AI models in use

Last reviewed: June 5, 2026. Reviewed quarterly.

This page names every AI provider and model family Corial routes production traffic to, where the inference happens per tier, the provider’s training-data policy, their published safety framework, and the independent third-party assessments their models have received.

We do not disclose which specific model handles which specific task inside Corial. That routing is part of how the product works and changes over time as the providers ship new versions. The set of providers and the model families we use are stable and public.

We cite third-party scores rather than invent our own. Specifically: the Stanford Foundation Model Transparency Index (FMTI) for developer transparency, and the MLCommons AILuminate safety benchmark for model safety behaviour. Both are independent. Both are imperfect. Both are better than a vendor self-assessment.

What Corial uses AI for

Across the product, AI handles natural-language understanding, extraction from unstructured input (voice notes, emails, documents), summarisation, classification, drafting, reasoning over customer-relationship context, and grounded web lookups. Image and PDF understanding for ingested documents. No autonomous customer-facing communication; every external message is drafted by AI and sent by a human.

Anthropic

Model families in use	Claude Opus, Claude Sonnet, Claude Haiku
Region of inference	Default tier: United States, via Anthropic’s API. EU residency tier: European Union, served on Google Cloud Vertex AI under licence.
Training on Customer Data	Contractually not used to train Anthropic’s models, under their commercial terms.
Published safety framework	Responsible Scaling Policy (versioned, public)
Stanford FMTI 2025	36 / 100. Middle group of the industry, alongside Google, OpenAI, Meta. Stanford CRFM
MLCommons AILuminate v1.0	Very Good (Claude 3.5 Sonnet and Claude 3.5 Haiku). Newer Claude 4.x versions have not yet been included in the public v1.0 benchmark. AILuminate

Google

Model families in use	Gemini Flash, Gemini Flash-Lite
Region of inference	Default tier: United States, via Google AI Studio. EU residency tier: European Union multi-region, via Google Cloud Vertex AI.
Training on Customer Data	Paid API and Vertex AI: contractually not used to train Google’s models. See Google Cloud Data Processing Addendum.
Published safety framework	Google DeepMind Frontier Safety Framework (versioned, public)
Stanford FMTI 2025	36 / 100. Middle group of the industry. Stanford CRFM
MLCommons AILuminate v1.0	Good (Gemini 1.5 Pro, Gemini 2.0 Flash, Gemini 2.0 Flash Lite). Gemini 3.x versions have not yet been included in the public v1.0 benchmark. AILuminate

Mistral AI

Model families in use	Mistral Large, Mistral Medium, Mistral Small
Region of inference	France (European Union), via Mistral La Plateforme.
Training on Customer Data	Contractually not used to train Mistral’s models, under their commercial terms.
Published safety framework	Mistral AI safety position (public; lighter than Anthropic or Google)
Stanford FMTI 2025	15 / 100. Bottom group of the industry. Mistral is notably less transparent about training data and evaluation than US peers; this is a real trade-off for the EU-sovereign benefit they offer. Stanford CRFM
MLCommons AILuminate v1.0	Mixed across versions. Mistral Large 2402 with output moderation: Very Good. Mistral Large 24.11 and Ministral 8B (API, no moderation): Fair. AILuminate

Honest notes on the scores

The 2025 FMTI industry average dropped to 40 out of 100, down from 58 in 2024. Stanford’s reading is that transparency in the AI industry has regressed. Anthropic, Google, and Mistral all scored lower in 2025 than they did the previous year. We use these scores anyway because they are the best independent measure available, not because they are uniformly good.
AILuminate v1.0 was released in early 2025. Frontier models released later (Claude 4.x, Gemini 3.x, Mistral Large 3) have not yet been added to the public benchmark. We expect later versions to be added over time and will update this page when they are.
The FMTI measures developer transparency, not model safety. AILuminate measures model safety behaviour, not developer transparency. They answer different questions and we publish both.

Review cadence

Model selection is reviewed quarterly against the criteria published in the Responsible AI Charter: capability, region, training-data policy, published safety framework, third-party transparency assessment, cost and latency. The set of providers and model families on this page reflects production state at the last review.

Sub-processor changes (adding or removing a provider) are notified to customers at least 30 days in advance.