Private, local-model AI.
For the buyers who can't use a public API.
A private, local-model configuration for buyers who cannot send prompts to public APIs. Custom engagement, hardware supplied. Nothing leaves your machine.
Book a 30-min discovery callSome buyers can't send their prompts to a public cloud API.
The standard AI assistant sends your prompts to a public cloud API. For most businesses that is fine. For some, it is not. Legal firms, health practitioners, defense-adjacent work, or anyone with genuine confidentiality constraints can't use a setup that hands their data to a third-party model provider. The Obsidian Executive Assistant is built for that buyer.
Distribution is the thing AI can't replicate for you.
Dheer Gupta makes the case that in an era of cheap models, the only durable moat is distribution. Your existing customers, your channels, your phone list, your reputation in your market. Those don't come from a vendor. The Obsidian Executive Assistant is for privacy-constrained operators who already have distribution and want AI running behind it without sending a single prompt off the machine.
"The only moat AI can't kill is the one you've already built: your distribution."
Paraphrased from Dheer Gupta, The Only Moat AI Can't Kill. The essay reads as a decent summary of why established operators have more to gain from AI than first-time builders do.The Obsidian Executive Assistant fits a privacy-constrained operator who already has customers, channels, or a name in the market. The investment makes sense because AI behind distribution compounds. AI without distribution doesn't.
What replaces what, at operation scale.
Daniel Miessler calls this "the fire of fires": AI is letting operators replace large parts of the SaaS stack rather than add another tool to it. At the Obsidian Executive Assistant tier, that same principle runs across your whole operation, role by role. Below is the typical pattern we see in discovery. Your mix varies. The point is that this is a teardown, not an install.
| Workflow | Current tool pattern | Executive Assistant agent replacement | Iteration cadence |
|---|---|---|---|
| Lead follow-up and nurture | CRM plus a sequencing tool plus a VA | Sales agent reads inbox, drafts, sends, logs to CRM | Quarterly |
| Meeting prep and notes | Meeting-notes AI plus a separate briefing doc | Briefing agent pre-reads, prep doc, post-meeting notes | Quarterly |
| Proposal and quote drafting | Template tool plus a writer plus rounds of edits | Proposal agent trained on your voice and pricing | Quarterly |
| Finance and bookkeeping triage | Bookkeeper plus category-guessing tool | CFO-style agent triages, flags, routes to your human bookkeeper | Quarterly |
| Compliance and document review | Partner review plus manual checklist | Compliance agent runs the checklist, flags exceptions | Quarterly |
| Customer onboarding | Forms tool plus project tool plus a coordinator | Onboarding agent runs the sequence, nudges clients, files docs | Quarterly |
| Internal reporting | Dashboard tool plus someone who builds the dashboard | Reporting agent pulls from source systems, writes the brief | Quarterly |
| Content and thought leadership | Agency or freelancer plus scheduling tool | Content agent trained on your voice, you approve, it schedules | Quarterly |
| Recruiting and pre-screening | ATS plus a recruiter plus screening calls | Recruiting agent filters, pre-interviews, shortlists | Quarterly |
| Knowledge base and SOPs | Wiki plus someone who keeps it updated | Knowledge agent maintains it from actual work output | Quarterly |
Not every row applies to every operation. Discovery is where we figure out which 6 to 10 actually fit your work and get built first.
How this differs from Tier 1 and Tier 2.
The next two paragraphs are technical. Skip to the next section if you're not a tech reader. Nothing below changes what you actually get.
The standard Digital Worker and Digital Assistant call the Anthropic Claude API for reasoning. Fast, capable, and your prompts travel to Anthropic's infrastructure under their API terms. Obsidian Executive Assistant runs everything on the machine we ship you. We rebuild our standard system to use a local model (Ollama or llama.cpp with a 70B to 405B class model depending on the hardware we spec for you) instead of Claude. Nothing leaves your machine.
The trade-off: local models today are slower and less capable than Claude on complex reasoning. We tell you exactly what your use case can and cannot do before you sign. No surprises. If your use case is a poor fit for local inference at today's model quality, we say so and recommend Tier 2 with a tighter data-handling contract instead.
How the year tends to unfold.
Discovery and first agents
Two-day discovery. Architecture doc. Private stack provisioned. First 3 to 4 agents live by end of Q1. Team trained on Telegram and Slack interfaces.
Real-work calibration
The first agents meet real edge cases. We tune voice, permissions, escalation rules. Next 2 to 3 agents added based on where the team is hitting friction.
Cover the remaining roles
By Q3 the pattern is clear and the team is asking for agents themselves. We add 2 to 3 more. Integrations deepen with whatever your team actually uses.
Handoff and forward plan
Full docs. Audit-friendly logs. You decide: take it in-house with your ops lead, or roll onto a monthly partnership for year two. Either way the system is yours.
Four quarterly roadmap reviews happen during the year. These aren't status calls. They're working sessions where scope for the next 90 days gets set together.
The full stack, yours, forever.
Custom architecture
We design the stack around YOUR workflows. Not a template. We sit with you and your team for 2 full days of discovery before we write a line of code.
Unlimited integrations in scope
Any tool with an API is fair game. Your CRM, ERP, industry vertical software, billing platform, compliance tracker, whatever you rely on.
Role-specific agents
Different agents for the CEO, the ops lead, the sales team, the finance person. Each with their own access, their own role, their own voice.
Multi-channel communication
Telegram, Slack, email, SMS. The agents reach you and your team wherever you already work. No new habits required.
Quarterly roadmap reviews
Every 90 days we sit down, review what's working, what's not, what's changed in your business, and we ship new capabilities accordingly.
12 months of partnership
A full year with us on the project. Updates, new agents, tuning, training, as your operation evolves. Not just a build and leave.
Compliance and data handling
Canadian privacy-law compliant (PIPEDA), with full access records, documented access policies. Ready for any client asking "how do you protect my data?"
Team training and documentation
Your team learns the system. Written docs, recorded walkthroughs, live sessions. Not a black box nobody understands but us.
Note on the AI models. At the Tier 3 tier the reasoning happens on the local model running on hardware we ship you. Nothing leaves the machine. No public API is called for reasoning. If you want a cloud model for some specific workload in addition, we can wire that in as an optional, scoped-out exception with your written sign-off. Full data-handling terms in our privacy policy.
What's included in the engagement
- 2-day on-site or over-video discovery with you and key team members
- Custom architecture document delivered within 2 weeks of contract start
- Local-inference hardware supplied (Mac-class machine, roughly $15K CAD at flagship spec)
- Our standard system rebuilt to run on a local AI model (Ollama or llama.cpp, 70B to 405B class)
- All requested integrations built during the 12-month engagement
- Role-specific custom agents for up to 8 named team members
- Multi-channel setup: Telegram, Slack, email, SMS as needed
- Quarterly roadmap reviews (4 over the year)
- Unlimited tuning and new-agent additions during the year
- Team training sessions and written operator docs
- Data-handling agreement, mutual NDA, audit-friendly access logs
- Priority 24-hour response SLA for issues during the year
Custom-priced after discovery. Hardware (Mac-class local-inference machine, approximately $15K CAD at flagship spec) is supplied and included. Scope priced per engagement. Payment terms: 50% at contract signing, 25% at 90-day milestone, 25% at 12-month close. Interac e-transfer, wire, or invoiced with NET 15.
Privacy-first buyers who can't hand prompts to a third-party model.
Legal and compliance-constrained firms
Solicitor-client privilege, litigation files, M&A diligence, regulated client data. You can't defensibly send any of that to a public cloud API. Local inference keeps every prompt on the machine.
Health-adjacent practitioners
Clinical notes, patient-adjacent workflows, PHI that touches your admin tools. If you'd rather not explain to a regulator why clinical context went through a third-party API, this is the tier.
Data-sensitive operators
Defense-adjacent work, financial advisors under regulator scrutiny, family offices, IP-heavy operators. Buyers whose material edge is what's in their head and their files, and who need that to stay there.
What buyers ask before signing.
What does the 2-day discovery look like?
Day one: we sit with you and the key operators, map the current workflows, find the highest-friction points. Day two: we sketch the architecture live with you and confirm what's in scope. You leave with a document you can share with your team.
What if our business changes significantly during the 12 months?
Expected and handled. The quarterly roadmap reviews exist for exactly that. New business unit? We add agents for it. Acquired a company? We absorb their workflows. No change-order fees within the original scope.
What's the difference between this and Digital Assistant (Tier 2)?
Digital Assistant is the standard template: up to 5 integrations, 3 custom agents, 90-day support. Runs on the Claude API. Great for buyers with no hard privacy constraint. The Obsidian Executive Assistant is for when you cannot send prompts to a public cloud API and you need the AI running on hardware you physically own. Different buyer, different build.
Do you take equity instead of cash?
Sometimes. If your operation is a strong strategic fit and you'd rather preserve cash, we'll talk equity + reduced cash combinations. Terms are case-by-case.
What happens at the end of the 12 months?
You own the infrastructure fully. We offer an ongoing partnership retainer at a monthly rate (typically $3,000-$5,000 per month depending on usage) for continuing updates, new features, and priority support. You can also walk away clean with the full docs and the running system.
How many of these engagements do you run at once?
Never more than 4 active engagements. This is deep work. Quality stays high because we don't overload ourselves.
Let's see if this is the right fit.
30-minute call. No hard pitch. We learn your business. You decide if the investment makes sense.
Book a 30-min discovery call