Retrieval reduces the distance to information. It does not reduce the distance to truth. Before an agent can use organisational context in consequential work, the sources still need accountable ownership, consistent meaning, known freshness, appropriate permission, traceable provenance and a route for correction.

Retrieval is not repair

An enterprise agent can search policy documents, product records, service tickets, contracts and operational systems in seconds. That is valuable. It removes a practical barrier between a person and the information required for a task. It can also create a persuasive answer from sources that disagree, use the same word differently or should never have been combined.

The retrieval layer decides what enters the model's working context. It can rank, filter, chunk and attach metadata. It cannot decide which business owner has authority when two sources conflict. It cannot make an obsolete policy current, restore missing consent or determine whether a local definition should become an enterprise one. Those are organisational and product decisions expressed through data, governance and operating practice.

Anthropic describes context engineering as the curation of instructions, tools, external data and interaction history within a finite resource. The useful lesson is selectivity. More accessible information is not automatically better context. The right question is whether the selected information is fit for this decision, for this user, at this moment and at this level of consequence.

SourcesAnthropic

A coherent answer can hide an incoherent organisation

Consider an account agent preparing a renewal. The customer record says the contract renews in September. A signed amendment in document storage says November. The billing platform contains a product code nobody currently owns. A support note promises a service level that the contract does not recognise. Search can retrieve all four records. Generation can turn them into one fluent account summary. Neither operation establishes which commitment governs the renewal.

A person who knows the estate may notice the contradiction and ask legal, finance or the account owner to resolve it. An agent optimised to complete the task may instead smooth the differences into one confident recommendation. The interface can make the answer feel more authoritative than the underlying information. Fragmentation has not disappeared. It has moved behind a synthesis layer and closer to action.

AWS's Generative AI Lens identifies fragmented environments, inconsistent governance and unclear ownership as constraints on generative AI scale. Its data-architecture guidance also treats quality, authorised access, versioning, lineage and freshness as continuing requirements for retrieval. The specific services are not the point. The broader implication is that access and readiness are different system properties.

SourcesAWS

Run six tests against the intended action

Do not begin with an enterprise-wide score for data maturity. Begin with the action the agent is expected to take and the exact sources it will use. A knowledge assistant that drafts an internal briefing can tolerate different uncertainty from an agent that changes a price, approves a claim or writes to a regulated record. Readiness is consequence-specific.

The following diagnostic turns context from a vague prerequisite into six conditions that can be inspected, owned and changed. A source does not need to be perfect. It does need to be good enough for the proposed action, with any remaining uncertainty made visible in the operating boundary.

  • Authority: Is there an accountable owner who can decide which source governs and resolve a conflict?
  • Meaning: Do important terms, entities, units and states mean the same thing across the sources being combined?
  • Freshness: Can the system identify when each source changed, when it expires and what supersedes it?
  • Permission: Is access evaluated for the user, agent, purpose and action at runtime rather than inherited from broad system access?
  • Provenance: Can a reader trace a material claim or action to the exact source, version and transformation that produced it?
  • Correction: Can an operator challenge the result, route the defect to an owner and stop the same error returning?
SourcesNIST; W3C; Microsoft

The operating boundary should reflect context confidence

A failed readiness test does not always require the use case to stop. It changes what the agent may do. When authority is unclear, the agent can present conflicting sources instead of selecting one. When freshness is uncertain, it can draft but require confirmation before communication. When provenance is incomplete, it can retrieve for discovery but remain outside the system of record. The permitted action should narrow as the evidence weakens or the consequence rises.

This is why context readiness belongs with product and workflow design, not only with a data remediation programme. The experience must show source, currency and uncertainty at the point where a person makes a decision. The workflow must include an escalation route. The agent's tools must enforce the chosen limit. Observability must retain enough identity, retrieval provenance, permission and outcome evidence to reconstruct what happened without collecting information indiscriminately.

NIST's AI Risk Management Framework places intended use, context, roles, monitoring and lifecycle governance around the complete AI system. Microsoft similarly recommends recording identity context, retrieval provenance, tool invocations, permissions and outputs. These sources do not prescribe one product experience. They support a practical design principle: context confidence must change system behaviour, not sit in a readiness spreadsheet.

SourcesNIST; Microsoft

Agents can support repair without owning the truth

There is a legitimate counterargument. Agents can compare document sets, cluster duplicate terms, identify stale references, propose mappings and surface records that appear inconsistent. Used carefully, those capabilities can make context repair faster and reveal problems that manual inventory would miss. The agent can be an effective diagnostic instrument.

The boundary is approval. A generated reconciliation should not become the new source of truth merely because it looks plausible. A domain owner must decide whether two terms are equivalent, whether an exception is deliberate and whether a source can be retired. The accepted decision then needs to return to the source, semantic model, policy or workflow so the improvement is reusable. Otherwise every agent session repeats the same interpretation.

This creates a productive sequence: expose, compare, decide, repair and observe. The agent helps with the first two steps. Accountable people and governed systems complete the decision and repair. Production evidence then shows whether the new context behaves as intended. That is a stronger use of AI than asking a retrieval layer to conceal unresolved organisational work.

Choose one context disposition before production

Finish the diagnostic with a decision, not an average score. Proceed when the six conditions are sufficient for the intended action. Proceed with constraints when the agent can assist safely through narrower tools, visible sources, confirmation or a reversible workflow. Repair before action when a missing owner, conflicting meaning or unreliable source would materially alter the result. Stop when the required context cannot be used lawfully, safely or accountably.

Record the source set, failed tests, operating restriction, repair owner and evidence needed to change the disposition. Reassess when a source, workflow, permission model or consequence changes. The result should be a small control attached to a real use case, not another enterprise assessment that becomes stale before delivery begins.

Agents are useful partly because they make the cost of fragmented context impossible to ignore. They show where knowledge is duplicated, where decisions have no durable home and where access has outrun accountability. Treat that exposure as evidence. Then repair the organisational system that produces context, rather than expecting the agent consuming it to become the system of record.

  • Disposition: Will the agent proceed, proceed with constraints, wait for repair or stop?
  • Intended action: Which decision or action will the context support, who is affected and what is the consequence of error?
  • Source set: Which sources, versions and owners are authoritative for this use case?
  • Failed conditions: Which tests of authority, meaning, freshness, permission, provenance or correction remain unresolved?
  • Operating restriction: Which narrower tools, visible sources, confirmations or reversible steps make constrained use acceptable?
  • Repair ownership: Who must resolve each material gap, and where will the accepted correction become durable?
  • Reassessment evidence: What must change before the disposition or permitted action can expand?
SourcesNIST; AWS

Sources and further reading

  1. AnthropicEffective context engineering for AI agents
  2. AWSData architecture, Generative AI Lens
  3. NISTAI Risk Management Framework Core
  4. W3CPROV Overview
  5. MicrosoftObservability for generative AI and agentic AI systems