The Information Supply Chain Problem

A New Architecture Problem — and a Broader One

Researchers from UC Berkeley's Sky Computing Lab published a paper at CIDR 2026 titled "Supporting Our AI Overlords: Redesigning Data Systems to be Agent-First" arguing that database infrastructure is not designed for the workloads AI agents generate, and needs to be rebuilt from the ground up. Their central concept is agentic speculation — the high-throughput, exploratory querying pattern agents use: issuing enormous volumes of probes to discover schemas, validate partial solutions, and establish grounding before they can formulate a response. Their empirical studies show this approach improves task accuracy by 14 to 70% over single-shot retrieval, and that providing grounding hints reduces agent query volume by up to 37%. Their proposed response — probe interfaces beyond SQL, optimizers that satisfice rather than fully solve, agentic memory stores — is the right research agenda for the database layer. Data systems will need to evolve in these ways.

But there is a complementary problem the paper does not address — one at the layer between retrieval and action — that in enterprise deployments may be the more consequential of the two.

A Day in the Life of an Agentic Business Workflow

Consider a five-stage autonomous order-to-cash process. At each stage an agent acts — checking credit, confirming inventory, generating a sales order, creating a shipping instruction, issuing an invoice. It writes to systems of record and makes commitments on behalf of the enterprise. Now introduce a single error.

The agent proceeds. It commits the full quantity, sets a ship date, generates the sales order, selects a carrier, issues the invoice. The workflow completed correctly in every visible sense. The customer received a confirmed ship date. And the enterprise is now committed to an order it cannot fulfill — with the error embedded in multiple systems of record and committed to an external counterparty before anyone realized something had gone wrong. This is not a failure of the data system layer in the sense the Berkeley paper addresses. The failure was in the information supply chain — the chain of verified, grounded, authoritative information the agent depended on at each stage.

The scenario above describes a single error. But agentic workflows expose a more structural problem: even individually reasonable per-stage accuracy rates produce unacceptable end-to-end results when errors compound.

If each stage of a five-stage business process operates at 85% accuracy — accounting for retrieval precision, entity resolution, and synthesis faithfulness — end-to-end accuracy is 0.85⁵: approximately 44%. The wrong inventory figure at Stage 2 becomes the basis for the sales order at Stage 3, the shipping instruction at Stage 4, the invoice at Stage 5. The error propagates silently.

The problem is structural. Getting above 90% end-to-end accuracy across five stages requires per-stage accuracy above 98% — a threshold no current LLM consistently achieves on complex enterprise data. And there is no incremental path out through model improvement: a hypothetical model achieving 95% per stage still produces only 77% end-to-end accuracy. The solution requires a mechanism that resets the error budget at each stage rather than allowing errors to compound.

The Information Supply Chain

Businesses have struggled with information quality for as long as they have run data-driven operations. Data professionals invest enormous effort cleansing, deduplicating, reconciling, and validating data across pipelines — and errors still surface at the point of consumption. The discovery of information quality failures at the point of consumption is one of the defining frustrations of enterprise data work.

Agentic AI does not change this underlying challenge. It makes it structurally more dangerous. When a human decision maker receives incorrect information, they may catch it through intuition, cross-reference, or domain experience. They can pause, question, and escalate. When an autonomous agent receives the same incorrect information, it acts. There is no pause, no intuition, no escalation trigger. The error propagates — silently, confidently, and at machine speed — through every downstream stage of the workflow.

The Berkeley paper addresses the infrastructure layer of this problem — the pipes through which information flows to agents. But a supply chain is not reliable because its logistics are efficient. It is reliable because every link delivers what it promises. The answer requires reasoning through cross-reference and verification: confirming that the specific claims an agent is about to act on are grounded in authoritative source documents, with every number traced to its exact origin and discrepancies surfaced before the agent acts.

Consider the difference between two responses to an agent querying available inventory:

The first response was retrieved. The second was reasoned — grounded against the source record, cross-referenced against related data, and delivered with a provenance chain that makes the discrepancy visible before the agent acts. The difference is not in the quality of the underlying model. It is in whether the information supply chain treats verification as a structural requirement at every stage.

Resetting the Error Budget at Each Stage

LAKEer is built on this principle: verification embedded at every pipeline stage, not bolted on at the end. An agentic workflow invoking LAKEer receives: ingestion into a heterogeneous knowledge graph (graph, vector, and tabular stores) with domain ontologies overlaid for entity resolution; six-channel retrieval fusing semantic search, keyword, graph traversal, and structured SQL; LLM synthesis of the fused evidence; claim-by-claim cross-reference of the synthesized answer against the knowledge graph, with unsupported inferences flagged or discarded; and provenance delivery — each claim linked to its exact source document, page, or table row. By verifying at the output of each retrieval invocation, LAKEer interrupts the compounding chain. Each workflow stage receives independently verified, provenance-backed information rather than unverified synthesis that cascades errors forward.

On Google DeepMind's FACTS Grounding Benchmark — 860 examples measuring whether AI-generated claims are traceable to source documents — LAKEer achieved 77.7% accuracy and 92% eligibility using GPT-4o-mini, exceeding Gemini 2.5 Pro's published score. The architectural implication is direct: the same underlying model without LAKEer's verification pipeline would score materially lower. Factual grounding is an architectural property, not a model capability.

Where the Problem Is Most Acute

The information supply chain problem is most severe where documents are complex and domain-specific, entity references are ambiguous, decisions are high-volume, and errors are irreversible. Three industries illustrate this:

In energy and oil and gas, drilling operations generate complex domain documents — well logs, NPT records, stratigraphic analyses — with specialized terminology and entity references that may be expressed inconsistently across sources. A confabulated answer may inform a well intervention decision or regulatory submission with consequences that are difficult to reverse. LAKEer's OSDU domain plugin addresses this with IADC-classified NPT extraction and stratigraphic correlation built into the graph construction pipeline.

In automotive and manufacturing, agentic procurement workflows must correctly resolve supplier entities across naming conventions, apply the correct contractual terms per component category, and confirm specification compliance. Failures go beyond incorrectly priced orders — misapplied contract terms create commercial liability, specification failures create quality risk, qualification errors create audit exposure. The information supply chain must be reliable at the claim level.

In financial compliance, regulatory reporting requires every figure to be traceable to an auditable source. Standard RAG systems can retrieve relevant documents and generate plausible summaries. They cannot certify that a specific claim corresponds to a specific figure in a specific authoritative source. Verified knowledge systems can — and in regulated industries, that capability is the difference between a system a regulator can audit and one that creates liability.

Two Problems, One Architecture Vision

The Berkeley paper and the information supply chain argument address different layers of the same challenge. The Berkeley group is correct that data systems need to evolve for agentic workloads — without this, agentic workflows are bottlenecked by infrastructure designed for human query patterns. But efficient data systems determine how quickly information flows to agents, not whether what arrives is correct. Without a verified information supply chain, agentic workflows are unreliable in a way no database engineering can fix.

Reliable agentic AI in the enterprise requires both: an agent-first data system that efficiently serves agentic probe workloads, and a verified information supply chain that ensures every claim is grounded before the agent acts on it. The organizations that solve both will deploy autonomous agents across complex business processes with confidence. Those that solve only one will find their workflows either too slow to scale or too unreliable to trust. The series reliability mathematics do not care which was neglected.

The productivity promise of agentic AI is autonomous execution of complex business processes at scale. That promise is only redeemable if the information agents act on is trustworthy at every stage.