Workday Buys Honeydew to Stop Three AI Agents From Reporting Three Different Headcounts
Workday said on November 18 it intends to acquire Honeydew, a semantic layer that compiles business definitions into governed queries so agents and BI tools return the same number. No price was given.
Workday said on November 18 that it intends to acquire Honeydew, a small semantic layer company, to fix a problem that shows up the moment two AI agents answer the same finance question differently. The price was not disclosed.
The example Workday uses in its announcement from Barcelona is the one every finance team recognises. Ask a dashboard, a chat assistant and an agent for headcount and you can get three numbers: one counts full time employees, one includes contractors, one counts approved but unfilled roles as staffed. None of the three is wrong. They disagree because each tool carries its own definition.
Definitions compiled before the model sees the question
Honeydew's product turns business definitions into governed queries, so the model is handed the question and the correct path to the answer instead of a pile of table schemas and SQL rules pasted into every prompt. Workday plans to run that compiler against live data in Workday Data Cloud, with the customer's permissions attached to the definitions, and to block invalid or unauthorised requests before a query executes.
That is a governance argument, not a modelling one. A language model can write a plausible revenue query. It cannot know whether the organisation counts bookings gross or net unless someone wrote that down once and the system enforced it.
Context, not reasoning, is the bottleneck
The executives quoted in the release make the same point from different angles. Workday chief technology officer Gabe Monroy says a model should not be improvising the calculation behind a payroll number. Yasmeen Ahmad, who runs Workday Data Cloud, says the bottleneck in enterprise AI is context rather than reasoning. Honeydew chief executive David Krakov says getting models to write queries was never the hard part.
Honeydew has shipped a semantic layer built as a Snowflake native app, with Pizza Hut as a named customer connecting Power BI and Excel to governed transactional data, and Snowflake Ventures invested in the company in 2025. Workday is buying the layer and the compiler rather than building both from scratch, which is the faster route to a definition layer that already runs on someone else's warehouse.
The transaction is expected to close in the fourth quarter of Workday's fiscal 2027, ending January 31, 2027, subject to customary conditions. Until then this is an intention, not a shipped product, and Workday has not said how customers on other warehouses would be served.
Why this lands on finance teams first
Every AI finance project eventually reaches the same test: can the system produce a number that survives an audit. Agents that read invoices and draft forecasts are easy to demo and hard to trust when the underlying metric is defined in four places. Buying a semantic layer is a bet that the definition, not the model, decides whether the output is usable. Our guide to AI for financial analysis covers where that boundary usually sits in practice.
Some offers on this page may be paid placements or contain affiliate links.
Finpresso: daily AI & finance brief
Free daily newsletter, read in 5 minutes.
Subscribe free