Artificial intelligence in construction has evolved from assistants that answer questions into agents that take multi-step action across estimating, scheduling, safety compliance, subcontractor management, and more. Understanding the terms behind it has become a practical necessity, especially in heavy civil and infrastructure work, where concepts like DOT pay items, prevailing wage, and multi-tier union labor rarely appear in a typical software glossary but separate a tool built for your jobs from one that fits.
This glossary breaks down the AI terms that matter most to general contractors, project managers, and owners, explaining what each one means and how it applies to real construction workflows.
AI (Artificial Intelligence)
In heavy construction, artificial intelligence is the use of computer systems that learn from project data to speed up estimating and bidding, flag scheduling and cost issues, support safety compliance, and increasingly take on multi-step tasks like drafting reports or estimates, with a person still making the final call.
In Construction: AI shows up as models learning from years of your estimating and job cost history to recognize patterns a rules-based system would miss, like which soil conditions or crew combinations tend to affect production rates on paving jobs.
Example: A model that reviews years of closed-out bids to flag which line items on a new highway resurfacing job are priced lower than similar past work.
AI Agent
An AI system that can break a goal into steps, use tools to carry them out, and work through a task rather than just answering a question.
In Construction: An AI agent goes beyond answering a question; it drafts an estimate, flags a cost variance, or writes a daily field report, then hands it to a person to review before it's finalized.
Example: An estimating agent that pulls historical unit costs and company libraries to draft a first-pass estimate for a bridge deck replacement, leaving final pricing strategy to the estimator.
AI Model
A system that has learned patterns from data so it can make predictions, answer questions, or create content. Think of it as the part of AI that “knows” how to solve a problem based on what it has learned.
In Construction: The model is the part of an AI tool that has actually learned from your estimating, job cost, and equipment data, which determines whether a recommendation reflects your business or the generic industry average.
Example: A model trained on years of HeavyJob production data that predicts how many tons of asphalt a paving crew can realistically place per day on a given DOT spec.
Agentic AI
AI systems that coordinate multiple agents to complete broader, multi-step tasks, with a person reviewing or directing the outcome at key points rather than the system acting on its own.
In Construction: Agentic AI can automatically triage RFIs, assign resources, and escalate issues based on learned patterns and project context.
Example: An agentic system that reviews an incoming RFI on a highway project, matches it to a similar past RFI and its resolution, drafts a response, and routes it to the PM for sign-off instead of just logging it in a queue.
API (Application Programming Interface)
A tool that lets different software programs talk to each other and share information. For example, an AI agent can use an API to connect with another app or service.
In Construction: APIs let an estimating or job cost platform automatically pull data from telematics, accounting, or safety systems, instead of someone re-entering it by hand.
Example: An API connection between a GPS telematics system and a fleet maintenance platform that feeds utilization data straight into maintenance schedules.
Anomaly Detection
AI that identifies unusual patterns or behaviors that deviate from normal operations, often indicating potential safety hazards or compliance issues.
In Construction: Anomaly detection monitors jobsites for safety violations, equipment malfunctions, or unusual activity that could indicate problems requiring immediate attention.
Example: Identifying equipment operating outside normal parameters before failures occur; Spotting unusual work patterns that might indicate safety risks.
Automation
Technology that performs tasks with minimal human intervention, following predefined rules or AI-driven decision-making.
In Construction: Automation handles the repetitive parts of a job, such as re-keying time card data or routing a submittal, so crews and office staff spend less time on data entry and more on decisions that require judgment.
Example: Automatically syncing daily field time entries into payroll and job costing without a clerk manually transferring hours from paper time cards.
Bias (in Machine Learning)
Unintended skew or prejudice, or unrepresentative data used to train the model.
In Construction: Bias shows up when a model is trained mostly on generic commercial building data and then applied to heavy civil work, resulting in recommendations that don't account for DOT specs, prevailing wage, or union labor rules.
Example: A generic estimating tool that consistently underprices earthwork because it was trained mostly on vertical construction projects rather than horizontal projects such as roadway or utility projects.
Chain of Thought (CoT) Prompting
A technique where a complex problem is broken down into a series of smaller, sequential steps within the prompt to guide the LLM's reasoning process and improve accuracy.
In Construction: Chain of thought prompting helps an AI tool reason through a multi-step construction question, such as calculating a change order's cost impact, instead of jumping straight to a possibly wrong answer.
Example: Prompting an AI assistant to first identify the affected cost codes, then calculate the labor and equipment impact, and finally total the change order value, rather than asking for the total outright.
Chatbot
A user interface that allows human users to communicate with and access AI models and associated tools.
In Construction: A chatbot gives field and office staff a way to ask plain-language questions about a job, such as the current budget status, without running a report or opening multiple screens.
Example: A superintendent typing "which cost codes are over budget on the I-35 widening job?" into a chatbot instead of pulling a variance report manually.
Clustering Analysis
A method in Machine Learning that finds groups or patterns in data. For example, it can show which factors often appear together in safety incidents.
In Construction: Clustering analysis can group safety incidents, equipment failures, or cost overruns by shared characteristics, surfacing patterns that would be easy to miss when reviewing incident reports one at a time.
Example: Grouping months of near-miss reports to reveal that most incidents cluster around a specific piece of equipment or shift time.
Complex (Task)
A task involving many interactions between inputs, resources, and outputs, where changes in one area have implications for many others, requiring judgment and understanding of the broader context. Humans are needed for complex tasks.
In Construction: Deciding a bid's final pricing strategy is complex because labor availability, competitor behavior, risk tolerance, and the owner relationship all interact, which is why that decision stays with the estimator rather than the AI.
Example: Setting the markup and contingency on a competitive DOT bid, where AI can surface historical data, but a person still has to weigh the bigger picture.
Complicated (Task)
A task with many steps or options, but where those steps and options are clear, independent, and do not interact in complex ways. AI is generally good at complicated tasks.
In Construction: Complicated but well-defined tasks, like quantity takeoffs from a spec sheet or reconciling time cards against a schedule of values, are where AI tends to save the most time because the steps are repeatable.
Example: Extracting pay item quantities from a DOT plan set, a task with many line items but a consistent, well-defined process.
Compliance Monitoring
Automated systems that continuously check operations against safety regulations, building codes, and company policies.
In Construction: Compliance monitoring flags jobsite conditions or paperwork that fall outside safety regulations, prevailing wage rules, or company policy before they become a citation or audit finding.
Example: A system that flags a crew working near a trench without a documented shoring inspection and matches it against OSHA excavation requirements.
Computer Vision (Machine Vision)
AI technology that enables computers to interpret and analyze visual information from images, videos, or real-time camera feeds.
In Construction: Computer vision lets a camera on a jobsite or on a piece of equipment recognize issues such as missing PPE, an unsafe swing radius, or completed work against schedule, without a person watching the footage.
Example: A jobsite camera system that flags a worker without a hard hat near an active excavation.
Context Engineering
The craft of carefully designing an AI model's inputs (prompts, retrieved information, and structure) to shape what it "knows" during a conversation or task, prioritizing relevance and placement of critical information.
In Construction: Context engineering determines whether an AI assistant answering a project question actually knows your cost codes, labor agreements, and job history, or is just guessing based on general construction knowledge.
Example: Structuring a prompt for an estimating assistant so it references your company's specific labor rate tables and equipment cost library rather than national averages.
Context Window
The amount of information (measured in tokens) that an LLM can handle and process simultaneously in a single interaction.
In Construction: The context window determines how much of a spec book, estimate, or daily log an AI tool can actually consider at once, which matters when you're asking it to reason across an entire bid package rather than one page.
Example: A context window large enough to hold an entire DOT special provisions document so the AI can answer questions spanning multiple sections without losing track partway through.
Custom Chatbots (e.g., Custom GPTs, Gems, Artifacts)
User-creatable versions of chatbots that allow saving work, organizing past chats, uploading documents, and setting project-specific or user-specific instructions (system prompts).
In Construction: A custom chatbot can be set up with a company's own specs, safety manual, or estimating standards so answers reflect how your company actually operates, not generic construction knowledge.
Example: A custom chatbot loaded with a company's safety manual, so foremen can ask specific policy questions and get answers sourced from that manual instead of a general web search.
Data Mining
AI techniques that discover patterns and insights from large datasets that would be impossible to identify through manual analysis.
In Construction: Data mining digs through years of closed-out jobs, change orders, and cost records to surface patterns, like which project types or owners tend to generate the most disputes, that would take a person weeks to find manually.
Example: Mining five years of job cost history to identify which subcontractor trades most consistently run over budget.
Data Wrangling
The process of organizing, cleaning, and labeling raw data to make it suitable for training AI models, especially deep learning models, often consuming 90% of the project effort.
In Construction: Data wrangling is the unglamorous work of cleaning up inconsistent cost codes, mismatched units, and duplicate records across estimating, job cost, and fleet systems so an AI tool has something reliable to learn from.
Example: Standardizing cost code names across estimating and job cost systems so that a variance report compares the same categories rather than near-matches that don't line up.
Deep Learning
A type of Machine Learning that uses neural networks organized in multiple layers to understand complicated real-world data, particularly in language and vision applications.
In Construction: Deep learning powers the more advanced construction AI features, like reading a scanned plan set or transcribing a foreman's spoken daily report, tasks too visually or linguistically complex for simpler statistical models.
Example: A deep learning model that transcribes and structures a foreman's verbal end-of-day report into a formatted daily log.
Generative AI
A type of Deep Learning that is pre-trained on vast datasets (like internet content) and is capable of understanding context to create new content, such as text, images, or video, by predicting the highest likelihood content for a "gap."
In Construction: Generative AI drafts a first-pass estimate narrative, daily report summary, or safety toolbox talk from a short prompt, giving a person a starting point instead of a blank page.
Example: Generating a draft scope-of-work summary from a set of plan notes and quantities for an estimator to review and refine.
Hallucinations
The phenomenon where a Large Language Model (LLM) fabricates information or gives a confidently incorrect answer because it is designed to provide an answer even when it does not know the true information.
In Construction: A hallucination is an AI tool stating a cost, code requirement, or quantity with total confidence when it's actually wrong, which is exactly the risk in a bid or invoice where a plausible-sounding error can slip through unnoticed.
Example: An AI assistant confidently cites a prevailing wage rate that doesn't match the actual DOT wage determination for that county.
Historical Data Analysis
AI examination of past project performance to identify patterns that inform future estimates and decisions.
In Construction: Historical data analysis is how an AI tool learns what a realistic production rate, unit cost, or crew size looks like on your jobs, instead of relying on industry-wide averages that don't reflect your equipment or labor.
Example: Analyzing three years of closed paving jobs to calibrate a more accurate tons-per-day production rate for a specific paving crew.
Incident Prediction
AI models that analyze job site conditions, worker behavior patterns, and historical incident data to predict where safety problems are most likely to occur.
In Construction: Incident prediction flags the jobsite conditions or crew patterns most associated with past injuries, giving safety managers a chance to intervene before an incident happens instead of after it's reported.
Example: Flagging that a crew working overtime hours in high heat has historically correlated with a higher rate of near-miss reports.
Judgment
The human ability to use experience, intuition, and context to make decisions in new or uncertain situations, which AI lacks.
In Construction: Judgment is the part of every AI-assisted estimate or report that still requires a person, such as deciding how aggressively to price a bid against a specific competitor or how to read a client's true intent behind a change order request.
Example: An estimator overrides an AI-drafted price for a line item because they have a specific supplier relationship the model doesn't know about.
Large Language Model (LLM)
A type of Generative AI model trained on massive amounts of text data from the internet and other sources, primarily used for language-based applications like chatbots.
In Construction: An LLM is the technology behind most construction chatbots and assistants that let you ask questions in plain English about a job, a spec, or a policy and get a written answer back.
Example: An LLM-powered assistant that answers "what's the retainage rate on this contract?" by reading the actual contract language rather than giving a generic definition.
Linear Regression
A statistical technique used in Machine Learning to find a predictive relationship between two or more variables, often resulting in an equation to predict outcomes.
In Construction: Linear regression can model something like the relationship between crew size and daily production, giving a simple, explainable way to predict output as conditions change.
Example: Using regression on past jobs to predict how added equipment hours affect earthwork production rates on a given soil type.
Machine Learning (ML)
A set of statistical techniques that enable systems to understand and use large quantities of numbers (data) to create predictive abilities, forming the foundation of all AI.
In Construction: Machine learning is the layer underneath most "smart" construction features, like flagged budget variances or suggested unit prices, that improves as more of your job data feeds into it over time.
Example: A machine learning model that gets better at predicting equipment maintenance needs the longer it has access to a fleet's telematics history.
Masking
A training technique for Generative AI where parts of the content are hidden ("masked"), and the model learns to predict the missing information, enabling it to fill in blanks or complete responses.
In Construction: Masking is part of how models get trained to fill in gaps, like predicting a missing quantity or cost code from surrounding context, a skill that shows up when an AI tool suggests how to complete a partially filled estimate line.
Example: A model trained by masking cost values in historical estimates and learning to predict them from the surrounding scope description.
Model Context Protocol (MCP)
An open standard for connecting AI assistants to outside data and tools, so one connector works across many AI products.
In Construction: MCP is what lets an AI assistant access your actual job cost, scheduling, or fleet systems mid-conversation, rather than only working from what was typed in the chat.
Example: An AI assistant using MCP to pull live budget-to-actual numbers directly from a job costing platform while answering a question in a chat window.
Natural Language Processing (NLP)
The field of AI focused on enabling computers to understand, interpret, and generate human language.
In Construction: NLP lets a foreman describe their day out loud, or a PM type a question in plain English, and have the system actually understand and act on it, rather than requiring rigid commands or dropdown menus.
Example: NLP turning a foreman's spoken description of the day's work into a structured daily report with quantities and crew hours.
Pattern Matching
The fundamental mechanism of Large Language Models (LLM), where they match user prompts and given context against patterns in their vast training data to generate responses.
In Construction: Pattern matching is why an LLM can sound knowledgeable about a construction topic without necessarily having verified facts about your specific job. It's recognizing familiar patterns in language, not checking your numbers.
Example: An LLM matching the phrasing of a question about DOT pay items to similar patterns seen in training, even if it has never seen your actual contract's pay item list.
Perceptrons
Early forms of artificial neurons, foundational to the development of neural networks in AI.
In Construction: Perceptrons themselves rarely come up directly in construction software, but they're the historical building block behind the neural networks powering today's computer vision and language features on the jobsite.
Example: The neural network behind a jobsite safety camera's object detection traces its lineage back to early perceptron models from the 1950s and 60s.
Predictive Analytics
AI techniques that analyze historical data to forecast future outcomes and identify potential risks before they occur.
In Construction: Predictive analytics uses past job performance to forecast things like which active jobs are trending toward a budget overrun, giving PMs a warning while there's still time to act.
Example: Forecasting that a bridge job's current concrete burn rate puts it on track to exceed budget by project completion if current trends continue.
Probability
A measure of how likely an outcome is. In AI, it helps the system decide which answer, prediction, or action is most likely correct based on the data it has learned from.
In Construction: Probability is behind a confidence score on an AI recommendation, telling you how sure the system is that a flagged variance or suggested price is accurate rather than a guess.
Example: An AI tool surfacing a cost estimate along with a stated 80% confidence level based on how closely the job matches historical patterns.
Prompt
The input text or query given to a Large Language Model by a user, which guides the LLM's pattern-matching process to generate a desired response.
In Construction: The prompt is what you actually type or say to an AI assistant, and how it's worded, vague or specific, with or without context, has a real effect on whether the answer is useful for your job.
Example: Asking "summarize today's field activity for the Route 9 job, including any safety incidents and equipment downtime" instead of just "summarize today."
Prompt Injection
A cybersecurity risk where a malicious prompt is inserted into a multi-turn conversation with an LLM to override or reverse its initial system prompt instructions.
In Construction: Prompt injection matters most for tools connected to real project data or systems, where a manipulated input could try to trick an AI assistant into revealing or altering information it shouldn't.
Example: A malicious message embedded in an uploaded document attempting to instruct an AI assistant to ignore its permissions and pull unrelated project financials.
Real-time Analytics
AI systems that process and analyze data as it's generated, providing immediate insights for decision-making.
In Construction: Real-time analytics turns field data, like time cards, GPS pings, or safety check-ins, into a current picture of the job as the day unfolds, rather than a report pulled together after the fact.
Example: A dashboard showing live labor and equipment costs against budget as crews clock hours throughout the day.
Retrieval Augmented Generation (RAG)
A technique that enhances LLM accuracy by allowing the model to retrieve information from a specific dataset (e.g., uploaded documents, external databases) before generating a response, thereby reducing hallucinations.
In Construction: RAG lets an AI assistant answer a question about your specific contract, spec, or job history accurately by first pulling the relevant document, rather than guessing from general training data.
Example: An assistant that retrieves a project's actual special provisions document before answering a question about a specific pay item, rather than relying on general DOT knowledge.
Symbolic Logic
An early approach to AI (1950s) that focused on representing knowledge and reasoning using logical symbols and rules.
In Construction: Symbolic logic underlies older rule-based software features, such as automated checks that flag a bid missing a required item, which follow explicit if-then rules rather than learning from data.
Example: A hard-coded rule flagging an estimate as incomplete if a required bond or insurance line item is missing, regardless of any learned pattern.
System Prompt
An invisible, pre-set set of instructions and rules embedded by the model maker or software vendor into an AI product, guiding the LLM's behavior and often overriding user prompts.
In Construction: The system prompt is what keeps a vendor's AI assistant consistent and within its intended guardrails, such as staying focused on job cost questions and declining to give legal advice, no matter how a user phrases their question.
Example: A vendor's system prompt instructs an assistant to always cite the source job or document behind any cost figure it provides.
Temperature
An internal setting in an LLM that controls its level of "creativity" or randomness in its responses. A higher temperature generally means the AI will generate more varied and potentially more creative outputs.
In Construction: Temperature affects how much variation you get from an AI writing tool. A low setting keeps a daily report summary close to the facts submitted, while a high setting can produce more creative but less predictable phrasing.
Example: Keeping an AI-generated safety incident summary at a low temperature setting so it sticks closely to the reported facts rather than embellishing details.
Tokens
Chunks of text, roughly 4 characters on average, that Large Language Models break language into for processing; a token can be a whole word, part of a word, or a punctuation mark. LLMs work with tokens, not directly with human words.
In Construction: Tokens are the unit an AI tool actually processes text in, which matters practically because a long spec book or lengthy daily log can use up a lot of tokens, affecting how much the AI can consider at once.
Example: A 200-page bid document breaking down into tens of thousands of tokens, which affects how much of it an AI assistant can review in a single pass.
Triggers
In artificial intelligence, a trigger is an event, condition, or signal that wakes an AI system, workflow, or agent and prompts it to act. It serves as the starting point that sets the AI into motion.
In Construction: A trigger is the specific event that sets an AI agent into motion, like a submitted daily report, a missed inspection deadline, or a cost code crossing a budget threshold.
Example: A budget variance crossing 10% over on a cost code triggers an AI agent to flag it and notify the PM automatically.
User System Prompt (Customization Prompt)
A customizable version of a system prompt that users can create in chatbots to set personal behavior guidelines, tone, length, or role for the AI.
In Construction: A user system prompt lets an estimator or PM set standing instructions for their AI assistant, such as always responding in a specific report format or always flagging DOT-specific pay items, so they don't have to repeat those instructions each time.
Example: A PM setting a custom instruction so their AI assistant always formats daily report summaries with separate sections for labor, equipment, and safety notes.
