Skip to content
SYS checking…
IPIIR
LIVE · indicators
Research lines6active programmes◆
Open repositories5on GitHub◆
Partner organisations4universities and institutes◆
Projects under way11across three lines◆
Datasets3openly available◆
Peer-reviewed papers4in preparation◆
Events held7seminars and colloquia◆
Pages of documentation240open knowledge base◆
Research lines6active programmes◆
Open repositories5on GitHub◆
Partner organisations4universities and institutes◆
Projects under way11across three lines◆
Datasets3openly available◆
Peer-reviewed papers4in preparation◆
Events held7seminars and colloquia◆
Pages of documentation240open knowledge base◆
Glossary

Glossary of applied AI and robotics

The IPIIR glossary is an open dictionary of the terms the institute uses in its own research on applied AI and robotics. Every entry gives one self-contained definition, a short context, and links to the work in which the term is examined.

33 terms

A

Adoption barriers

Барьеры внедрения

Also known as: adoption barriers · барьеры адопции

Adoption barriers are the conditions that stop an AI tool from being adopted in an organisation: cost, missing skills, distrust, regulatory uncertainty, and incompatibility with existing processes.

The institute studies barriers with field methods — a survey across industries and regions and a case-study protocol — rather than expert opinion. The gap between declared and actual use is recorded separately: on inspection, some reported deployments turn out to have been one-off trials. The study is to produce quantitative adoption benchmarks and an open dataset.

Work by the institute
Related terms
SMEs in the context of AIInternal championDeployment post-mortem

AI as an operational layer

Операционный слой ИИ

Also known as: AI as an operational layer · операционный контур ИИ

The operational layer of AI is the set of points in an organisation's processes where a model is embedded in everyday work and its output leads to an action rather than remaining a demonstration.

The framing is useful because it moves the question from technology to process: what matters is not that a model is present but which decision is taken on its output and who is accountable for that decision. The institute maintains an open catalogue of patterns for integrating large language models into operational processes, recording for each pattern its conditions of applicability, typical architecture, effect metrics and characteristic causes of failure.

Work by the institute
Related terms
Applied AILLM integration patternReturn on investment of automation

Applied AI

Прикладной ИИ

Also known as: applied AI · прикладной искусственный интеллект

Applied AI is the use of artificial-intelligence methods to solve specific tasks inside the existing work processes of organisations, as distinct from research into the models themselves.

The subject of applied AI is not model architecture but what happens where a tool meets people's work: who accepts or rejects the system's suggestion, what maintenance costs, how responsibilities are redistributed. IPIIR studies that layer with field methods — by observing real deployments rather than laboratory set-ups. The institute's six research lines approach it from different angles, from human-AI interaction to regulation and machine perception.

Related terms
AI as an operational layerLLM integration patternAdoption barriers

↑ Back to the index

B

Beyond-benchmark quality metric

Метрика качества за пределами бенчмарка

Also known as: beyond-benchmark metrics · эксплуатационные метрики

A beyond-benchmark quality metric is an operational measure of a model taken under the conditions of the real process rather than on a reference dataset.

Examples include robustness to lighting drift, the cost of a false alarm against that of a missed defect, and degradation on rare defect classes. What they share is a link to the consequences of an error in a specific process rather than to averaged accuracy. The institute applies the same principle to measuring the effect of automation: what counts is the effect inside the process, not the figure on the test bench.

Work by the institute
Related terms
Computer vision on the production lineFalse positive on the production lineReturn on investment of automation

↑ Back to the index

C

Computer vision on the production line

Компьютерное зрение на линии

Also known as: computer vision on the line · машинное зрение на линии · визуальный контроль качества

Computer vision on the production line is the use of image-recognition models for quality control directly in the production flow, at the speed of the line and under its real lighting and load.

Laboratory accuracy on a reference dataset is a poor predictor of such a system's fitness: the line brings lighting drift, rare defect classes and variable load. The institute proposes a set of operational metrics in place of a single accuracy figure, with trials planned on industrial data from the line's partners.

Work by the institute
Related terms
Machine perceptionFalse positive on the production lineBeyond-benchmark quality metric

↑ Back to the index

D

Data specification

Спецификация данных

Also known as: data specification · словарь данных · data dictionary

A data specification is a document that fixes, before collection begins, the fields of a future dataset, the dictionary of values, the anonymisation procedures and the distribution licence.

Publishing the specification in advance lets the industry and other researchers comment on the structure while it can still be changed, and limits the temptation to fit the data to a desired result. For the dataset on AI use in small and medium-sized businesses, version 0.1 of the specification is open for comment. The institute publishes study protocols before going into the field for the same reason.

Work by the institute
Related terms
Open datasetDatasheet for a datasetSelf-reports versus telemetry

Datasheet for a dataset

Документирование данных (datasheet)

Also known as: datasheet · паспорт набора данных · документирование данных

A datasheet for a dataset is a description explaining how the data were collected, what they contain, which limitations and assumptions they carry, and on what terms they are distributed.

A datasheet answers the questions that cannot be reconstructed from the data themselves: how the sample was formed, what was excluded from it, how anonymisation was carried out. For the institute's open dataset that role is played by a published specification with a field dictionary and a CC BY 4.0 licence. In regulated environments documenting the data belongs to the minimum standard on equal terms with documenting the model.

Work by the institute
Related terms
Model cardData specificationOpen dataset

Deployment post-mortem

Постмортем внедрения

Also known as: post-mortem · разбор провала внедрения

A deployment post-mortem is a structured review of a failed AI deployment: interviews with the participants, a classification of the causes of failure, and a final report following a common template.

The institute separates causes into technical, organisational and economic ones — it is the conflation of these layers that leads a company to the wrong conclusions about its own experience. The protocol includes anonymisation rules that allow reviews to be published without harming the participants. Collected post-mortems feed the open catalogue of integration patterns and anti-patterns.

Work by the institute
Related terms
Integration anti-patternPilot projectAdoption barriers

↑ Back to the index

E

Explainability

Объяснимость

Also known as: explainability · XAI · интерпретируемость

Explainability is the property of an AI system to give a user or an auditor intelligible grounds for its output: sources, degree of confidence, and the limits of applicability.

In regulated environments explainability has stopped being a preference: requirements for the reproducibility and explainability of decisions are being written into Russian regulation, which the institute's report systematises. In consumer interfaces it takes the form of confidence estimates, links to sources and labelling of generated content. The institute distinguishes explainability from its imitation — presentation that creates a false sense of reliability while explaining nothing.

Work by the institute
Related terms
Trust interfaceModel cardRegulated environment

↑ Back to the index

F

False positive on the production line

Ложноположительное срабатывание на линии

Also known as: false positive · ложная тревога · ложное срабатывание

A false positive on the production line is a case in which the inspection system flags a good part as defective, causing a stoppage, a re-check or a rejection without cause.

The cost of this error and the cost of a missed defect differ by an order of magnitude and depend on the process, so the decision threshold is tuned by the cost of consequences rather than by model accuracy. Frequent false alarms also devalue the system in the operators' eyes: the signal starts being ignored and control stops working altogether. The institute includes the balance of these two errors in its set of operational metrics.

Work by the institute
Related terms
Computer vision on the production lineBeyond-benchmark quality metricHuman-in-the-loop

Federal Law 243-FZ

243-ФЗ

Also known as: Федеральный закон № 243-ФЗ · Federal Law 243-FZ

Federal Law 243-FZ, adopted on 26 July 2026 and in force from 1 September 2026, moves the regulation of artificial intelligence in Russia from strategy documents to binding rules.

The law marks the shift from soft law to positive legislation and changes the operational duties of organisations in regulated environments. For medical organisations that use clinical decision support systems, the institute has translated its requirements into concrete duties in a separate policy brief. The wider regulatory picture — experimental legal regimes, sectoral requirements, the allocation of liability — is examined in the report on the legal contour of applied AI.

Work by the institute
Related terms
Regulated environmentModel cardExplainability

↑ Back to the index

H

Human-in-the-loop

Human-in-the-loop

Also known as: HITL · человек в контуре принятия решения · человек в цикле

Human-in-the-loop is an arrangement in which a person is a required part of an AI system's decision loop, confirming, correcting or rejecting the model's output before it results in an action.

Mandatory confirmation before irreversible actions is one of the mechanisms that improve the quality of joint human-model decisions. The reverse side is habituation: once confirmation becomes a ritual, oversight exists on paper only. The institute examines such mechanics in its review of trust interfaces and observes them in a field study of office staff working with AI assistants.

Work by the institute
Related terms
Trust in an AI assistantTrust interfaceExplainability

↑ Back to the index

I

Integration anti-pattern

Анти-паттерн внедрения

Also known as: integration anti-pattern · анти-паттерн интеграции

An integration anti-pattern is a recurring decision that looks reasonable at the outset and systematically drives a deployment to failure.

Anti-patterns accumulate from post-mortems and field observation and are published alongside the patterns: without them a catalogue gives the one-sided picture in which everything works. Typical examples are automating a process that would be cheaper to fix organisationally, and an interface that creates a false sense of reliability. In service robotisation the same role is played by anti-scenarios.

Work by the institute
Related terms
LLM integration patternDeployment post-mortemRobotisation anti-scenario

Internal champion

Внутренний энтузиаст

Also known as: internal champion · champion · локальный лидер внедрения

An internal champion is an employee who in practice carries an AI deployment inside a company without a formal mandate or dedicated resources.

In small companies this role more often than not decides the outcome of the first ninety days. Dependence on a single person makes the result fragile: once they leave, the tool is usually abandoned even if it worked. The institute observes this role in its first-ninety-days framework and in its study of adoption barriers in small and medium-sized businesses.

Work by the institute
Related terms
Adoption barriersPilot projectSMEs in the context of AI

↑ Back to the index

L

Land surface temperature (LST)

LST (температура поверхности)

Also known as: LST · land surface temperature · температура поверхности

Land surface temperature (LST) is the temperature of the ground computed from the thermal bands of satellite imagery; it is not air temperature and in hot weather runs markedly higher.

LST is the base quantity in heat island maps: zonal statistics within administrative boundaries are computed from it. An image captures the moment of the satellite pass, so a single scene says nothing about the daily cycle. The institute's pipeline passes a publication gate with a root-mean-square error threshold of 2.5 K; below that threshold nothing is published.

Work by the institute
Related terms
Urban heat islandPMTilesBeyond-benchmark quality metric

LLM integration pattern

Паттерн внедрения LLM

Also known as: LLM integration pattern · паттерн интеграции LLM

An LLM integration pattern is a reproducible way of embedding a large language model into an operational process, described through its conditions of applicability, typical architecture, effect metrics and characteristic causes of failure.

The IPIIR catalogue is built bottom-up, from reviews of real deployments, predominantly in small and medium-sized businesses. A pattern counts as described only once its boundaries — the conditions under which it stops working — are recorded alongside the successful cases. Version 0.1 of the catalogue is open and grows as verified cases accumulate.

Work by the institute
Related terms
Integration anti-patternAI as an operational layerReturn on investment of automation

↑ Back to the index

M

Machine perception

Машинное восприятие

Also known as: machine perception · промышленное машинное восприятие

Machine perception is a technical system's ability to derive from sensor signals enough information about the state of the physical environment to act within it.

In industrial and logistics environments perception does not operate under laboratory conditions: lighting changes, rare defect classes appear, equipment ages, and the pace is set by the line rather than by the model. The institute's research line studies which properties of perception systems determine their fitness on a real site. Computer vision is the most developed part of this area but not the only one.

Work by the institute
Related terms
Computer vision on the production lineBeyond-benchmark quality metricService robotisation

Model card

Документирование моделей (model card)

Also known as: model card · карточка модели · документирование моделей

A model card is the minimum set of information about a model that must accompany the delivery of a solution: intended use, training data, quality metrics and known limitations.

Without such a document an audit or an incident investigation is impossible in principle: after the fact there is usually no one left who can reconstruct what the model was trained on and what for. The institute proposes a minimum documentation standard for regulated environments, drawing on international practice in model cards and datasheets and adapting it to Russian requirements. The draft is open for industry comment.

Work by the institute
Related terms
Datasheet for a datasetRegulated environmentExplainability

↑ Back to the index

O

Open dataset

Открытый набор данных

Also known as: open dataset · открытые данные · open data

An open dataset is research data published under a free licence together with a field dictionary and a description of the collection procedure, so that the conclusions can be re-checked independently.

The institute publishes data as a result in their own right rather than as an appendix to a text: the dataset on the use of AI in small and medium-sized businesses is distributed under CC BY 4.0, and the urban heat island monitoring results are reproducible from the source imagery through the project's pipeline. Openness here is a condition of verifiability, not a gesture of goodwill.

Work by the institute
Related terms
Data specificationDatasheet for a datasetUrban heat island

↑ Back to the index

P

Pilot project

Пилотный проект

Also known as: pilot project · пилот · proof of concept · PoC

A pilot project is a deployment of an AI or robotic solution limited in scale and duration whose purpose is to test the solution's fitness under real conditions rather than to demonstrate the technology.

The institute's field observation of robotisation pilots reveals a stable set of errors: choosing a task that would be cheaper to fix by changing the process, underestimating the cost of maintenance and staff training, and having no stopping criterion. A pilot with no failure condition stated in advance becomes a demonstration: it cannot be failed, and so nothing can be learned from it.

Work by the institute
Related terms
Deployment post-mortemRobotisation anti-scenarioInternal champion

PMTiles

PMTiles

Also known as: PMTiles · облачно-оптимизированные тайлы

PMTiles is a single-file map storage format that allows interactive layers to be served from static hosting without a tile server.

The format removes the running cost of publishing maps: the file sits next to the static assets and the browser requests only the fragments it needs. The institute publishes the urban heat island map layers as PMTiles. For a research organisation this is also a question of longevity: the map outlives the project because it does not depend on a running server.

Work by the institute
Related terms
Urban heat islandLand surface temperature (LST)Open dataset

↑ Back to the index

R

Regulated environment

Регулируемая среда

Also known as: regulated environment · регулируемая отрасль

A regulated environment is an industry in which the use of AI is constrained by sector-specific legislation and supervision: healthcare, financial markets and education.

Such environments impose requirements on reproducibility, explainability and the explicit allocation of liability that do not apply in the general case. The institute has systematised the Russian legal regime for the use of AI in three regulated environments over the period from October 2019 to August 2026. The practical consequence for a developer is a minimum documentation standard for the model and the data, without which an audit is impossible.

Work by the institute
Related terms
Federal Law 243-FZModel cardExplainability

Return on investment of automation

ROI автоматизации

Also known as: ROI · return on investment · окупаемость автоматизации

The return on investment of automation is the ratio of the real gain from an AI or robotic deployment to its total cost, calculated with account of the work that did not disappear after deployment but moved to other operations.

The usual error is to book the operator's saved time as a straight saving. Part of the freed time goes into checking the model's output, reworking results and handling exceptions, and therefore stays inside the process. The institute proposes a set of metrics that separate real gains from a repackaging of effort, and tests them on the deployments in its pattern catalogue.

Work by the institute
Related terms
Work redistribution versus savingTotal cost of ownership of a robotised solutionBeyond-benchmark quality metric

Robotisation anti-scenario

Анти-сценарий роботизации

Also known as: robotisation anti-scenario · анти-сценарий

A robotisation anti-scenario is a recurring situation in which robotisation does not solve the original task and calls for the process itself to be reconsidered.

The catalogue of anti-scenarios is the core of the assessment framework for service environments: it answers the question of when not to robotise, which industry material usually leaves out. A frequent example from field observation is a task that is cheaper to remove by changing the process than to hand to a robot. An anti-scenario differs from a failed pilot: it is a property of the task, not of the execution.

Work by the institute
Related terms
Service robotisationIntegration anti-patternTotal cost of ownership of a robotised solution

↑ Back to the index

S

Self-reports versus telemetry

Самоотчёты против телеметрии

Also known as: self-reports vs telemetry · декларируемое и фактическое использование

Self-reports versus telemetry is the methodological opposition between two ways of measuring the use of AI: asking employees about their work and automatically recording what they do.

A persistent discrepancy appears between declared and actual use, and it runs in both directions. Each method yields valid data under its own conditions: a self-report explains motives and the context of a decision, telemetry records frequency and outcome. The institute compares them in a dedicated paper and builds the measurement layer of its field studies on a combination of both.

Work by the institute
Related terms
Usage telemetryAdoption barriersData specification

Service robotisation

Сервисная роботизация

Also known as: service robotisation · сервисные роботы

Service robotisation is the use of robots in environments that serve people — hospitality, cleaning, delivery — where the robot works alongside humans in unstructured space.

The decisive question here is not technical feasibility but the division of tasks between human and robot and the total cost of ownership against the manual process. The institute maintains an assessment framework for such scenarios and a catalogue of anti-scenarios, drawing on field observation in hospitality: room delivery, cleaning, reception areas.

Work by the institute
Related terms
Robotisation anti-scenarioTotal cost of ownership of a robotised solutionPilot project

SMEs in the context of AI

МСБ в контексте ИИ

Also known as: МСБ · малый и средний бизнес · SME · SMB

SMEs in the context of AI are small and medium-sized companies whose adoption of AI is limited not by access to models but by the absence of resources for integration, maintenance and staff training.

Small and medium-sized businesses are barely represented in industry surveys of AI adoption, although they make up the bulk of all organisations. IPIIR devotes a research line to them: a field study of adoption barriers, an observation framework for the first ninety days, and an open dataset on the use of AI tools. The dataset specification was published before collection began so that the industry could comment on its structure in advance.

Work by the institute
Related terms
Adoption barriersInternal championOpen dataset

↑ Back to the index

T

Total cost of ownership of a robotised solution

TCO роботизированного решения

Also known as: TCO · total cost of ownership · полная стоимость владения

The total cost of ownership of a robotised solution is the full cost of running a robot over its service life: purchase, integration, maintenance, downtime, staff training, and the cost of reverting to the manual process.

Robotisation decisions are almost always made on the price of the device, whereas most of the spending appears after installation. The institute's methodology computes TCO over a three-to-five-year horizon and compares the robotised process with the manual one rather than with an idealised scenario. The calculation sheets are published in an open format and are being tested on pilot cases.

Work by the institute
Related terms
Return on investment of automationService robotisationRobotisation anti-scenario

Trust in an AI assistant

Доверие к ИИ-ассистенту

Also known as: trust in AI assistants · калиброванное доверие

Trust in an AI assistant is an employee's readiness to accept the system's recommendation without independent verification, in proportion to how reliable that system actually is.

What helps is not maximal trust but calibrated trust: too much of it leads to errors being accepted uncritically, too little makes the tool worthless. IPIIR measures how staff in office roles accept, correct and reject assistants' recommendations, combining diary self-reports, semi-structured interviews and usage telemetry. The study protocol was published before data collection began, so that hypotheses and metrics were fixed in advance.

Work by the institute
Related terms
Human-in-the-loopTrust interfaceUsage telemetry

Trust interface

Интерфейс доверия

Also known as: trust interface · интерфейс объяснимости

A trust interface is the set of interface elements through which an AI tool shows the user the limits of its competence: confidence estimates, links to sources, labelling of generated content, and confirmation before irreversible actions.

The institute's review systematises these elements and separates the patterns that genuinely improve joint human-model decisions from the anti-patterns that manufacture a false sense of reliability. A trust interface is the practical side of explainability: the user gets a signal at the moment of work rather than a report after the fact.

Work by the institute
Related terms
ExplainabilityHuman-in-the-loopIntegration anti-pattern

↑ Back to the index

U

Urban heat island

Остров тепла

Also known as: urban heat island · UHI · городской остров тепла

An urban heat island is an area of a city that stays persistently warmer than the surrounding land because of dense development, paved surfaces and a shortage of vegetation.

The institute builds open maps of urban heat islands in Russia from Landsat and Sentinel-2 imagery at 30-metre resolution. Across ten cities over 2015-2025 the results include the highest mean surface temperature in Astrakhan (43.6 °C) and the fastest growth in Moscow (+1.25 °C a year by the city median). Every figure is reproducible from the source data through the project's pipeline.

Work by the institute
Related terms
Land surface temperature (LST)PMTilesOpen dataset

Usage telemetry

Телеметрия использования

Also known as: usage telemetry · телеметрия рабочего места

Usage telemetry is automatically collected data on what an employee actually does with an AI tool: how often it is invoked, which suggestions are accepted or rejected, and how long it is used.

Telemetry closes the gap between what employees report about their use of AI and what actually happens. The institute treats it together with its ethical constraints: informed consent, anonymisation, and the employee's right to see the data collected about them. Without those conditions workplace telemetry is unfit for research, however precise the measurement.

Work by the institute
Related terms
Self-reports versus telemetryTrust in an AI assistantBeyond-benchmark quality metric

↑ Back to the index

W

Work redistribution versus saving

Перераспределение работы против экономии

Also known as: work redistribution · переупаковка трудозатрат

Work redistribution is the effect whereby the time freed by automation does not leave the process but moves to checking the model's output, reworking results and handling exceptions.

Redistribution can be told apart from saving only by measuring at the level of the process rather than the single operation, because at the level of the operation the gain is always visible. Without that measurement the reported effect of a deployment turns out to be a repackaging of effort rather than a return. The institute proposes metrics that separate the two cases.

Work by the institute
Related terms
Return on investment of automationBeyond-benchmark quality metricHuman-in-the-loop

↑ Back to the index

How to use this glossary

Every term has a permanent link: ipiir.ru/research/glossary#tco goes straight to the definition of the total cost of ownership of a robotised solution. Such links are convenient to cite in articles and internal documents.

The glossary records terms as they are used in the institute's own work and grows together with it. It is neither an industry standard nor a translation of external terminology: where a definition departs from common usage, the context explains why.

Research lines → All publications →