sufficient by design.
Included
[ 03 - For the IT director ]

The architecture

Open-weight models, standard enterprise infrastructure, signed supply chain. Nothing your IT can’t audit.

Reasoning · embeddings · OCR · speech
  • Gemma 3 / Gemma 4
  • BGE-M3 / Nomic Embed
  • Tesseract / PaddleOCR
  • Whisper (self-hosted)
  • Firm LoRA
Sits on top of what you already use
  • Clio · Actionstep · ProLaw
  • iManage · NetDocuments
  • SharePoint · Microsoft 365
  • Outlook · Gmail archives
  • Active Directory · SSO
Retrieval · storage · runtime
  • Qdrant / FAISS
  • LangChain / LlamaIndex
  • FastAPI · encrypted Postgres
  • BIZON chassis · NVIDIA GPUs
  • Docker · Prometheus
Permissions · audit · sovereignty
  • SSO / SAML
  • RBAC enforced at retrieval
  • Full audit log of every query
  • Signed binaries · immutable snapshots
  • No outbound network · air-gap option
[ Risk · mitigation · on the same page ] What a partner will ask before signing

Every failure mode, already answered.

Drive failure RAID 1 or RAID 10 redundancy on every storage volume. Hot-swap drives on Firm and National tiers.
Power outage UPS with graceful shutdown. Optional generator integration on the National tier.
Disaster recovery Local encrypted backups with off-site rotation, controlled by the firm. Keys never leave the firm.
Ransomware Immutable snapshots and an air-gapped backup tier the encryption-malware can’t reach.
User abuse Full audit log of every query, file access, and AI response. Exportable for client questionnaires.
Permission abuse Role-based access control enforced inside the retrieval layer — not bolted on at the UI.
AI data leakage Architectural: no external network access from inference nodes. The model literally cannot phone home.
Hardware failure On-site spare components covered under the support SLA. RMA shepherded by us, not your IT.

All client data, embeddings, and model artifacts remain encrypted at rest. The firm owns the hardware, the data, and the encryption keys. We don’t hold a copy of any of the three.

[ Run the math · a firm of 80 lawyers ]

The return is arithmetic.

Hours saved · per lawyer / week
0 h
Drafting, summarising, review
Payback · at $400 avg rate
0
Weeks of recovered billable time
Loi 25 exposure · avoided
$0
Maximum fine per serious breach
3-yr TCO · Firm tier, 80 lawyers
$0k
80 lawyers × $29 / mo × 36 mo = $83,520 over 3 years · Copilot ~$115k, US/EU compute · hardware audited separately
Payback ladder · first-year subscription recovered in billable time
Firm size
Hours saved / yr
Recovered value
Payback
30 lawyers
6,300 h
$2,520,000
4 days
60 lawyers
12,600 h
$5,040,000
2 days
80 lawyers
16,800 h
$6,720,000
1.4 days
150 lawyers
31,500 h
$12,600,000
< 1 day
300 lawyers
63,000 h
$25,200,000
< 1 day
Client data sent to outside servers
0

Nothing the lawyers type, ask, or get back ever leaves the office. The next client questionnaire about “where is our data processed” becomes a one-page answer.

Outside companies seeing your data
0

Not OpenAI. Not Microsoft. Not Google. Not Anthropic. Not Bilbs. Loi 25 has nothing to say about data that never crosses a border.

Surprise price hikes from vendors
0

Your subscription is locked in at $29 / lawyer / month for the first 36 months — no surprise hikes, no per-token surcharges. Add services later only if you choose to.

* Hours saved based on 4.2 h/lawyer/week average across drafting, summarising, document review, and research. Value recovered at an illustrative $400/hr blended rate (your firm’s actual number goes here). Loi 25 fine figure is the statutory maximum (Art. 90.12). Full methodology in the .

Junior associates
1–3 h / day

Saved on first-pass drafting, precedent retrieval, and summaries from the DMS - the work the firm currently writes off in junior billables.

Senior lawyers
30–120 min / day

Saved on contract summarisation, internal precedent search across thirty years of matter memory, and knowledge retrieval that didn’t exist before.

Paralegals & admin
2–4 h / day

Saved on intake automation, draft generation, document review checklists, and the administrative support layer that runs the firm.

[ The elevator pitch ]

We help law firms deploy a private AI server that keeps confidential client data inside the firm while giving every lawyer instant access to the firm’s collective knowledge. Instead of unmanaged AI usage across ChatGPT and Copilot, we centralize AI securely — with governance, permissions, an audit log, and integrations designed for the way law firms actually work.

Bilbs AI · a private AI server for Québec law firms · Montréal-based

[ The pattern · already proven ]

We’re new. In-house AI isn’t.

Bilbs is a small company. In-house AI is not a small idea. The institutions with the strictest confidentiality rules on the planet — JPMorgan, Morgan Stanley, Bloomberg, Apple — already run their own AI on their own hardware, trained on their own files. They chose private AI because they couldn’t defend any other answer. The question for a Québec law firm is no longer whether this is possible. It’s who installs it, at a price the partnership can sign.

JPMorgan Chase 2024
230,000
employees · in-house LLM Suite

JPMC built LLM Suite entirely in-house because banking-secrecy and data-privacy obligations made external LLMs impossible. The same confidentiality logic applies to solicitor-client privilege - and to Loi 25.

Source · The Digital Banker
Morgan Stanley 2024
98%
advisor adoption · 16k users

AI @ Morgan Stanley Assistant lives behind the firewall, trained on a 100,000-document internal corpus. Advisors find what they need in seconds - the same pattern that makes a lawyer find the 2017 memo in seconds.

Source · Morgan Stanley press
Bloomberg 2023
363B
tokens · proprietary corpus

BloombergGPT - a 50B-parameter model trained on 40 years of the firm's proprietary financial data. The landmark paper proving a domain-specific, in-house LLM beats off-the-shelf on the tasks that matter to you.

Source · arXiv 2303.17564
Klarna 2024
700 FTE
workload absorbed by one AI

Klarna's in-house AI agent handles millions of conversations monthly - the work of 700 full-time agents, with customer-satisfaction scores on par with humans. Customer service moved from cost centre to line-item savings.

Source · Klarna / industry coverage
Meta 2023–2024
50k+
engineers · internal code LLM

CodeCompose / Metamate is Meta's internal AI coding assistant, trained on their monorepo and used by every engineer. Built in-house precisely because third-party code assistants couldn't see the codebase.

Source · Meta Engineering
Apple 2024
On-device
by default · Private Cloud Compute

Apple Intelligence runs most AI on the device itself; anything that can't is routed to Apple-owned Private Cloud Compute. Not a bolt-on to OpenAI - a deliberate architecture where the weights stay with the user.

Source · Apple Security

Every one of these institutions chose private AI — because their confidentiality obligations made no other answer defensible. Your firm answers to the same kind of obligations.

[ 06 - Foundation · partners · standards ]

Built on real things.

Bilbs is a young company. What the Box is built on, however, is not - it’s Google’s Gemma 4 open-weight family fine-tuned on the firm’s files, the server chassis that BIZON and Supermicro ship every week, and the Québec-based operating practices your IT director already runs the firm on.

Hardware
  • NVIDIA Blackwell RTX PRO 6000 · B200 · GB200
  • NVIDIA Hopper H200 NVL / SXM5
  • AMD EPYC · Intel Xeon host CPU platforms
  • Samsung · Solidigm enterprise NVMe + SSD
  • Mellanox / ConnectX 25 / 100 / 400 GbE NICs
Models & runtime
  • Gemma 4 Google · the only LLM family we ship
  • Gemma 4 4B / 12B / 27B / 70B size by tier
  • Gemma 4 Embedding retrieval over the document system
  • Gemma ToU commercial open licence
  • vLLM inference runtime
  • PyTorch training + fine-tune
Standards we follow
  • Loi 25 · Québec by design · no cross-border
  • Barreau du Québec confidentiality §3.06
  • SOC 2 Type I attested · Type II Q4 ’26
  • PIPEDA · federal DPA on file
  • FIPS-140-3 National tier
  • SLSA · cosign signed supply chain
Design partners

Our first design partners — Québec law firms and adjacent professional bodies with strict confidentiality obligations — helped us spec the product for the legal market. References available under mutual NDA once we’re past the 45-minute call.

[ The alternatives ]

What about Copilot?

Your lawyers already use AI. The real question is where it runs, what it knows, and what you can prove to a client.

What most firms do today ChatGPT, Copilot, or wait-and-see
With Bilbs A private server in your office
Where the AI runs A server in the United States.
Where the AI runs Your server room.
Loi 25 — data crosses a border Yes — every single query.
Loi 25 — data crosses a border No. Never.
Does it know your firm’s precedents No — only the public internet.
Does it know your firm’s precedents Yes — trained on every matter.
Record of every question asked None — it happens off-network.
Record of every question asked Every question, on your server.
What it costs the firm Hidden, per seat, growing.
What it costs the firm From $29 / lawyer / month.
Answer to a client AI questionnaire Nothing that holds up.
Answer to a client AI questionnaire One page, signed by the MP.

Copilot keeps your prompts “inside Microsoft’s data boundary” — but that boundary is still a US or EU data centre Microsoft operates. For a Québec firm with privileged-material obligations, that’s not the same conversation as a server in your own office. We’re not saying Copilot is wrong for every firm — we’re saying it isn’t private AI.

[ 06 - Competitive positioning ]

Compared to everything else on the table.

vs. ChatGPT / Copilot

No client data leaves the firm.

No exposure to third-party model-training pipelines. Citations are mandatory. ChatGPT and Copilot are good general-purpose tools — Bilbs is the confidential layer the firm puts around them so privileged work stops leaking into someone else’s data centre.

vs. Traditional legal ERPs

The AI actively understands the corpus.

A practice management system is a database. Bilbs is a system that reads the database — and the documents, and the emails, and the meeting transcripts — and answers in natural language with citations. It’s not a passive store; it’s a queryable memory.

vs. Cloud-based legal AI tools

Fully offline architecture.

Aligned with Loi 25 by design, not by promise. The cloud-legal-AI category answers data-residency questions with contracts and trust-centre pages. Bilbs answers them with a server sitting in your office.

vs. In-house IT building it themselves

Productized, supported, financed.

A one-off internal project ships, then drifts. Bilbs is continuously improved — model refreshes, security patches, runbook updates — under a support SLA. Your IT director gets to focus on the firm, not on babysitting a custom AI stack.

[ References ]

We’ll introduce you to firms like yours.

Our customer firms are Québec law firms with privileged-material obligations — they don’t put their names on a public website, and we don’t ask them to. After the free 45-minute call, we’ll set up a confidential reference call with a managing partner from a firm of your size. They’ll tell you what worked, what didn’t, and what they’d do differently. No marketing in the middle.

Mutual NDA within 24 hours · references after
[ 04 - How it rolls out ]

Nine weeks, associates on day one

Fixed-scope deployment. Kill-switch after week one if the shape is wrong. Billing is never disrupted.

01 Week 1–2

Free 45-minute call

A first call with the managing partner. A working session with your IT director. A written plan you can take to the partnership. Nothing is ordered until you sign it.

02 Weeks 3–5

It learns your firm

Your IT director gives us secure access to the document system. The AI reads every contract, memo, and matter your firm has handled — and learns how the firm writes. Nothing leaves the building.

03 Weeks 6–7

We install the server

The hardware arrives in a flight case. We rack it in your server room with your IT director. We wire it to the firm’s sign-on, the document system, and the practice management software. Tested before any lawyer sees it.

04 Weeks 8–9

Pilot group, then everyone

One practice group tries it first — usually corporate or litigation. We run two training sessions with the lawyers and one with admin. When the pilot group signs off, the rest of the firm rolls in. Billing and time entry stay exactly as they are.

[ 07 — Pricing ]

One line. From $29 per lawyer.

You’re not paying for GPUs or models. You’re paying for confidentiality, governance, and the institutional memory layer. $29 / lawyer / month covers audit, deployment, training, updates and support. Hardware is audited separately — yours, or we source it. Canadian dollars. No hidden line items.

Ship-or-refund guarantee

If the pilot group doesn’t hit the acceptance criteria we agreed on in the first week, the firm gets a 100% refund — no meetings, no lawyers, no awkward email. We carry the risk.

A typical invoice
Line 01
Bilbs — everything to run private AI for the firm
Audit, deployment, indexing, on-site training, updates, model refresh, retraining when you add new files, security patches, monitoring, backups, support within 4 hours, 24/7 emergency line, the audit log. Goes up only if the firm adds optional services (below). Cancellable on 30 days’ notice.
from $29 Below Copilot
/ lawyer / month
Hardware
The server itself — yours, or we source it
After the audit, the firm chooses: deploy onto a GPU-capable server you already own, or have us source and install one (separate, transparent quote at OEM cost, with the OEM’s 3-year warranty). Either way, the per-lawyer line stays the same. We don’t mark up the GPU.
Audit-first
your call after

How the math works: Everything Bilbs does for the firm — audit, deploy, train, update, support — sits in one per-lawyer line. The hardware is the firm’s asset, not a vendor-locked rental: it’s either already in your server room, or we’ll quote you the OEM list price for it. Our margin lives in the per-lawyer line, the same economics as Microsoft Copilot — on your own server.

Day one
Audit

No upfront fee. We audit the firm, propose the deployment plan, and tell you which hardware path fits. You sign only when you’ve seen the spec.

Every month · 60-lawyer firm
$1,740 / mo

60 lawyers × $29. One line on the invoice. Hardware audited separately — yours if you have a server, OEM-cost quote if we source it.

If you add services
À la carte

Fine-tuning on a niche practice area, high-volume transcription, multi-office sync, storage expansion — only added when the firm asks. Each one on its own line.

[ Optional · if and when you want them ]
Storage expansion

When the corpus outgrows the chassis.

Drop-in NVMe expansion or an additional storage shelf. Priced at hardware cost plus install. The firm’s indexed memory keeps growing without touching the tier.

Additional office

Per-office node, central sync.

For multi-office firms: a node per location with encrypted cross-office synchronization over private network. The Toronto office queries the Montréal precedents — without either dataset leaving Canada.

AI fine-tuning

Premium tier · custom model behaviour.

For firms that want the model to write in a very specific house style, or that operate in a niche practice area. We fine-tune on the firm’s corpus, weight critical document classes, and re-evaluate every quarter.

Transcription credits

High-volume meeting capture.

The base tier handles regular meeting load. For litigation firms running 80+ depositions a month, additional Whisper capacity is billed as usage-based credits — still entirely on-premise, no cloud transcription ever.

None of these are required. The base subscription — from $29 / lawyer / month — covers what most firms need for years. Expansion lines exist so the firm doesn’t outgrow the platform on the corpus side or the workload side. Each one shows on its own invoice line, never folded into the base.

[ 08 - What the partnership will ask ]

The real questions.

“This sounds expensive.” +

The firm is not paying for GPUs or models — it is paying for confidentiality, governance, operational leverage, and risk reduction. The price the firm sees on the invoice is one line: from $29 / lawyer / month. That covers the audit, the deployment, on-site training, updates, support, the audit log, the 24/7 pager — everything that makes the AI run for the firm. It goes up only if the firm chooses additional services. Hardware is audited separately: if you already have a GPU-capable server, we use it; if you don’t, we quote you the OEM cost on a separate page, no markup. For a 60-lawyer firm: 60 × $29 = $1,740 / month. Below Microsoft Copilot. One day of avoided billable-time leakage pays for a quarter of subscription. A single Loi 25 incident carries a maximum fine of $10M or 2% of worldwide revenue — that’s the real frame on the cost question.

“We already have an AI policy.” +

A policy describes what people should do. A system decides what people can do. Survey after survey shows most lawyers use AI in spite of the firm’s policy — because the policy says no, but the tool that’s in front of them at 9 pm says yes. The Box flips that around: it gives lawyers an AI they actually prefer (because it knows the firm’s own precedents and writes in the firm’s voice). The policy becomes something the system can enforce, not just a paragraph in an email.

“We already have Microsoft Copilot. Why add Bilbs?” +

Bilbs is not competing with Copilot. Most of our customer firms keep Copilot for the M365-native productivity it’s good at (Word, Excel, Outlook drafting on non-privileged content). What Copilot can’t be is the firm’s confidential AI layer: it processes prompts in Microsoft-operated data centres (cross-border under Loi 25), it can’t be fine-tuned on the firm’s precedents, it has no on-prem audit log, and it has no governance over the ChatGPT use that’s already happening in browser tabs. Bilbs sits around your existing AI - including Copilot if you want to keep it - and gives the firm the secure, governed, knowledge-connected layer that none of the off-the-shelf chatbots can be. Side-by-side analysis in the compliance audit.

“What if it gives wrong answers?” +

Every answer cites the exact file it came from. If the AI cannot verify a source in the firm’s own indexed material, it doesn’t produce an answer — the retrieval pipeline architecturally prevents the hallucinated-paragraph problem that gets lawyers in trouble with ChatGPT. The lawyer clicks through and reads the source before anything goes out — the same review a senior partner does on a junior’s draft today, just on a draft that’s 90 seconds old instead of 4 hours old. The Box is a first drafter, never a decider. Human oversight remains mandatory on every legal work product — that statement sits in the application, the onboarding kit, and the terms of service.

“What happens if Bilbs goes out of business?” +

The server is in your building. The trained AI is on it. The software is licensed to your firm and sits in your IT’s repository. A 30-50 page printed runbook lives on the shared drive. If we close shop tomorrow, everything your firm uses today keeps running. That’s the whole point of installing it in your office: the firm depends on the firm, not on a vendor.

“Our associates won’t adopt it.” +

Your associates are already using AI — they’re just doing it in browser tabs the firm can’t see. The real question is: will they prefer the firm’s AI to the public one? Yes, because the firm’s AI knows your precedents and your matters. It can answer “how did we handle the Smith file in 2022” — ChatGPT can’t. In the law firms we’ve deployed in, adoption runs ahead of public AI within a month. Training is included: two sessions per practice group, on-site.

“Too disruptive to deploy.” +

Nine weeks from the signed plan to the day the AI goes live for the firm. Day one looks identical to the day before — lawyers open a tab in the browser, just like they open Outlook. Billing, time entry, and the document system don’t change. One practice group tries it first; the rest of the firm rolls in when that group signs off. Your IT director’s weekend isn’t on the hook — we install, we stay on call, we own the runbook.

“We’re going to wait and see what happens with AI.” +

Waiting is itself a decision — and it costs something. Every month, your associates keep typing client matters into ChatGPT. Every month, another firm on Bay Street or Côte-des-Neiges deploys this and starts answering client AI questionnaires with a better answer than yours. The free 45-minute call is the lowest-commitment way to find out where your firm actually stands — you walk away with a written assessment whether you proceed or not.

“Paperwork, IP, warranty?” +

Mutual NDA signed within 24 hours of your first call. The MSA and the privacy agreement are bilingual (French + English) and already Loi 25-aligned. The firm owns the trained AI outright at the end of the lease — full IP transfer on final payment. The hardware carries the manufacturer’s 3-year warranty; we handle the paperwork if anything breaks. Based in Montréal, Eastern time. Response under 4 hours, emergencies same-day.

[ The problem ]

Your associates are already using AI. The firm is not protected.

01 · The shadow AI

Client files are already going into ChatGPT.

A policy on the wall doesn’t stop a junior associate pasting a contract into ChatGPT at 11 pm. Only an in-house alternative the lawyers actually prefer does. Right now, most partners find out their associates use AI after the fact — if at all.

→ Give them an in-house tool they prefer over ChatGPT.
02 · The Loi 25 risk

One incident. One complaint. One La Presse headline.

Every prompt that leaves the firm — to OpenAI, to Microsoft, to anyone — is privileged material crossing a border. Under Loi 25, fines reach $10 million or 2 % of worldwide revenue, and the partners are on the hook personally.

→ Nothing leaves the building. Ever.
03 · The client’s question

Your biggest client just asked how you use AI.

Banks and in-house counsel now send questionnaires asking where the AI runs, who can see the data, and how it’s logged. “We have a policy” is no longer an answer. A real answer fits on one page — but only if the AI actually lives in your office.

→ A one-page answer signed by the managing partner.

The question isn’t whether your associates will use AI. They already do. The question is whether the firm is protected — and whether the managing partner has an answer ready the next time a client, a regulator, or a journalist asks.

[ 09 - Long-term vision ]

A private operating system for institutional intelligence.

Québec law firms are the entry market because they share the hardest version of one structural problem — highly sensitive data, decades of institutional knowledge trapped in unstructured files, growing regulatory pressure, and no acceptable cloud AI path. The same architecture, security posture, and commercial model translate directly to other regulated, high-trust industries. The roadmap is on the wall.

Entry market
Law firms

Québec, then francophone Canada and broader Canada.

Vertical 2
Medical · dental

Clinics under professional secrecy and Loi 25 health-data rules.

Vertical 3
Accounting · audit

Working papers, client files, retention obligations.

Vertical 4
Wealth · family offices

Private wealth managers and multi-family offices.

Vertical 5
Insurance brokers

Policy archives, claims histories, underwriting memory.

Vertical 6
Public sector

Municipal and provincial departments with sovereign-data mandates.

We’re not building an ERP. We’re building the private operating system for institutional intelligence in regulated industries. Law firms come first because the constraint is hardest there: Loi 25, professional secrecy, and decades of partner judgment locked inside PDFs and inboxes. Every vertical above has the same shape — sensitive data, no acceptable cloud AI path, and accumulated knowledge that walks out the door when a senior leaves.

[ 09b - Pilot program ]

A limited cohort of Québec firms.

Each cohort is intentionally small so deployments overlap only on our time, not the firm’s. The structure is the same every quarter — five steps from the first call to the day the AI is live for the firm.

01

Discovery call

20 minutes, free. The managing partner and the IT director on one call. We listen, we ask, we tell you whether the fit is real.

02

Needs analysis

A short infrastructure review of the firm’s document system, practice management, identity directory, and network. Written assessment in your inbox.

03

On-prem deployment

Server racked in your office. Controlled indexing of existing documents — no migration, the files stay where they are. Permissions wired to your directory.

04

Onboarding

Staff onboarding and permissions configuration. Two sessions per practice group, on-site. The firm’s policy becomes something the system can enforce.

05

Feedback loop

Continuous feedback during the pilot window. We optimize on what the firm’s lawyers actually do with it. The platform learns the firm; we learn the firm.

Reasons to be in this cohort

The firms that join the early cohorts get the lowest deployment friction, the most engineering attention, and the right to shape the roadmap. The market is starting to divide — between firms that operationalize AI internally and firms that don’t. We’re taking a handful of partner firms per quarter, and after that the calendar rolls.

Q3 2026 cohort · free-call calendar open

Another firm in
Montréal is already on the calendar.

We take a small number of Québec law firms per cohort so deployments overlap only on Bilbs’s time, not yours. The 45-minute call is free and commits you to nothing. You leave with a written assessment of where your firm stands on shadow AI, Loi 25, and the gap between the AI usage already happening and the AI usage the partnership can actually defend.

Download whitepaper
No payment today · 45-min call within one business day · No pitch, no deck
Parlons-en

BOOK THE AUDIT. MEET THE PARTNERSHIP. PROTECT THE FIRM. KEEP YOUR FILES.