Where governance is not a nice-to-have.

Situation

An agent is to review goodwill requests and create refunds in the ERP. The requests contain bank details and medical certificates; every booking moves real money.

With ATF

• The PII vault masks IBAN, name and diagnosis before every model call
• The MCP Gate offers only erp_customer_lookup and erp_create_refund – the booking pauses until approved
• Policy: health data only via the EU provider; no decision means rejection
• Every case ends in an Ed25519-signed evidence bundle for internal audit

Situation

Claims arrive as PDF, photo and email. Handlers should decide faster, but payouts and rejections must stay justified and auditable.

With ATF

• Extract turns forms, expert reports and emails into structured Markdown with page references
• Classifier and router sort by line of business and urgency – without embedded code
• Payouts via the Gate only with approval, optionally four-eyes
• Shadow mode measures new prompt versions on real traffic before they go live

Situation

Manuals, maintenance plans and incident reports live on a file server in the plant. The cloud may not access them; answers must point to the right page.

With ATF

• The Outpost mirrors the plant folder into the platform – it pushes, the cloud never pulls
• Extract keeps tables and page numbers; figures stay referenceable
• Search answers hybrid with citations, only from documents the person asking may see
• Local models via Ollama or vLLM, data never leaves the plant

Situation

Matters and files are strictly separated. AI research must never mix content from another matter into an answer – and every source has to be traceable.

With ATF

• Tenant and permission separation via row-level security in PostgreSQL, not in the UI
• The verified identity is authoritative; a client cannot claim permissions
• Document diffs between points in time and tracing a file number through time
• Explain card per answer: which sources, which model, which policy

Situation

An agent is to triage tickets, narrow down causes and prepare changes. It may read a lot but change things only with approval – and token costs must stay predictable.

With ATF

• Flow with classifier, router and parallel analysis – testable in the Studio, reviewable as YAML
• The Gate separates ask and act: monitoring queries free, rollbacks only with approval
• Per-run budgets and a model cascade: cheap model first, escalate on weak evidence
• Live graph and debug trace show which branch ran in which version

Situation

An assistant is to summarize findings in plain language and answer questions. Health data (GDPR Art. 9) may only be processed under strict conditions.

With ATF

• Data class health in the flow; the policy only allows approved providers and regions
• PII vault fail-closed: if detection fails, nothing reaches the model
• Output guardrails and groundedness checks against the source
• Erasure across all services with a signed receipt; a dossier for the DPO