How Achi operates as a human-governed, deterministic AI workflow platform — protecting customer data, maintaining strict tenant isolation, and committing to no model training on your content.
Core standard: Achi is an AI-powered operational assistant, never an autonomous decision-maker. AI generation is always a draft; every operational transition, approval, and state change stays human-authorised and rule-bound.
01 — Foundational AI Principles
Architectural Controls & Ethical Framework
Human-in-the-loop. AI generates proposals, summaries, and recommendations; authorised humans retain sole authority to publish, approve, or execute.
Grounding & safe fallback. AI answers are grounded in the org’s own records and form context; where context is missing, Achi falls back safely rather than inventing facts.
Transparency. AI-generated content and suggestions are labelled in the interface.
No model training on your data. Customer prompts, submissions, uploaded documents, and workflow logs are not used to train, fine-tune, or evaluate foundation models.
Deterministic rule enforcement. Business logic, routing, permissions, and SLA calculations run in a deterministic database engine, independent of the AI.
Role-bound. AI features operate strictly within the authenticated permissions and business role of the requesting user.
02 — Functional Boundaries
What AI Can and Cannot Do in Achi
What AI can do
Draft form schemas, section layouts, and field validation from a natural-language prompt.
Propose multi-stage workflows, business roles, and SLA targets for your review.
Summarise multi-page submissions, contracts, and attachments for human reviewers.
Detect missing information or inconsistent fields in a submission.
Answer questions about your own records via the “Ask ACHI” copilot.
What AI cannot do
Execute payments, approve requisitions, or sign contracts autonomously.
Publish or deploy forms and workflows without an authorised human confirming it.
Override approval hierarchies, business roles, or security policies.
Bypass form validation, required reviewer checklists, or role permissions.
Expose restricted data to unauthorised users during generation or summary.
03 — Data Protection & Model Providers
Subprocessor Handling & Data Lineage
Achi orchestrates AI through a single provider gateway. We apply the following safeguards with the underlying model providers (e.g. Anthropic, OpenAI, Google):
No training / no retention. Prompts sent to model subprocessors are configured for no-retention processing and are not stored or used by vendors for quality control or retraining.
Data minimisation. Only the data needed for the task (e.g. field context for a form draft) is sent; system secrets, credentials, and database keys are never included.
Provider failover. If a model provider is unavailable, Achi routes to a fallback provider or degrades safely to manual, non-AI operation.
Dormot Technologies Ltd · Achi AI Governance Policy