Engineering
Principles that guide how we designed the BooleanMaths Agentic Stack
Pointing an AI agent at raw marketing data creates a false sense of precision. We designed BooleanMaths around a different premise: accuracy is a context and structure problem, not a query-generation problem. Here are the principles behind our agentic stack — and the independent validation that followed.
Our Principles
Four layers, one governed answer
Most wrong answers in AI analytics come from three failure modes: the agent picks the wrong entity, the data or definition has gone stale, or the right information exists but the agent never finds it. We built three layers to close those gaps — and a fourth to verify every answer and keep the system honest as brands and data change.
LAYER 01 — GOVERNED DATA
Marketing models built into the dataset
Revenue, ROAS, CM1/CM2 and attribution don't resolve to forty plausible tables — they resolve to one governed definition. Multi-touch attribution, SKU-level P&L and our survey-calibrated corrections are computed as the semantic layer, with India-specific realities (cart_token loss, RTO/COD noise, long journeys) already baked in. Ambiguity collapses before the agent ever searches.
LAYER 02 — RESTRICTED MCP TOOLS
A bounded tool set that answers vague queries deterministically
Instead of letting a model write arbitrary SQL against a warehouse, the BooleanMaths MCP exposes a curated set of tools — channel attribution, product contribution margin, RTO breakdown, creative intelligence and more. A loose question like "why did margins drop last month" maps to a bounded set of governed tools that return the same answer every time. Structure, not raw access, is what makes it reliable.
LAYER 03 — BRAND CONTEXT
Measurement done inside the business context
An agent that doesn't understand your business answers what you asked, not what you meant. BooleanMaths carries each brand's economics, prior decisions and terminology into every analysis, so a number arrives already interpreted against the brand's own context. This is the layer most tools skip — and the one that turns raw measurement into a decision.
LAYER 04 — VALIDATION
Continuous evals with a direct line to support
Governed answers only stay trustworthy if you keep checking them. We run internal evals against the MCP tools and marketing models to catch regressions before a brand ever sees one, and route every correction from client support straight back into those evals and reference docs. It's the loop that keeps answers accurate as integrations, definitions and brands change — and the layer that catches the silent, plausible-looking wrong answer.
Independent Validation
Anthropic's own data team reached the same conclusions
We shipped the BooleanMaths agentic stack in April 2026. In June 2026, Anthropic published how its own data science team enables self-service analytics with Claude — and the architecture they recommend is the one we had already built. Their headline: pointing an agent at a raw warehouse creates a false sense of precision, because accuracy is a context and structure problem, not a code-generation one.
APRIL 2026
BooleanMaths ships the agentic stack
Governed marketing models, a restricted MCP tool layer, per-brand context, and continuous validation — live in production for D2C brands.
JUNE 2026
Anthropic publishes matching findings
Anthropic's data team documents the same three-layer approach — governed sources of truth, structured tool routing, and business context — as the way to make AI analytics trustworthy.
The same three failure modes — and the same fixes
ENTITY AMBIGUITY
Anthropic: collapse to one governed answer
Their fix is a small set of canonical datasets and a semantic layer the agent must use first. Ours is the governed marketing-model layer — attribution, CM1/CM2 and corrections resolve to a single definition.
RETRIEVAL FAILURE
Anthropic: structure beats raw access
Their key finding: giving an agent raw access to thousands of prior queries barely moved accuracy — structure did. Our restricted MCP tool set routes every question to the right governed tool instead of open-ended SQL.
MISSING CONTEXT
Anthropic: the layer most teams skip
They call business context the piece most teams skip and underrate longest. It's built into BooleanMaths from the start — each brand's economics, decisions and terminology travel with every query.
Activate your Marketing Data with BooleanMaths
LAYER 04 — VALIDATION