Responsible AI

Responsible AI at BHASM

How decisions get made. What's explainable. How to audit.

THE ENGINE

What the engine actually does

BHASM reads observable customer behaviour — purchase history, engagement signals, cycle position, world context — and produces three things per customer per cycle: a score, a recommended action, and the reasoning behind both.

It does not invent intent. It does not synthesize facts. Every score is derived from data the operator’s own systems already record.

Message text is generated with the assistance of AI, calibrated to the operator’s brand voice, and passed through automated content safeguards before any send. Scoring and timing decisions come from BHASM’s internal engine — no external model decides who gets contacted or when. Operators can review, edit, or stop any message from the terminal.

PER CUSTOMER

What’s explainable per customer

For any customer the operator can ask “why this score?” — and BHASM returns a list of contributing factors with relative significance.

example explanation
01

Silence duration

45 days since last order · primary signal

02

Cycle elapsed

1.5 cycles overdue · significant

03

Spend trajectory

last three orders declining · moderate

04

World context

pre-festival window · moderate

05

Relationship balance

steady, responsive · stable

Every decision is recorded with the world signals live at the time and the safeguard that let it through, and the record exports as a stream you can hand to an auditor. What you cannot do is re-run a past decision against today’s engine — a replay would answer with today’s model, not the one that decided.

PROTECTED

What’s protected

Exact model coefficients, aggregation formulas, and proprietary calibration logic are protected intellectual property. This ensures the engine remains effective, continuously improving, and difficult to replicate. All decisions remain fully auditable for compliance and review needs.

THE SAFEGUARD STACK

How decisions are checked

Every outbound action passes through a safeguard stack before firing. The stack enforces:

01

Default silence — no action without an observable signal

02

Permanent opt-out respect — once a customer says stop, that decision never reverses

03

Frequency protection — no over-contact in any window

04

Sentiment safeguards — open complaints halt all outreach

05

Account scope — every query names the account it may read

MEASUREMENT

How accuracy gets measured

BHASM records each outcome — returned, lost, or still inconclusive — and reports precision, recall and false-positive rate for the window you pick. Rows with no settled outcome are excluded rather than counted as wins. Below thirty settled outcomes the page reports nothing at all: it names the sample it has and waits, because an accuracy figure off twelve rows is a number, not evidence. Ask us for the methodology.

COMPLIANCE

Compliance posture

DPDP Act 2023 (India) — enforced

GDPR (EU) — enforced

CCPA (California) — enforced

Isolated Airlock encryption — every tenant under its own derived key

Data Processing Agreement available on request — write to hello@bhasm.ai

For audit log access, compliance documentation, or to ask a specific question about how the engine handled a specific decision — hello@bhasm.ai.