← intelligencewarehouse.com Request a demo
Intelligence Warehouse · Research

How the warehouse actually works

Five papers on the architecture, written for people who want to check the reasoning rather than take the claim. Each one carries its own seeded simulation or worked examples, states what it assumes, and ends with a ledger of where it could be wrong. Every figure is reproducible from the script that produced it.

Authors Akhil Singh · Nidhi Prabhu · Aniruddha Tammewar Published July – September 2026 Papers 5
Paper 01·September 2026·Seeded simulation, real graph builds
How AI finds the right answer in a big company
Answers are decided by what reaches the model. Similarity search fills that space with near-duplicates from other departments; walking a typed graph reaches the right material instead, at a cost that does not grow with the company.
21%of needed rules survive fixed-budget retrieval
−3.5%context growth as the graph grows 60×
−50.9%cost per query across year one
Paper 02·September 2026·420 circuits, 12 months
Where your decision rules actually live
Your recurring decisions have no system of record: some of it is written down, most of it is in people's heads, the rest only shows up when a case goes wrong. This is how to capture all three.
81%of failures trace to rules no document contains
33.5% → 1.9%intervention rate over one year
98%of the rulebook captured by month 12
Paper 03·September 2026·200 cells × 52 weeks
Getting more from the models you already have
The forecast is decent and still moves the business less than it should, because the number arrives stripped of its error bars, blind to what other teams know. Same model, better decisions.
±0.0%change to the model itself
−20%decision cost across the year
−92%festival stockouts once the first became a rule
Paper 04·July 2026·Seeded Monte Carlo, n=1,000
Why enterprise AI costs so much
Most enterprise AI puts the whole program inside the prompt, so every routing step is a paid inference over the whole working set. Run that as ordinary code instead and the bill changes shape.
95.97%fewer tokens processed per task
87.54%lower run-cost after caching the baseline
24.8×token throughput multiple
Paper 05·September 2026·Method paper, worked examples
Starting with what each team is for
Grounding, not intelligence, is the bottleneck, and nobody is trained to explain how their team works. Every team has a goal, every goal has milestones, every milestone has workflows. Start there.
3statements in the whole model
8questions to capture one team
45 minto goal, milestones and workflows

How they fit together

Start with what each team is for. It is the method for getting a company's goals, milestones and workflows written down in days rather than months. Where your decision rules live then covers how one recurring decision gets captured and kept current, and how AI finds the right answer covers what happens when someone actually asks a question of everything those two produced.

The models you already have is about plugging existing ML into the same picture, and why enterprise AI costs so much is the cost argument underneath all of it: what changes when the control flow runs as ordinary code instead of as tokens.

If you only read one, read how AI finds the right answer. It is the load-bearing claim, and the other four either feed it or price it.

intelligencewarehouse.com