Product · Architecture & security
For the reader who checks the claims
How a question becomes a governed answer, where it runs, and the mechanism the whole three-audience model rests on.
Config-driven ETL
New sources onboard from configuration, not code — POS, loyalty, ERP, DMS, handheld — in days. This is what “live in weeks” rests on.
Multi-tenancy
The Advanced Analytics Platform is multi-tenant by design: every tenant works against their own permissioned view — their products, outlets and numbers, baked in. Without this, letting outside parties in is a policy; with it, it is a mechanism.
Multi-agent decomposition
A commercial question is decomposed — scope resolution against the governed model, metric retrieval, comparison, narrative — and each step is constrained to governed definitions. The answer carries its scope because the scope was resolved first, not inferred afterwards.
MCP integration
Platform capabilities — Report Builder among them — are exposed to Caisy as MCP tools. A conversation can build a report; the toolset extends without retraining anything.
Retrieval over your own data
Answers are computed from the customer’s own harmonised data at query time — not from anything memorised in a model. No customer data is used to train models.
Model-agnostic, within your Bedrock estate
Models are swappable within your own cloud environment — on AWS, that means Bedrock — so you get current engines without chasing whatever launched last week, and without data leaving your account.
Cloud-agnostic, your account
Deployed in your own cloud environment — AWS preferred, and the deployment we know best — alongside the existing warehouse or lake. Databricks and warehouse estates are complemented, not replaced. Data residency follows your account.
Workflow automation
Extensible workflow automation connects answers to actions — alerts, triggered actions, distribution of results — configured, not coded.
Permissions the AI cannot get wrong
Caisy never chooses what you are allowed to see. Every user queries a view with their permissions already baked in, and the platform inserts the entitlements itself — deterministically, on every query. The AI answers the question; it never gets to decide the scope.
The unauthorised rows are not filtered out of the answer — they are never in the view the query runs against.
There is no wrong view for the AI to pick.
There is no wrong view for the AI to pick.
Security posture
Enterprise-grade AI. Secured, certified, yours.
Every model runs inside an audited cloud perimeter, on your own data — certified, compliant, and fully under your control.
ISO 27001
ISO 42001
SOC 2 Type 2
GDPR-aligned
Your data, your tenant
Role-based access control
Microsoft Entra ID SSO
Deployed in your own cloud
No customer data used to train models