# AI, ARIA assistant, document scan & draft, and AI setup

> Screenshots use representative demo data captured from a live environment. AI features depend on a configured provider. Never send real customer data to a model without that tenant's explicit data-consent, and never show API keys.

How AI works in ColorsX74 ERP: the ARIA assistant and its skills, document scanning, document drafting, AI configuration, and the guided five-step setup wizard.

[← Public docs landing](README.md) · [Open HTML version](ai-automation.html)

## How AI is positioned

AI is ERP-aware help embedded in the workflow, branded **ARIA**. It answers questions about your data, scans documents into structured records, and drafts business documents — governed by a per-tenant provider configuration and consent posture. It is built around discrete *skills* you toggle, and a setup wizard guides administrators from hardware detection to activation.

## ARIA — the AI assistant

- **What.** A conversational assistant over your ERP data, with a *Skills* panel (each ON/OFF) and a *Provider* panel (model, provider, status). It offers starter prompts.
- **How.** Open *ARIA*, type a question or click a prompt (e.g. "Show this month's top 5 customers by revenue", "Which contracts expire in the next 30 days?"). In the demo, PDF reports, Email draft, Contract draft, Document scanner, and Summaries are ON; Forecast is OFF. Enter to send, choose a model.
- **Wisely.** Treat ARIA as a fast analyst — verify numbers against the source report. Enable only needed skills. Ask concretely. Keep demos on demo data.

![ARIA assistant](assets/screenshots/web/ai.png)

## Document scan

- **What.** Doc Scanner turns documents (invoices, receipts) into structured data for review.
- **How.** Upload a clean demo file, review the extracted fields, correct, then commit. Extraction is always a draft for a human.
- **Wisely.** Always review amounts, dates, and tax before committing. Use clean, non-confidential samples in demos.

![Document Scan](assets/screenshots/web/ai-scan.png)

## Document draft

- **What.** Turns a prompt + template into a first draft of a business document.
- **How.** Choose a template, describe what you need, edit and finalise the draft.
- **Wisely.** Use drafting to skip the blank page, not the review — read every generated clause. Keep a small set of good templates.

![Document Draft](assets/screenshots/web/ai-draft.png)

## AI Settings — provider, models, skills, hardware, privacy

- **What.** Configure the tenant's AI across tabs: *Provider*, *Models*, *Skills*, *Hardware*, *Sessions*, *Privacy*; links to the setup wizard.
- **How.** On *Provider*, pick a provider (e.g. Anthropic Claude), choose/enter a model, set endpoint and optional API key, then *Test connection* before saving. Use other tabs for skills, hardware tier, session history, and consent.
- **Wisely.** Always *Test connection* first. Keep the API key in secure config, never on a shared screen. Match the model to the hardware tier the Hardware tab reports.

![AI Settings](assets/screenshots/web/ai-settings.png)

![Settings → AI Configuration](assets/screenshots/web/settings-ai.png)

## The AI Setup Wizard

A guided five-step path: **1 Detect hardware → 2 Choose provider → 3 Choose model → 4 Skills & confirmation → 5 Test & finish**. Step one inspects CPU, cores, RAM, and GPU and recommends a provider tier (e.g. llama.cpp with a small model on a CPU-only machine, with a RAM advisory).

![AI Setup Wizard](assets/screenshots/web/ai-setup.png)

**Wisely.** Run the wizard once per environment before opening AI to users, and heed RAM/GPU advisories — picking a model the hardware cannot serve is the top cause of slow/failing AI. Re-run after a hardware change.

## Consent, safety, and ARIA handoff

User data is only sent to a model where the tenant has granted consent; scanning/drafting produce drafts for human review; sessions are recorded for auditability. When discussing ARIA, emphasise human confirmation, permissions, and job history over invisible automation.

**Wisely.** Make consent an explicit per-tenant decision, keep a human in the loop, and demo only with demo data and sandbox credentials.

**Enabling cloud AI ("enable cloud content" prompt).** Local providers need no consent; cloud providers — OpenAI, Anthropic, Google Gemini, **OpenRouter** — only become selectable/usable once an admin grants this company's *AI data consent* (AI Settings → Privacy, or the wizard's provider step). That toggle clears the "enable cloud content / consent required" message and unlocks the cloud cards, and still gates every actual request per tenant.

**OpenRouter.** Unified OpenAI-compatible gateway (one key → OpenAI/Anthropic/Google/Mistral/…). Pick *OpenRouter* (endpoint prefilled to `https://openrouter.ai/api/v1`), paste your OpenRouter API key, choose a model id like `openai/gpt-4o-mini`, then *Test & finish*.

## Related pages

- [Overview, login, and dashboard](overview.md)
- [Accounting, reports, banking, tax, and payments](finance.md)
- [Sales, purchase, contracts, and rental](commercial.md)
- [Inventory, manufacturing, marketplace, and add-ons](operations.md)
- [Admin, settings, security, module access, audit, and approvals](admin-security.md)
- [Troubleshooting, browser requirements, and support handoff](troubleshooting.md)