A building business can call the same thing by four names. The estimator writes “owner-supplied plumbing.” The selection sheet says “client-provided fixtures.” Purchasing records “no PO.” The project manager types “by owner” in a schedule note. A person who knows the job can reconcile those phrases. An AI workflow may treat them as four unrelated conditions and produce a confident, incomplete answer.

OpenAI's August 2026 enterprise report describes the shift from asking AI questions to delegating work across sources and systems. That shift raises the cost of inconsistent language. When an agent only drafts a paragraph, a vocabulary mismatch is annoying. When it compares an estimate to a specification, prepares a purchase list, or flags a project exception, the same mismatch can change what work appears complete. The Datum interpretation is simple: before you automate a cross-system workflow, define the terms that must mean the same thing everywhere.

Find The Terms That Change Decisions

Do not start by cleaning every field in every system. Start with one decision the AI will support. If the workflow checks whether selections are ready for ordering, list the terms that determine readiness: approved, allowance, alternate, owner-supplied, field-verified, long-lead, discontinued, released, and on hold. If the workflow reviews change orders, focus on scope status, pricing status, authorization, schedule impact, and billing status.

Choose terms with operational consequences. A label belongs in the first vocabulary when a mismatch could change cost, scope, schedule, responsibility, compliance, customer communication, or the next authorized action. Cosmetic naming differences can wait.

Build A Small Translation Table

  • Canonical term: the approved name the workflow will use in outputs and status fields.
  • Definition: one plain-language sentence that tells a team member what the term includes and excludes.
  • Known aliases: the shorthand, legacy labels, vendor terms, and software values that mean the same thing.
  • Disallowed collisions: similar phrases that must not be merged because they carry different responsibility or risk.
  • System of record: where the authoritative current value lives.
  • Owner: the role allowed to change the definition or resolve an ambiguous match.

Keep the first table narrow enough to review in one meeting. Twenty well-governed terms tied to a real workflow are more useful than a 500-row glossary nobody owns. Store the table where people and software can both reach it, preserve revision history, and give every change an effective date.

Separate A Synonym From A Business Rule

“Client-provided” and “owner-supplied” may be synonyms. “Allowance” is not. An allowance describes a pricing condition; owner-supplied describes responsibility for procurement. A fixture can be both, one, or neither depending on the contract and project record. If an AI system collapses those terms because they often appear together, it can invent a responsibility that the visible source never assigned.

Write these distinctions as testable rules. For example: do not infer procurement responsibility from an allowance field; do not treat approved as released for purchase; do not treat delivered as installed; do not treat a vendor quote as a committed purchase order. The useful vocabulary is not just a list of words. It is a map of the decisions those words are allowed to support.

Learn From The Industry's Data Work

buildingSMART USA describes the US Data Dictionary as an effort to centralize and harmonize built-asset terminology and map it to classifications and schemas. buildingSMART International's data dictionary likewise exists to share consistent definitions for the built environment across software. A small contractor does not need to implement an international standard to learn from that architecture. The transferable lesson is that definitions, mappings, ownership, and machine access are infrastructure, not clerical cleanup.

Use recognized classifications when they genuinely match the workflow, especially when data must move between outside partners or software. Keep company-specific terms where they reflect real operating practice. Record the mapping rather than pretending every internal label is universal.

Make Ambiguity A Visible State

The agent should not silently choose the closest term when two meanings remain plausible. Give it an unresolved state and a review route. Show the source phrase, the candidate mappings, the affected record, why the distinction matters, and the person authorized to resolve it. Once reviewed, add the result to the translation table only if it should apply beyond that single project.

Log every automatic mapping with its source system, source value, canonical value, vocabulary version, and confidence or rule used. That audit trail makes it possible to trace a bad output back to the exact translation instead of arguing about whether the model misunderstood the whole project.

Test With Messy Historical Records

Replay completed jobs, not polished demo data. Include abbreviations, blank fields, renamed products, old cost codes, copied scope notes, vendor terminology, and project-specific exceptions. For each record, check whether the workflow chose the correct canonical term, preserved important distinctions, cited the authoritative source, and escalated genuine ambiguity.

Track false merges and missed matches separately. A missed alias may leave work out of a report. A false merge can be worse because it makes unlike conditions appear safely reconciled. Expand the vocabulary only after the workflow passes a representative set of real records and a responsible operator accepts the remaining exception rate.

Publish Definitions That Help Buyers Decide

Google says AI Overviews and AI Mode do not require special AI-only schema. Helpful, original content, clear internal links, accessible text, and structured data that matches the visible page remain the foundation. For a building-industry company, clearly explaining terms customers routinely confuse can serve both people and search systems—but the page should go beyond a generic definition. Show how your company applies the term, what decision it changes, what evidence controls it, and where responsibility shifts.

The Datum Rule

Do not connect an AI agent to three systems and hope the model reconciles their language. Give the workflow a small, owned vocabulary tied to one decision. Define aliases, preserve distinctions, expose ambiguity, and test the mappings against real jobs. Shared vocabulary is not glamorous, but it is what lets delegated work stay attached to the same operational truth your team uses.

Build The Grounding Layer

Sources Read

Next step, if this note maps to a problem on your desk: Book Private Training.

Related Field Notes