# Osolix · Hybrid AI Mode
## The only Fixed Asset Management platform that runs Claude AND your local model in parallel — and picks the higher-confidence answer per request

> **Audience:** CIO · CISO · CFO · Head of Internal Audit
> **Read time:** 3 minutes
> **Demo time:** 10 minutes (live extraction of a real scanned invoice)
> **Audit-evidence pack available:** SOX 404 control matrix mapped to code paths

---

## The problem every FAM buyer faces

Every Fixed Asset Management evaluation in 2026 hits the same fork:

| Path | What you get | What you give up |
|---|---|---|
| **Cloud-AI-only** (Mojodat · Asset Panda · MaintainX) | High accuracy on invoice extraction, asset classification, predictive maintenance | Per-call API cost (cents per asset × 200 K assets = real money), tenant data leaves the boundary, vendor lock-in to whoever powers their AI |
| **No-AI / rule-based** (Snipe-IT · legacy on-prem) | Zero cost, zero data egress, audit-defensible | OCR fails on scanned invoices, no anomaly detection, no offline-capable insight, no learning from history |

You're forced to pick one. **Until now.**

---

## What Osolix does differently

Osolix Hybrid AI Mode runs **both tiers in parallel on every single request** and returns the answer with the higher confidence — *with the source badge stamped on every field*.

```
                 ┌─── Online (Anthropic Claude) ─── confidence 0.91 ──┐
                 │                                                     │
   user upload   │                                                     ├─►  pick
   invoice.pdf ──┤                                                     │   higher
                 │                                                     ├─►  confidence
                 │                                                     │
                 └─── Offline (llava-phi3 in-tenant) ── confidence 0.84 ┘
                                                          │
                                                          └─► reasoning string:
                                                              "Hybrid: Online won
                                                               (0.91 vs 0.84)"
```

Three fundamental differences from every competitor:

1. **You never lose data sovereignty.** The offline tier runs Ollama / vLLM / llama.cpp on the customer's own server. Tenant data stays inside the tenant boundary in both tiers — Anthropic's "no training on customer data" is the floor, not the ceiling.
2. **You never lose the cost-of-API floor.** When the offline model wins, the request cost is zero. Claude only gets called on the requests that actually need its accuracy edge.
3. **You never lose to a model deprecation.** When Anthropic deprecates a model or a customer flips to Sovereign mode, the offline tier keeps every workflow alive without code changes.

---

## How it works in practice

### Configuration is per-tenant + per-AI-mode (Online · Offline · Hybrid · Sovereign)

```
Online      →  Claude calls every request. Falls back to Offline on rate-limit.
Offline     →  Local model only. Air-gapped tenants. Zero outbound calls.
Hybrid      →  Both tiers in parallel; higher-confidence wins.
Sovereign   →  Hybrid with PreferOfflineOnTie=true. Equal-confidence ties go
                to local. For regulated tenants who want Claude only when it
                strictly outperforms.
```

### Verified live during this product release

```json
{
  "vendorName": "ACME Computer LLC",
  "invoiceNumber": "INV-2026-0042",
  "lineItems": [...],
  "source": "Offline",
  "confidence": 0.84,
  "reasoning": "Hybrid: Offline tier won (confidence 0.84 vs Online 0.20).
                Local vision model 'llava-phi3' read 1 page image of
                invoice.jpg directly inside the tenant boundary —
                no outbound calls."
}
```

### Where Hybrid Mode plugs in

Every Osolix AI feature can run in any mode — the resolver pattern is uniform across the product:

| Feature | Online tier | Offline tier |
|---|---|---|
| **Smart Invoice Wizard** | Claude vision | llava-phi3 / moondream / qwen2-vl |
| **Asset Classifier** (category + sub-category suggestion) | Claude Haiku semantic ranking | In-tenant rules + history k-NN |
| **CAPEX Co-pilot** (justification + similar projects) | Claude Sonnet | History k-NN + standards retrieval |
| **Maintenance Co-pilot** (root cause + similar incidents) | Claude Sonnet | Heuristic + maintenance history |
| **Retirement Co-pilot** (disposal optimisation) | Claude Sonnet | Residual value + tax rule engine |
| **AI Chat Assistant** | Claude Sonnet | Local LLM endpoint (configurable) |
| **16 internal AI agents** (data cleansing · IFRS compliance · depreciation auditor · etc.) | Claude where useful | Heuristic-only for the deterministic ones |

---

## What competitors can't match

The Hybrid architecture rests on three pieces of Osolix engineering that no SaaS competitor can copy in <12 months:

1. **Per-tenant offline-model hosting integration.** Osolix talks OpenAI-compatible chat completions to whatever runtime the customer installs (Ollama, vLLM, LM Studio, llama.cpp). The customer brings the GPU; Osolix brings the prompts.
2. **JSON-salvaging parser tuned for small open-weights models.** Small CPU-friendly models (moondream 1.7B, llava-phi3 3.8B) routinely emit prose-wrapped JSON. Osolix's salvager extracts the structured object regardless of wrapping — critical for reliability when running offline.
3. **Compact-prompt fallback for sub-7B models.** When the configured offline model is small, the resolver swaps to a tighter prompt that small models actually follow. Without this, offline accuracy collapses on tenants with modest hardware.

> **Mojodat / Asset Panda / Maximo / Oracle FA all use cloud-only AI.** They are architected around per-call OpenAI / Anthropic / OpenAI-Azure spend and shared-tenant infrastructure. To match Osolix's Hybrid Mode they'd need to: (a) add per-tenant offline-model hosting, (b) author a parallel-invocation resolver, (c) write the JSON salvager + compact prompts, (d) instrument the cost telemetry, (e) audit-log both tiers separately. That's a 12-18 month architectural pivot.

---

## What enterprise security questionnaires now answer

The Hybrid + Sovereign architecture closes the standard enterprise-AI security questionnaire in one breath:

| Question | Osolix answer |
|---|---|
| Does customer data leave the tenant boundary for AI? | **Configurable per tenant.** Sovereign + Offline modes: never. Hybrid + Online modes: only when the customer explicitly enables Anthropic, with no-training-on-customer-data contractual floor. |
| Can AI be disabled for regulated workflows? | Yes — per AI agent toggle (`AgentSetting` table) AND per-mode global flag. |
| What happens if the AI provider goes down? | Tenants in Hybrid / Online mode auto-fall back to Offline; tenants in Sovereign / Offline mode are never affected. Zero-incident architecture. |
| Where does the AI confidence score come from? | Per-tier, per-field, exposed in API response and stamped on every UI badge. Audit-defensible. |
| How do you log AI usage for SOX traceability? | Every AI call writes `AuditLog · EntityType='Auth' OR 'Scim' OR 'AI'` with provider, mode, and outcome. Per `docs/31-sox-control-matrix.md` ITGC-A2. |

---

## Cost transparency

Hybrid Mode runs both tiers per request — the AI Settings page surfaces a clear cost-impact banner so finance teams understand the spend profile before the contract signs:

> "Hybrid mode invokes both Online and Offline tiers per request, so Anthropic API spend is roughly equivalent to Online-only mode for the same volume. Use **Offline** for zero-cost / sovereign deployments, **Online** when Anthropic accuracy is the priority, and **Hybrid** when you want best-of-both and accept the cost trade-off. Wins are recorded in the AI source badge so you can later see which tier actually answered."

For the Osolix Sovereign SKU (air-gapped tenants, no internet), the Hybrid Mode tie-break flag (`Tenant.PreferOfflineOnTie`) is on by default, and Online tier calls fail-fast to the Offline path.

---

## Available SKUs

| SKU | AI tier | Per-asset / month | Best fit |
|---|---|---|---|
| **Pro**          | Online (Claude) | included | SaaS-first tenants on AED budget |
| **Pro Secure**   | Online + Hybrid | included + per-call passthrough | Mid-enterprise wanting cost optionality |
| **Suite**        | Hybrid (default) | included + per-call passthrough | Enterprise with 50K+ assets + Big-4 audit posture |
| **Sovereign**    | Offline only · air-gapped | included (no per-call) | Government · defence · regulated banking |

---

## Demo flow (10 minutes)

1. **30 sec:** Open the Smart Invoice Wizard. Drop in a scanned PDF the prospect supplies on the call.
2. **2 min:** Watch the Hybrid resolver fire both tiers. Show the AI source badge stamp on every extracted field — vendor, invoice number, line items.
3. **3 min:** Walk through `docs/31-sox-control-matrix.md` § "Audit log integration" — show the `audit.AuditLogs` row written per AI call.
4. **2 min:** Flip the tenant to Sovereign mode (`AiMode=Hybrid` + `PreferOfflineOnTie=true`). Re-upload the same invoice. Show the same fields extracted, this time with `source:"Offline"` and zero Anthropic spend.
5. **2 min:** Open the AI Settings page. Show the per-agent toggle (16 agents). Point out the cost-transparency banner.
6. **30 sec:** Close on the moat: "No competitor offers this. The architecture takes 12-18 months to copy. The first SaaS contract you sign for FAM either commits you to per-call AI cost forever, or to no-AI forever — except Osolix."

---

## Pricing leverage

Hybrid Mode is the **single highest-priced feature in the SKU ladder** because it's the only way to combine maximum accuracy with maximum cost flexibility. Strategy Head guidance: list price for Suite SKU should be 30-40% above Pro Secure on per-asset basis, and Sovereign should be 80-100% above Pro Secure with a flat-fee floor for air-gapped deployments where per-asset pricing doesn't model cleanly.

---

## Closing line

> "Osolix is the only Fixed Asset Management platform built so the AI architecture is your competitive advantage, not your vendor's. You get Claude when Claude is right. You get your local model when local is right. You always get the audit trail. You never pay twice."

— Filed by Head of Corporate Strategy · 2026-05-02