AI Terms Construction Owners Should Know

The complete glossary of AI terms every construction owner should know is a controlled vocabulary covering artificial intelligence, data environments, BIM, project controls, risk and asset information. It explains what each term means in an owner’s design, construction and operations workflow, what it does not mean, and how to test a vendor’s claim. The practical purpose is not to memorise technology language. It is to define the data, permissions, calculations and human approvals that make AI useful and governable across a capital programme.

Why This Glossary Exists

Construction terminology is shifting because AI is being placed alongside systems that already manage drawings, specifications, schedules, cost, contracts, quality, safety and asset information. Construction was ranked second-lowest among industries on McKinsey Global Institute’s Industry Digitization Index, based on 2016 data cited in Reinventing construction: A route to higher productivity (February 2017). At the same time, 73% of surveyed UK firms reported using BIM on at least some projects in the NBS Digital Construction Report 2024 (April 2024), compared with 13% in 2011.

That combination creates a terminology problem. A project team may use “CDE” to mean an ISO 19650 information-management environment, while a vendor uses it to describe a document repository. “AI agent” may mean a multi-step tool-using workflow in one proposal and a chat interface in another. “Digital twin” may refer to a dynamically updated asset representation or simply a 3D model.

These distinctions affect procurement, reporting and accountability. ISO 19650 emphasises consistent information requirements, naming and information states. NIST’s AI Risk Management Framework 1.0 (January 2023) identifies data quality and well-defined context as important to reliable AI. If “risk”, “issue”, “defect” and “change event” are used interchangeably, an AI model cannot reliably classify records or produce comparable portfolio reporting.

A glossary therefore belongs in an owner’s information requirements, RFP, data model and AI governance policy. It should define the entities, statuses, calculations and actions that apply across designers, general contractors, consultants, PMCs and internal teams. It should also identify which outputs are advisory and which require a named human approver.

Foundational Terms

Project Management Information System (PMIS)

Definition: A PMIS is a system used to collect, integrate and disseminate project information such as documents, schedules, cost data and communications.

A PMIS supports project management processes; PMI discusses the concept in the PMBOK® Guide – Seventh Edition (2021). It is not automatically a portfolio management system, an enterprise resource planning system or an asset-lifecycle platform. In practice, an owner may use a PMIS to consolidate project controls and approvals while separate BIM, finance or facilities systems remain authoritative for particular records.

For an owner, the evaluation question is: which information is held, who owns it, and how is it transferred into portfolio and asset reporting? AI can only answer questions consistently when the PMIS exposes structured project, contract, cost, schedule and document relationships.

Common Data Environment (CDE)

Definition: ISO 19650-1:2018 defines a CDE as a “single source of information used to collect, manage and disseminate all relevant approved project documents for multidisciplinary teams”.

A CDE is more than a folder or document-management label. ISO 19650 information management includes requirements for information containers and states such as work in progress, shared, published and archived. It is not the same thing as BIM, which describes information-rich digital representations and processes, nor does the term alone prove that a platform manages cost, procurement, risk or assets.

In practice, an owner may connect a BIM authoring environment, a BIM collaboration environment, a project management platform and an owner-side capital project system. Ask where the source of record sits, how revisions are controlled, and whether AI retrieves approved information with its status and references intact.

Building Information Modelling (BIM)

Definition: BIM is an information-management approach using structured digital representations of built assets across their lifecycle, guided by the ISO 19650 series and related open standards.

BIM is not simply a three-dimensional model. The information may support design coordination, construction planning, quantities, handover and operations. buildingSMART’s IFC standard provides an open data format for exchanging built-asset information; BCF and IDS address other collaboration and information-requirement needs.

The owner’s concern is whether model information can be connected to requirements, approvals, cost, schedule, quality and handover data. AI may use BIM objects and their relationships as context for clash, constructability or information checks, but a model is not automatically a complete project record.

Earned Value Management (EVM)

Definition: EVM is a project-management technique that measures performance and progress by integrating scope, time and cost data.

PMI’s Practice Standard for Earned Value Management (second edition, 2019) defines the principal measures as Planned Value (PV), Earned Value (EV) and Actual Cost (AC). Cost Variance is CV = EV − AC; Schedule Variance is SV = EV − PV; Cost Performance Index is CPI = EV ÷ AC; and Schedule Performance Index is SPI = EV ÷ PV. Estimate at Completion (EAC) is the forecast total cost at completion, while Estimate to Complete (ETC) is the forecast cost required to finish the remaining work.

EVM is not a prediction model. It is a structured measurement method. Predictive analytics can use EVM trends as inputs, but an owner should verify the baseline, progress rules, accrual treatment and forecast formula before comparing projects.

Critical Path Method (CPM)

Definition: CPM identifies the longest path of dependent activities and calculates the earliest and latest start and finish dates that avoid delaying the project.

AACE International Recommended Practice 52R-06, updated in 2018, addresses time planning and schedule development. CPM is not the same as a progress dashboard or a list of late activities. A late activity matters differently depending on its logic, float and relationship to completion milestones.

For owners, AI-assisted schedule analysis should expose the activities, dependencies, logic changes and assumptions behind a forecast. A probability of delay without that evidence is not a substitute for schedule review.

Digital Twin

Definition: A digital twin is a digital representation of a physical asset, system or process that is dynamically updated from real-time data and uses simulation, machine learning and reasoning to support decisions.

This definition comes from the Centre for Digital Built Britain and the UK National Digital Twin programme (2020). A digital twin is not merely a BIM model or a static as-built database. BIM can provide important structured asset information, while a twin adds continuing updates and decision-oriented analysis across the asset lifecycle.

For an owner, the test is operational: what physical or process data updates the representation, how frequently, which decisions does it support, and who validates the information? Without defined sources and update requirements, “digital twin” remains too broad for an RFP or handover obligation.

AI & Automation Terms

Artificial Intelligence and Machine Learning

Definition: OECD defines an AI system as a machine-based system that infers from received inputs how to generate predictions, content, recommendations or decisions that can influence physical or virtual environments.

Machine learning (ML) is a subset of AI that uses statistical techniques to enable systems to learn from data rather than relying solely on explicitly programmed instructions, as described by ISO/IEC 22989:2022. ML may support predictions of delay, cost overrun, safety incidents, quality defects or equipment failure when suitable historical and current data exists.

AI is not synonymous with automation. A rules engine follows predetermined if-then logic; robotic process automation repeats defined actions; ML estimates outcomes from data. The owner should ask which of these is being used, what data was used for training or calibration, and how performance is measured after deployment.

Generative AI

Definition: Generative AI produces new content such as text, images or code from patterns learned during training.

NIST’s AI RMF 1.0 (January 2023) includes generative AI within its AI risk discussions. In construction, documented use cases include drafting RFIs, summarising site reports, suggesting contract-clause language, analysing documents and producing design alternatives when connected to BIM information.

Generative AI is not automatically factually correct, contractually authorised or grounded in the project record. An owner should require source references, confidence or uncertainty indicators where appropriate, access controls and human sign-off for consequential outputs.

Large Language Model (LLM)

Definition: An LLM is a generative AI model trained on very large text corpora to predict the next token in a sequence, enabling language understanding and generation.

LLMs can support semantic search across project documents, contract extraction, RFI and meeting-minute summaries, and conversational questions about project information. An LLM is not the same as a CDE, a project database or a verified calculation engine. It generates language; it does not, by itself, establish that an answer is supported by the latest approved drawing.

In an owner workflow, the useful distinction is between a general language capability and a grounded application. Ask whether the model can identify the source document, revision, clause or data record behind an answer and whether users can correct or reject the output.

AI Agent and Agentic Workflow

Definition: In current enterprise practice, an AI agent is an autonomous or semi-autonomous software entity that perceives information, reasons about a goal and takes actions through tools or APIs.

An agentic workflow is a multi-step chain in which an LLM or similar model calls tools sequentially to complete a task. Microsoft’s AutoGen paper (2023) and OpenAI’s function-calling documentation (2023–2024) describe relevant approaches. There is no single formal standard definition for “agentic workflow”.

An agent is not necessarily a chatbot. A chatbot may answer one question; a rules engine may route a form; a workflow may send a notification. An agentic process might detect a contract risk, create a risk-register item, assign it to a package owner and request review. For an owner, the critical controls are action permissions, audit logs, tool scope, escalation rules and mandatory approval before a consequential action.

Predictive Analytics

Definition: Predictive analytics combines historical and real-time data with statistical models or ML to estimate future outcomes such as delay, cost overrun, safety incident, quality defect, equipment failure or claims risk.

Predictive analytics is not certainty and does not eliminate professional judgement. McKinsey reported that AI-based predictive analytics could reduce project delays and cost overruns by up to 10–20% in some adoption cases in its June 2018 article; that is case-based and should not be treated as a universal result.

Deloitte identified schedule optimisation, cost estimation, design clash detection, quality and safety analytics, predictive maintenance and document or contract analytics among leading E&C AI use cases in 2023. Owners should ask what outcome is predicted, over what timeframe, against which baseline, with what precision, and how often the model is refreshed.

Computer Vision, Natural Language Processing and RPA

Computer vision applies AI to images or video, such as site imagery for quality or safety analysis. Natural language processing (NLP) enables software to classify, extract, search or generate human language, including contract and RFI content. Robotic process automation (RPA) executes predefined, repeatable actions across software interfaces.

These terms describe techniques, not proof of project value. A vendor should explain the input data, confidence thresholds, exception path and accountable reviewer. Image recognition that flags a possible issue is different from an approved quality record; document extraction that identifies a clause is different from a legal decision.

Project Controls Terms

Baseline, variance, forecast, EAC and ETC

Definition: A baseline is the approved reference plan against which performance is measured; variance is the difference between actual or current status and that reference; a forecast is an estimate of future performance; EAC forecasts total cost at completion; and ETC forecasts the remaining cost to finish.

These terms are not interchangeable. A variance describes what has happened relative to a baseline. A forecast describes what may happen next. AI models need both the original baseline and the version history of approved changes; otherwise a changed target can be mistaken for performance improvement or deterioration.

Risk, issue, assumption and dependency

Definition: A risk is an uncertain event or condition that may affect objectives; an issue is a present problem requiring management; an assumption is information accepted as true for planning; and a dependency is a relationship in which one activity, decision or deliverable relies on another.

A risk register records identified risks, likelihood, impact, owner, mitigation and status. This aligns with ISO 31000:2018 and PMI’s PMBOK® Guide – Seventh Edition (2021). Risk exposure is commonly expressed as probability multiplied by impact, although the scoring method must be defined by the owner.

Mixing these categories damages reporting and AI classification. A possible authority approval delay belongs in a risk workflow before it occurs; an overdue approval is an issue. An owner should define statuses such as open, closed and deferred, along with who may change them.

Change event, change order, variation, contingency and allowance

Definition: A change event is a potential alteration requiring assessment; a change order or variation is an authorised change to scope, price or time; contingency is an allocated reserve for defined uncertainty; and an allowance is a budget provision for work or cost that is not yet fully defined.

Contract terminology varies by jurisdiction and contract form, so the owner’s definitions should be written into the project procedures and commercial documents. AI can identify related correspondence, clauses and cost or schedule effects, but it should not convert a detected possibility into an authorised variation without the required review and approval.

Portfolio, programme and project

Definition: A project is a defined temporary delivery effort; a programme coordinates related projects or workstreams; and a portfolio groups investments for governance and strategic decision-making.

The distinction matters because an owner may need project-level RFI detail, programme-level interface risk and portfolio-level CapEx or forecast reporting. A model trained only on one project cannot automatically answer a portfolio question. The data model must preserve relationships between assets, programmes, projects, contracts, packages and controls.

How to Use This Glossary When Evaluating Vendors

Use the glossary as a procurement and governance instrument, not only as onboarding material. Put the definitions into the RFP, information requirements schedule or contract annex. Require bidders to map their fields and statuses to the owner’s terms and identify where a term has a different meaning.

Ask vendors these questions:

• Which capabilities use deterministic rules, ML, an LLM or an agentic workflow?
• What project, portfolio and asset data does each capability use, and what is the source of record?
• Can the system show data lineage from the source record to the dashboard, prediction or generated text?
• How are model versions, training or calibration data, known limitations and performance documented?
• What confidence, exception and approval controls apply before an action affects cost, time, contract administration or payment?
• Can the platform preserve document revision, information status and citations when answering a project question?

Do not accept “AI-native” as a sufficient technical description. A meaningful evaluation should connect the claim to a data model, a workflow and an accountable role. For example, ask a vendor to demonstrate how an RFI is retrieved against the approved drawing and specification, how a possible risk is recorded, how a prediction is evaluated, and where the project manager signs off.

For interoperability, test the mapping between BIM, CDE, PMIS, ERP and asset systems. ISO 19650, buildingSMART open standards and the owner’s own information requirements provide useful reference points. For governance, NIST AI RMF and the OECD AI principles provide a basis for discussing reliability, transparency, accountability, privacy and fairness.

Track the glossary itself. Useful measures include the percentage of active projects using the standard definitions; the percentage of key systems mapped to them; the percentage of records correctly classified; duplicate or ambiguous category rates; time to answer an owner question; manual reporting effort; and the precision or accuracy of predictions after standardisation. Connect those measures to project controls: CPI, SPI, EAC, schedule adherence, change-order value, high-impact risk exposure, mitigation-plan coverage, rework, defects and TRIR.

A controlled vocabulary can also be implemented in an AI-enabled owner platform. Zepth is one example of an AI-native platform built around a common data environment and structured project, procurement, asset and financial information, with human approval retained for consequential actions. The architectural principle is more important than the product name: AI should work against governed project context rather than operate as an untraceable layer beside it.

FAQ

Use this glossary as the starting point for an RFP annex, information-management schedule or internal AI policy. Subscribe to Zepth Insights, download the framework and checklist, or book a walkthrough to see how the definitions can be applied to governed capital-project workflows.