AI glossary for surveyors

Plain-English definitions of the AI and compliance terms the RICS AI standard uses — written for surveying practices, not data scientists. 27 terms.

AI system

Any software that performs tasks associated with intelligent behaviour — from large language models to measurement OCR to features embedded inside estimating or BIM platforms. The RICS standard covers embedded AI explicitly: how a capability arrived matters less than what it does to your service.

AI systems register

The written register the RICS standard requires for AI systems with material impact: the system, its purpose, the date first used and the next review date. See the full requirements and the free template.

Appropriateness assessment

The written assessment of whether AI is the most appropriate tool for a task, considering alternatives, data risks, error and bias risk, and environmental and stakeholder impacts. A periodically reviewed written policy or standing statement can serve as this assessment.

Audit trail

A chronological, tamper-resistant record of decisions and actions. For AI compliance it is what turns "we're careful" into evidence — who decided what, when, and why, in a form that survives scrutiny.

Baseline knowledge

Section 2 of the RICS standard: anyone using AI on surveying services must understand, at minimum, the types of AI and how they work, their limitations and failure modes, the risks of erroneous output and bias, and data risks.

Bias

Systematic skew in an AI system's outputs, arising from training data, algorithm design or the context of use. One of the four overarching risks the AI risk register must document.

Dip sampling

Randomised checking of a proportion of outputs at regular intervals, required by the standard where AI produces automated or high-volume output — per-output review being disproportionate, but accountability remaining with the firm. See outputs and reliance.

Due diligence (AI)

The written pre-procurement process RICS 4.1 requires for materially-impacting AI: six minimum written information requests to the supplier, recorded answers, recorded fitness-for-purpose testing, and gaps carried to the risk register. See the six questions.

Embedded AI

AI functionality inside software adopted for another purpose — transcription in a meetings platform, suggestion engines in estimating tools. In scope for the standard, and the easiest kind to miss because nobody installed it deliberately.

Erroneous output

Output that is simply wrong — a hallucinated citation, a mis-measured quantity, a mistranscribed figure. An inherent AI risk the standard requires firms to manage and document rather than assume away.

Explainability

The ability to provide, on request, written information about the AI systems used: type, basic ways of working and limitations, due diligence done, risk management, and reliability decisions made (RICS 4.4). See the five items.

Generative AI

AI that produces new content — text, images, code — rather than classifying or measuring existing material. Drafting assistants and chatbots are generative; measurement OCR generally is not.

Hallucination

A generative AI failure mode where the system produces confident, plausible, false output — invented case law, non-existent standards clauses, fabricated figures. The reason human review of AI-drafted professional content is non-negotiable.

Large language model (LLM)

The AI architecture behind tools like ChatGPT and Copilot: a model trained on very large text corpora to predict likely continuations. Strong at drafting and summarising; structurally prone to hallucination; understanding this trade-off is part of baseline knowledge.

Machine learning

The family of techniques where software learns patterns from data rather than following hand-written rules. Most modern AI, including LLMs and measurement models, is machine learning.

Material impact

The RICS standard's threshold test: whether an AI use materially affects delivery of the surveying service. Material use triggers the register, due diligence, client notices and output reviews. The firm makes the call — and must record the determination and reasoning. See how to decide.

Model

The trained artefact inside an AI system — the thing that actually maps inputs to outputs. Vendors swap and update models inside the same product name, which is why register review dates and repeat due diligence matter.

OCR (optical character recognition)

Technology that extracts text and measurements from images and drawings. Modern OCR is AI-driven, and in QS work its output often feeds cost plans directly — making it a frequent 'material impact' candidate.

Professional scepticism

One of the four components of professional judgement the standard names for output reliability decisions (with knowledge, skills and experience): treating AI output as a junior's work to be checked, not an oracle's answer.

Prompt

The input given to a generative AI system. Prompts can contain confidential material, which is why data rules ('what may be entered into which tools') belong in every firm's AI policy.

RAG rating

Red / amber / green risk categorisation. The RICS standard requires each AI risk register entry to carry a RAG rating or similar. (Unrelated to the AI technique 'retrieval-augmented generation', which shares the acronym.)

Risk appetite

The level of risk a firm is prepared to accept in a given area — a required field for each entry in the AI risk register, because 'amber' means nothing without knowing what the firm tolerates.

Risk register (AI)

The living document RICS 3.3 requires: risks described, scored for likelihood and impact, mitigation planned, appetite stated, status updated, RAG-rated — and reviewed at least quarterly. See what goes in it and the free template.

Shadow AI

AI used by staff without the firm's knowledge or governance — personal ChatGPT accounts on client work being the classic case. The strongest argument against blanket bans: prohibition doesn't stop use, it stops visibility.

Systems register

See AI systems register.

Terms of engagement (AI clauses)

Where AI use is material, engagement documents must set out six things in writing: when AI is involved, which parts of the process, PI cover (if available), and the client's routes to contest, seek redress and opt out (if at all). See the six clauses.

Training data

The data an AI model learned from. Its accuracy, relevance, diversity, known gaps and bias risks are among the six things firms must ask suppliers about in writing during due diligence.

Definitions are general information in the context of the RICS AI standard, not legal advice. ComplyQS is not affiliated with or endorsed by RICS.