Making business data accessible to an AI agent is not the same as making it safe to use. The agent must retrieve the right figure for the right entity and period, with the correct unit and enough context to show where the answer came from.
That distinction matters whenever an agent answers customer questions, prepares management reports, recommends actions or initiates workflows. A confidently delivered number can still be wrong because it came from an obsolete spreadsheet, mixed monthly and annual periods, used the wrong regional definition or omitted an important qualification.
TechCrunch reports that the United Nations is turning to Google to help make its global data ready for AI agents. According to the report, the initiative follows a UNICEF test in which leading AI models reportedly struggled to retrieve global development statistics accurately.
The public account is a useful warning, but it is not evidence that a particular technical solution works. It does not identify the models tested, quantify their error rates, disclose Google’s architecture or provide implementation results. Managers should therefore treat it as a narrow case study in the difficulty of retrieving statistics, not as validation of a vendor or platform.
The broader operational lesson is that AI readiness must be tested at the answer level. It is not enough to confirm that an agent can connect to a database.
Start with defined authoritative sources
Many organisations have several plausible versions of the same fact. Revenue may appear in the accounting system, a sales dashboard, a planning workbook and a board presentation. Customer status may differ between the billing platform and customer relationship management system. An agent given access to all of them has more information, but not necessarily a reliable way to choose.
For each business concept the agent may use, designate an authoritative source and document its scope. Managers should be able to answer:
- Which system owns the approved value?
- Does that authority apply to historical reporting, current operations or both?
- Which source takes precedence when systems disagree?
- Who can approve exceptions or corrections?
Authority can depend on the question. The finance ledger may govern recognised revenue, while the sales platform governs the current pipeline. That distinction must be explicit rather than left for the agent to infer from column names.
This source register should remain selective. Connecting every available folder and application can increase the number of conflicting or obsolete records presented during retrieval. Businesses still organising their underlying information can begin with a broader knowledge-system readiness process for autonomous AI.
Document what each value means
A value without its definition is not an answer. Before making a dataset available, document the business meaning of important fields and indicators. At minimum, cover:
- Definition: What is included and excluded?
- Unit: Is the value expressed as euros, thousands of euros, percentages, people, orders or another measure?
- Geography or organisational scope: Does “Europe” mean the EU, a sales territory or every European country in which the company operates?
- Reporting period: Is the value daily, monthly, quarterly, year-to-date or based on a rolling period?
- Status: Is it provisional, forecast, revised or final?
- Calculation method: Is it recorded directly or derived from other fields?
Definitions must also distinguish terms that employees may use interchangeably. “Customers” could mean registered accounts, paying organisations, active users or contracts. “Margin” could refer to gross margin, contribution margin or another internal measure. If the business has not resolved the ambiguity, the agent cannot reliably resolve it on the business’s behalf.
The UNICEF Data Quality Framework discusses accuracy through dimensions including completeness, precision, reliability and coherence. It also explains the difficulty of assessing accuracy when the true value is unknown. For a business, the practical implication is to evaluate whether data and answers are coherent, sufficiently complete and fit for the stated decision rather than treating accuracy as a simple binary property.
Attach metadata and provenance to results
Metadata should travel with the data rather than live only in a separate manual. An agent retrieving a value should also be able to retrieve its definition, source, period, unit, status and last update time.
Provenance creates an auditable route from an answer back to its origin. Depending on the system, this route might identify the source application, dataset, table or document, record, transformation and publication timestamp. It should allow a reviewer to investigate how the answer was produced without relying on the model to reconstruct its reasoning.
Provenance is particularly important when information has been transformed. A dashboard total may depend on filters, currency conversion, exclusions and calculated fields. Linking only to the dashboard homepage may not reveal those steps. The trace should be specific enough for an authorised reviewer to reproduce or verify the result.
Give every dataset an update owner
AI agents can make stale information easier to consume, which can make it more dangerous. A polished answer may conceal the fact that its source has not been refreshed.
Assign a named business owner for each consequential dataset. That owner should define the expected update frequency, acceptable delay, correction process and treatment of late or incomplete inputs. The operating rules should also say what the agent does when freshness requirements are not met.
Depending on the use case, the safe response may be to label the value as outdated, use the most recent approved period with a warning, request human confirmation or decline to answer. Silently substituting old data should not be the default.
Control access at retrieval and action levels
An employee’s permission to ask an agent a question should not automatically grant access to every source the agent can reach. Retrieval should respect the underlying user’s role, purpose and applicable restrictions.
Managers should separately define permission to:
- discover that a dataset exists;
- retrieve individual records or aggregated values;
- view sensitive metadata or source documents;
- combine data from different systems;
- export or retain an answer;
- take an action based on the answer.
This separation matters because aggregation is not always harmless, and combining two permitted datasets can reveal information that neither exposes alone. Customer-facing agents need especially narrow boundaries so that one customer’s records cannot appear in another customer’s response.
Test retrieval with representative questions
A connection test proves that the agent can reach a source. A readiness test asks whether it returns the correct answer in context.
Create an evaluation set from real questions users are likely to ask. Include straightforward lookups, comparisons, ambiguous wording, requests spanning multiple periods and cases where no approved answer exists. Add questions designed to expose predictable mistakes, such as similarly named products, revised figures, missing months, conflicting documents and values expressed in different currencies or units.
For every test, prepare an approved reference answer that records the expected value, acceptable wording, source, unit, scope, period and any necessary qualification. Subject-matter owners, not the agent itself, should approve this reference set.
Testing should evaluate more than whether the expected number appears. Useful measures include:
- the share of answers with the correct value, unit, entity and period;
- the share supported by the designated authoritative source;
- unsupported answers for which no source justifies the claim;
- contextually wrong answers that contain a real figure but apply it incorrectly;
- failures to disclose stale, provisional or incomplete data;
- permission breaches;
- appropriate abstention when evidence is insufficient.
Results should also be reviewed by question type. A single overall score can hide serious weaknesses in high-consequence areas. Passing routine product lookups does not compensate for unreliable financial or compliance answers.
Define escalation before deployment
An agent needs explicit rules for uncertainty and conflict. It should escalate when authoritative sources disagree, required metadata is missing, a dataset is outside its freshness limit, the user requests an unauthorised combination or the answer could materially affect a person, payment, contract or regulatory obligation.
Human review should be mandatory for consequential decisions unless the organisation has separately established an appropriately governed automation process. The reviewer needs the proposed answer, supporting sources, relevant metadata and identified conflicts, not merely a request to approve an unexplained output.
Escalations should become inputs to improvement. Repeated confusion over the same metric may reveal an unresolved definition problem. Frequent stale-data warnings may point to weak update ownership. Unsupported answers may indicate that retrieval scope is too broad or that the agent is answering when it should abstain.
Set an evidence-based readiness gate
Before release, require business, data, security and process owners to approve a defined scope. A dataset is ready only when its authority, meaning, provenance, update controls and permissions are documented, and when the agent has passed representative retrieval tests against thresholds proportionate to the use case’s risks.
The UN and UNICEF context illustrates why this work cannot stop at accessibility. It does not establish which architecture businesses should adopt or prove that a specific platform will solve retrieval accuracy. The defensible approach is to begin with a limited set of important questions, establish approved answers, test the complete retrieval path and expand only when observed performance supports doing so.
For managers, the decisive question is not “Can the agent read our data?” It is “Can we demonstrate that it selects the right source, preserves the context and declines or escalates when a reliable answer is unavailable?”