# VARIA

> VARIA is a decision-flow simulator. Compose decision-model questions, LLM steps, code rules, confidence gates and human handoffs, run 10,000 items, and see cost per 1M, p50/p95 wait and how much work still lands on people — before you build.

VARIA is in preview (October 2026). Costs follow public pricing (Jev by TypeSafe: $0.042 per 1M input tokens, output free) plus the LLM price the user enters; answer shares, confidence and latency are user assumptions until calibrated on labelled data. VARIA is an independent project, not affiliated with TypeSafe. Modelled example (demo support flow, $3 per human touch as an assumption): machines cost $1,037 per million tickets, while 28% of tickets still reach a person ($904k at $3 per human touch).

## Pages

- [VARIA home (Markdown)](https://heyvaria.com/index.md): the whole landing page as Markdown — what VARIA is, the people bill, how it works, the what-if story, agents (MCP/CLI/WebMCP), why p50 jumps, use cases, comparison, field notes, FAQ, glossary
- [VARIA home (HTML)](https://heyvaria.com/): the same page with the interactive simulator

## Key answers

- [What is VARIA?](https://heyvaria.com/index.md#what-is-varia): VARIA is a decision-flow simulator for AI pipelines. You compose decision-model questions, LLM steps, code rules, confidence gates and human handoffs; VARIA runs 10,000 items through the flow and shows cost per million, p50 and p95 wait, and the share of work that still reaches a person, before you build it.
- [What does an AI support flow cost per million tickets?](https://heyvaria.com/index.md#the-people-bill): In the VARIA demo flow (modelled), the machines cost about $1,037 per million tickets: one decision-model call per ticket, plus an LLM that drafts refund replies. People still touch 28% of those tickets, which is 301,000 touches per million because some tickets need two. At $3 a touch that is about $904k, or $0.90 per ticket against $0.0010 for the machines. If the model’s confidence is calibrated, another 13.9% of tickets land in the wrong queue: $416k at $3 per reroute. All in, $1.32M per million, against $4.17M if people triage everything. LLM price calculators count only the machine part.
- [What does AI customer support cost per ticket? How much of a queue can AI handle? How much does a Jev request cost?](https://heyvaria.com/index.md#faq): answered in the FAQ with modelled numbers and public pricing
- [Why does p50 jump in what-if scenarios?](https://heyvaria.com/index.md#why-p50-jumps): Because a median only moves when half of the items change path. In the demo flow the fast path takes 0.3 s and the LLM path 6.6 s. While fewer than half of the tickets take the LLM path, p50 stays at 0.3 s; past half, it jumps to 6.6 s. The mean grows smoothly. Read p50 next to p95 and the mean, and watch the share on the slow path.
- [VARIA vs spreadsheets, Flowcost.ai, vendor ROI calculators, LLM observability, workflow builders](https://heyvaria.com/index.md#compare): Most teams estimate in a spreadsheet, trust a vendor’s ROI calculator, or build first and measure with observability. Each answers a different question.

## For agents

- [WebMCP tools on the home page](https://heyvaria.com/index.md#for-coding-agents): simulate_support_flow, get_pricing_assumptions, explain_term, list_use_cases, prefill_flow_request (never submits; the human does)
- [AI catalog](https://heyvaria.com/.well-known/ai-catalog.json): MCP server card for VARIA (preview; not publicly available, no remote endpoint)
- [Request flow modelling](https://heyvaria.com/#model-my-flow): a form with WebMCP declarative attributes (toolname="request_flow_modelling")

## Optional

- [Full text for LLMs](https://heyvaria.com/llms-full.txt): this file plus the full Markdown page, FAQ and glossary in one document
- [Classic edition](https://heyvaria.com/classic) and [DOS edition](https://heyvaria.com/dos): the same page in a plain product-site design and as a text-mode program; the home page is the canonical one