AI Transparency Cards

Guide to AI Transparency Cards

Everything you need to know about our AI features, in one place.

An AI Transparency Card is a one page summary of a single AI feature in a Bentley product. Use At a glance, What the feature does, and Trust & safety to find the basics, how the AI works, and the controls that protect your data. For common questions, see the FAQs below.

At a glance

Product

The Bentley application that includes the feature. We publish a separate card for each product and feature pair, because enablement and safeguards can vary by product.

Feature name

The feature's official name, for example Bentley Copilot, Label Organizer, or Intelligent PID Import. The name stays the same across every product that includes the feature.

Description

A plain language summary of what the feature does.

What the feature does

Feature functionality

The kind of work the AI performs. Each feature lists up to three of these, so you can see its main role at a glance.

AutomatePerforms repetitive tasks and can repeat them at scale.
AnalyzeExamines your data and provides actionable insights.
AugmentEnhances your creative exploration and problem solving, while you stay in control.
ClassifySorts or labels inputs into predefined categories.
PredictForecasts numeric values or the likelihood of future events from input data.
GenerateCreates original content, such as text, sound, images, or video, from your prompt.
ActAn AI agent that can reason and carry out multi step actions, with or without a human confirming each step.

Model source

Who built and owns the underlying AI model. Ownership determines who controls the model and, in turn, how your inputs are governed and protected.

ProprietaryBuilt and owned by Bentley. Bentley controls the full training pipeline.
Fit-for-accountA model trained or fine tuned specifically for your account, using data you have agreed to provide.
LicensedA third party model used under license. Bentley's contracts with these providers govern how your inputs are handled.
OpenA publicly released open-source model that anyone can inspect and run. Bentley operates it within its own secured infrastructure.

Primary technique

The main AI technology behind the feature. Cards list the first technique that applies, even when more than one technique is used.

Large language modelUnderstands and generates language, and sometimes audio or images. It powers conversational features such as Bentley Copilot.
Diffusion modelCreates images or other data by progressively refining noise.
TransformerA machine learning architecture for understanding sequences and structure.
Neural networkA deep learning technique that learns patterns from data through layered, interconnected nodes.
EncodingConverts data into a representation that models can process efficiently.
ClassificationSupervised learning that assigns items to predefined categories.
PredictorLearns from historical data to forecast outcomes or trends.
Genetic algorithmAn optimization method inspired by natural selection, used to explore constrained or unconstrained design spaces.
Statistical analysisUses statistical methods to find patterns and relationships in data, without a trained model.

Input format

The types of content the feature takes in to do its work. Knowing what a feature reads helps you understand what data it needs and what it can access.

TextWritten language the feature reads, such as prompts, questions, or descriptions.
ImageStill pictures the feature reads, such as photos, scans, or screenshots.
AudioSound the feature listens to, such as speech or recordings.
VideoMoving images the feature reads, with or without sound.
Drawing sheetAn engineering drawing sheet from the design file.
Time series dataValues recorded over time, such as sensor readings.
AlertsEvent or alarm notifications, such as IoT sensor alerts.
Python scriptExecutable Python code supplied to the feature.
PDFA PDF document the feature reads.

Output format

The types of content the feature returns, so you know what to expect back and in what form.

TextWritten language the feature produces, such as summaries, answers, or descriptions.
ImageStill pictures the feature produces, such as renderings or generated images.
AudioSound the feature produces, such as synthesized speech.
Drawing sheetAn engineering drawing sheet the feature produces.
Time series dataValues over time the feature produces, such as forecasts.
TagsLabels or keywords the feature applies to content.
Statistical predictionsNumeric forecasts or probabilities the feature produces.
Python scriptExecutable Python code the feature generates.
JSONStructured data in JSON format.
iModelA Bentley iModel containing engineering data.

Trust & safety

Enablement

How the feature gets turned on or off, and how much control you and your administrators have.

Opt-in at install timeOff by default, and someone must deliberately enable it during installation.
Opt-in at opening timeOff by default, and you choose to enable it when opening the application or file.
Opt-out at project creationOn by default for new projects, but it can be switched off when the project is created.
Opt-out at install timeOn by default, but it can be disabled during installation.
Opt-in by administrator at installUnavailable to users unless an administrator enables it before installation.
Opt-in by administratorUnavailable to users unless an administrator enables it, and that can happen at any time.
Opt-out by userAvailable by default, and each user can disable it.
Opt-out by administratorAvailable by default, and an administrator can disable it for all users.
AI-first productThe product is built around AI, so purchasing the software is considered opting in.
N/AThe feature is an inherent part of the product or service and has no separate switch.

Training data source

What data Bentley used to train the model. Check here to understand how your own content relates to model training. For licensed third party models, anything you add through prompts or retrieval is not training data and is not listed here.

N/ABentley did no training at all. The feature uses a third party model as is, and your data was not used to train anything.
ProprietaryTrained only on Bentley-owned data, such as Bentley documentation or internally produced datasets.
Purchased/LicensedTrained on data acquired from users or partners under an explicit license.
SyntheticTrained on artificially generated data.
Subscriber/User dataTrained on subscriber data — used only for fit-for-account models that you commission for your own account.
Open-sourceTrained on freely available open data.

Fine-tunable

Whether the generic model can be fine tuned with your account data, at your request, to create a fit for account model.

YesThe generic model can be fine tuned with your account data, at your request, to create a fit for account model.
NoThe model is fixed and is never adjusted with customer data.

Personal data

Whether any personal data is involved in training the model or when you use the feature.

NoThe feature does not learn from or process personal data.
YesPersonal data is involved in training or when you use the feature, and the safeguards below explain how that data is protected and how it is handled.

Feedback mechanism

The ways you can tell Bentley whether an AI response was helpful. Your feedback guides how we improve the feature over time.

Thumbs upA quick way to mark an AI response as helpful.
Thumbs downA quick way to flag an AI response as not helpful.
Written commentsA free-text box for sharing more detailed feedback in your own words.
NoThe feature has no feedback channel.

Feedback enablement

Whether and how your feedback is collected, and who controls it, so nothing is captured without a deliberate choice.

Opt-inFeedback is only collected when you actively choose to send it. Pressing the feedback button is the opt in, and nothing is collected silently.
Opt-in by administratorThe feedback channel is off until an administrator enables it.
Opt-in by user with opt-out by administratorYou opt in by giving feedback, but your administrator can turn the channel off for the whole organization.
No feedbackNo feedback is collected.
N/ANot applicable, because the feature has no feedback mechanism.

Encryption at rest

Whether data handled by the feature is encrypted while it is stored.

YesIndustry standard encryption is applied to your data while it is stored, so it stays protected at rest.
NoEncryption is not applied at this stage, even though there is data that could be encrypted.
N/AThere is nothing to encrypt: the feature doesn't store your data (at-rest) or nothing leaves your machine (in-transit). A stronger privacy statement than Yes.

Encryption in transit

Whether data handled by the feature is encrypted while it moves over the network.

YesIndustry standard encryption is applied to your data while it moves over the network, so it stays protected in transit.
NoEncryption is not applied at this stage, even though there is data that could be encrypted.
N/AThere is nothing to encrypt: the feature doesn't store your data (at-rest) or nothing leaves your machine (in-transit). A stronger privacy statement than Yes.

Risk / bias mitigation

The strategies in place to keep the AI's output safe, fair, and reliable, especially where people need to trust the result.

Human oversightThe AI does not take consequential actions on its own. You review and confirm, so AI output is a suggestion rather than an unattended decision.
Bias testing/auditsThe model and its data are tested or audited for bias.
Input validationInputs are checked and constrained before reaching the model, which helps reduce misuse and poor output.
Rate-limitingUsage is throttled to prevent abuse and runaway behavior.
ExplainabilityThe feature shows its reasoning or sources, for example citations, so you can verify answers instead of taking them on trust.
User consent mechanismsThe feature asks for your consent before specific data uses or actions.

Other safeguards

Extra privacy and data-protection measures a feature applies on top of standard encryption and access controls, such as anonymization or differential privacy.

AnonymizationData is stripped of identifying information so it cannot be traced back to a person.
DeidentificationDirect identifiers are removed or masked before the data is processed.
Differential privacyMathematical noise is added so individual records cannot be reconstructed.
Access controlThe AI can only see content that you are already authorized to see. It does not widen access to your projects.
N/ANo additional safeguards beyond those already listed are needed for this feature.