AI-Ready Data Without a Curated Warehouse

Kuudo makes Amazon data AI-ready for agentic analytics by landing source-shaped data in low-cost, customer-owned storage, keeping freshness and lineage beside it, pairing that context with domain expertise encoded as Skills, and using Agent Flow to govern how each workload reads and returns data. A curated warehouse remains valuable for stable shared metrics. It no longer has to predict every question before an agent can work.

Microsoft Copilot, Perplexity, ChatGPT, and other broadly available assistants bring capable general reasoning, and your competitors use them too. They know nothing about your business out of the box. Kuudo gives those same tools your data, your rules, and the way your business operates through Amazon MCPs, Skills, Atlas, and approval controls: your account as it is right now, your judgment running every time, and your call before anything changes.

AI-ready data starts with low-cost landed storage

Agent use cases do not arrive like a dashboard roadmap. A settlement discrepancy may matter for one afternoon. A catalog issue may begin with one ASIN and expand across a marketplace. An attribution question may disappear when its lookback window closes. Modeling every possible question before the data can be used puts the slowest step at the front of the system.

Kuudo starts by landing source-shaped Amazon data in low-cost storage the customer owns. That preserves the detail, source state, and history a use case may need before anyone knows the final question. Curation can happen later wherever repetition makes it worthwhile. The landed layer remains the durable base for questions that are new, temporary, or too specific to deserve a permanent model.

MotherDuck's research supports this separation. In its original vendor-run study of 460 questions, raw schema produced 29.8% accuracy and column comments produced 30.1%. Precomputed views reached 86.6%, while a simple prompt plus macros reached 93.2%. Comments alone did not supply enough meaning in that setup, but the strong view result also shows why curated models remain useful for repeated questions.

The later work moved richer context beside the data. MotherDuck described raw tables paired with a hierarchical domain layer and reported 100% across 445 questions after excluding five. It then described versioned Guides and reported 418 of 419 held-out questions correct with 55% lower cost than its discovery baseline. These are vendor-run, model-coupled evaluations. They are not independent results or a Kuudo benchmark, and Kuudo has not run an equivalent benchmark.

Kuudo reads that progression as an architecture signal: preserve the source-shaped data, put useful context beside it, and curate the questions that prove stable. A warehouse becomes one valuable product of the foundation instead of the admission ticket to every use case.

Context makes mutable commerce data safe to interpret

Landed rows are necessary, but they are not self-explanatory. Amazon records change, reporting surfaces use different clocks, and some metrics remain incomplete or are restated after first publication. The context beside the data must say where a value came from, when it was observed, and how Amazon defines it.

Amazon Agent Atlas grounds four examples that make the requirement concrete:

1. In the Vendor Orders documentation retrieved by Atlas, Amazon may change a purchase order after creating it. The changed order carries purchaseOrderChangedDate, the time Amazon updated it. Snapshots taken before and after that timestamp represent different source states.
2. Under delivery-date reserves, Payments Dashboard sales components may not match Business Reports because statements and settlements include transactions when funds are released. The same order can appear under different dates across the surfaces. Neither surface is necessarily wrong; each uses a different clock.
3. Amazon Ads says conversions may take 12 hours to appear and attribution metrics for a report date remain incomplete until the lookback window ends. Attributed sales are reported on the interaction date even when the sale occurs later within that window.
4. The Sponsored Products methodology says reporting is adjusted for offline sophisticated invalid traffic and restated within a 30-day timeframe, with rare quality revisions within 90 days. That statement applies to the documented Sponsored Products methodology, not every Amazon report.

These facts do not make star schemas useless, and they do not mean every Amazon record changes. They show why time and lineage belong in the meaning of a commerce metric. “Sales for Tuesday” remains incomplete until the source surface, observation time, and derivation are attached.

Amazon Agent Flow's documented contract keeps dataset references, field definitions and classifications, freshness, and lineage available to the agent. Atlas supplies the Amazon-specific rule context. One tells the agent which state it has; the other explains how Amazon says to interpret that state.

Skills turn domain expertise into reusable infrastructure

Freshness and lineage explain the source. They do not capture how a business decides. Domain expertise includes the operator's classifications, exclusions, thresholds, known exceptions, comparison logic, and the sequence that turns a finding into action. Kuudo encodes that expertise as Skills beside the landed data.

A Skill can state that settlement-date sales require reconciliation before comparison with an order-date KPI. It can define which attribution window must close before a pacing decision becomes final. It can package a curated analysis, schedule, and operation contract so the same judgment runs again without being rebuilt from a chat transcript.

Atlas and Skills play different roles. Atlas supplies Amazon-specific grounding from the documented source. Skills hold the customer's application of that knowledge: which rule matters here, which exception the business accepts, and what the workflow should return. Keeping both beside the data makes the context useful without baking every decision into a physical table.

That expertise also travels. An authorized AI client can use the same Skill because the contract lives with the customer's system rather than inside one conversation. Models and interfaces can change while the business keeps its definitions, exceptions, and operating sequence. Private. Trusted. Owned.

Agent Flow is the runtime ephemeral workloads require

An ephemeral use case can still be large. An agent may inspect schema, test a hypothesis, revise the query, expand from one entity to thousands, and request another result seconds later. The question may vanish after the decision, but its workload still needs limits while it exists.

MotherDuck's analysis of agent query behavior found agents issuing 29 times as many queries as humans, with a four-second median gap between agent queries versus 60 seconds for people. Agent queries scanned a median of about 7,000 rows versus 200,000 for human queries. Those figures describe MotherDuck's observed workload. They are not a universal agent law and do not predict a Kuudo customer's bill.

Kuudo worked backward from that workload shape to Amazon Agent Flow. The foundation must let an authorized agent discover landed datasets and their context, inspect the proposed work, run within an enforced boundary, and keep large results out of the conversation. Agent Flow supplies that runtime contract:

• A scan estimate exposes the bytes a proposed query would inspect.
• A scan ceiling rejects a governed query that exceeds its allowed budget.
• A bounded preview returns enough evidence to inspect the result.
• A durable result handle keeps the full output in the customer's environment.
• Server-side analysis lets later steps work without moving the full table into chat.

The ceiling applies to each governed query, not total cloud spend across an investigation. Every later query is checked again. This makes experimental and short-lived use cases governable without first turning each one into a permanent analytics product. If a use case becomes recurring, its Skill and schedule can scale the operating pattern; if its metric becomes broadly stable, that output can graduate into a curated model.

What happens next

Build the foundation in the order agent work requires it. Land Amazon data in customer-owned storage first. Preserve source state, freshness, and lineage. Attach Atlas grounding and encode domain expertise as a Skill. Then define the per-query scan ceiling, preview boundary, result handling, and approval path in Agent Flow.

Use curation where the workload earns it:

This order keeps optionality in the data and moves repeatable knowledge into context. A one-time reconciliation can run without preliminary curation. A catalog-wide audit can reuse the same foundation at a larger scope. A stable metric can become a view without cutting the agent off from the landed detail that explains an exception.

That is the game changer: low-cost landed data, source-aware context, customer-owned domain expertise, and a runtime built for the shape of agent work. The curated warehouse still has a job. It simply stops being the gate every new question must pass through.

Next: why everyone has the same AI, and where your defensible advantage actually lives.