En
← All articles

A correct metric can still be an answer an AI should not give

Semantic definitions settle the calculation, not permission. Enterprise AI needs identity, time, cache and action boundaries tested together.

ColumnAI FrontiersAuthorOpen Market NotesTypeArticle
Branching data lines pass through a gate with a red keyhole

A manager asks: “What is the average bonus per person in my department this quarter?” A system can use the right formula, retrieve current records and produce a perfectly reproducible number, yet still be wrong to answer. The manager may have just moved to another department. Or the filtered population may contain only one person. Numerical correctness and authorized disclosure are different properties.

That distinction is the useful lens on Workday’s proposed purchase of Honeydew. Its November 18 announcement describes an intent to acquire, with conditional closing expected in the fiscal quarter ending January 31, 2027. It is not evidence of a completed integration. The important question is whether business meaning and query permissions can both become executable rules, rather than whether another AI company joins the portfolio.

A measuring rule is not an access pass

Think of a semantic layer as an executable contract for a metric: its numerator and denominator, relevant business period, permitted relationships between records and definition version. An average bonus might use quarter-end headcount or average full-time-equivalent staffing. Those measures serve different purposes. A model should not silently substitute one for the other while composing an answer.

Consistent meaning does not require every situation to produce one number. It requires explicit names, owners and conditions for different definitions. If “this quarter” could mean a fiscal or calendar quarter, the system should clarify. Stable SQL that silently selects the wrong interpretation is still wrong. Determinism makes calculations repeatable; it does not establish that the chosen inputs express the right business intent.

Authorization asks different questions: who is asking, whom the agent represents, which rows and fields are allowed, for what purpose, and whether access remains valid now. Moving the formula into a compiler does not automatically answer them. If every request ultimately uses one broad service account, different names in the chat interface do not mean the database recognizes different access scopes.

Three counterexamples beat a polished demonstration

The following are architecture tests, not allegations of incidents at Workday or any other vendor.

First, a manager has access in the morning, moves departments in the afternoon, and the assistant still serves a cached bonus result from the morning. The formula and original query may both have been correct. The failure is whether the cached result becomes unavailable when permission changes. A cache key needs to account for access scope and policy version; matching only the question text can turn a once-authorized query into a redistributable data copy.

Second, a user can see a department total and a total excluding one employee. Subtracting two individually permitted aggregates may reveal that employee’s bonus. Row filters and field masking do not automatically prevent inference across results. Sensitive use cases may need minimum group sizes, constraints on query combinations or human review. Refusing an answer can be more faithful to the real policy than returning a nominally permitted field.

Third, a live query supplies today’s organizational structure to explain last quarter’s cost per employee. Fresh is not necessarily appropriate. Historical analysis needs the assignments effective at the time, the metric version and a data cutoff. For a closed-period report, a reproducible approved snapshot may be more useful than a constantly changing live figure.

These failures concern authorization lifetime, combinations of outputs and business time. A model that writes better SQL does not directly repair any of them.

There is more than one implementation path

Workday proposes tighter native integration of semantics with its business permissions. Unreleased capabilities should still be treated as plans, not present guarantees. Another path enforces policy within an existing data platform. Snowflake’s documentation describes row-access policies determining returned rows at query runtime. Query-time governance is therefore not a single-vendor technique, and putting data in a warehouse does not inherently remove governance.

A further route combines shared metrics with explicit credential mapping. dbt’s documentation links service tokens to underlying data-platform credentials. The procurement question is whose access scope ultimately executes, not simply whether the product has a semantic layer. Team-scoped credentials and user-level identities each bring operating costs. A unified entry point does not make those choices disappear.

For a small set of high-risk tasks, preapproved query templates may be preferable. They sacrifice flexibility but narrow the executable space and simplify audit. A warehouse can remain practical for cross-system business analysis, provided the organization accepts the costs of refresh lag, permission synchronization and historical versioning. A native live-data approach must demonstrate how identities and business definitions connect across systems, not merely show one successful query inside a single application.

Acceptance testing should follow an evidence chain

A procurement test should go beyond “the answer matches the dashboard.” Run the same question under different identities, before and after a role change, at different cutoff dates and with cache hits. Record the metric version selected, authorization principal, data scope, policy decision and result timestamp. A refusal also needs an auditable reason that does not itself disclose sensitive information.

Then test actions separately. Seeing a bonus budget does not authorize changing bonuses; understanding accounts payable does not authorize making a payment. Read access, write access and approval responsibility need distinct checks. A semantic contract can help an agent understand a number without granting authority over the resources that number represents.

The commercial value to watch is therefore not just a shorter prompt. It is whether less duplicated metric maintenance can coexist with reliable boundaries through permission changes, historical reproduction and calls from multiple tools. Workday’s announcement supplies a direction. Delivery evidence must show that the system knows not only how to calculate, but why this person may receive this answer now.

Information only. No investment, legal, tax, or financial advice.