Capability records, perspective 4: Data, knowledge, and memory

Part of the SEI AI Adoption Framework (AIAF).

4.1 Data quality, lineage, and provenance

Purpose

Prevent AI from being built or run on data whose source, meaning, permissions, or quality limits are unknown, so that errors and misuse cannot be traced.

Intended outcome

Important data can be traced to its source, meaning, permissions, transformations, and quality limitations.

Decisions enabled

  • whether data is fit for the intended AI use;
  • which sources may be used and for what;
  • what quality issues must be fixed or disclosed; and
  • who is accountable for correcting a data problem.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • identify the data an AI use depends on and its source and owner;
  • record meaning, permitted purposes, transformations, and known limitations;
  • define quality rules and check them at points where data enters the AI use;
  • trace outputs back to inputs where consequence requires it;
  • log, assign, and resolve data issues; and
  • reflect source changes in dependent uses.

Required evidence

  • lineage and provenance records;
  • quality rules and results;
  • permission and purpose records;
  • issue log with resolution; and
  • change notifications to dependent uses.

Common failure patterns

  • data used because it is available;
  • lineage lost through copying and transformation;
  • quality measured at the source but not where it is used;
  • issues fixed locally without telling the owner; and
  • purpose limits not recorded.

Capability levels

Level Observable condition
Emerging Data is used as found, and its origin and quality are known only to the people who prepared it.
Defined In-scope data sources have recorded owners, purposes, and basic quality rules.
Managed Lineage and quality checks cover the path into the AI use, and issues are tracked and resolved with the owners.
Adaptive Quality and lineage controls adapt to observed AI errors, and the organization detects source change before it affects outputs.

Review and exit triggers

  • a source changes meaning, format, or permission;
  • quality falls below the rule;
  • outputs cannot be explained from inputs;
  • a new purpose is proposed for existing data; or
  • an owner changes.

4.2 Training, tuning, and evaluation data

Purpose

Prevent models and evaluations from being built on data that is unlawful, unrepresentative, contaminated, or untraceable, which makes results unreliable or unusable.

Intended outcome

Data used to train, tune, test, or compare AI is suitable, permitted, representative, and versioned.

Decisions enabled

  • whether data may be used to train, tune, test, or compare;
  • whether the data represents the population and conditions of use;
  • how to keep evaluation data independent of training data; and
  • which versions of data support which results.

Applicability

Workloads: All
Delivery modes: A, D
Typical state: Required

Key activities

  • record the origin, license, and permitted use of each dataset;
  • check consent, confidentiality, and contractual limits before use;
  • assess representativeness against the intended use and affected groups;
  • keep evaluation data separate from tuning data and protect it from leakage;
  • version datasets and link them to models and results; and
  • remove or replace data that cannot be used.

Required evidence

  • dataset records with license and permitted use;
  • representativeness review;
  • separation and leakage controls;
  • version links to models and evaluations; and
  • removal or replacement records.

Common failure patterns

  • scraped or purchased data used without rights review;
  • evaluation data reused for tuning until results look good;
  • datasets changed without a new version;
  • data that reflects only the easiest cases; and
  • no ability to remove data on request.

Capability levels

Level Observable condition
Emerging Datasets are gathered by teams and used without consistent records.
Defined Datasets used for training, tuning, and evaluation are recorded with source, license, and version.
Managed Representativeness and separation are reviewed, versions link to results, and unsuitable data is removed.
Adaptive Data practice adapts to drift and to findings from evaluation, and the organization refreshes datasets before they stop representing the work.

Review and exit triggers

  • a dataset's license or consent basis is challenged;
  • the population or conditions of use change;
  • evidence of leakage appears;
  • performance in operation departs from evaluation; or
  • data must be removed.

4.3 AI information classification and handling

Purpose

Prevent sensitive information from leaking or being misused through places that ordinary data rules do not reach, such as prompts, retrieved text, logs, memory, actions, and generated material.

Intended outcome

Rules cover prompts, inputs, outputs, retrieval sources, logs, memory, actions, and generated artifacts.

Decisions enabled

  • which information may enter an AI use;
  • how outputs and logs are classified and handled;
  • which tools and locations are permitted for each class; and
  • when an exception may be approved.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • extend classification rules to prompts, inputs, outputs, retrieval sources, logs, memory, actions, and artifacts;
  • set handling controls for each class and delivery mode;
  • tell people what they may enter into which tools;
  • apply technical controls where possible, such as filtering, masking, and access limits;
  • handle exceptions through recorded approval; and
  • review the rules when tools or data change.

Required evidence

  • classification rules covering AI artifacts;
  • handling controls by class and mode;
  • user guidance;
  • exception records; and
  • control test results.

Common failure patterns

  • rules written for documents but not for prompts;
  • outputs inherit no classification from inputs;
  • logs that keep sensitive prompts without protection;
  • technical controls absent and policy relied on alone; and
  • exceptions that become the norm.

Capability levels

Level Observable condition
Emerging People decide case by case what to enter and how to treat outputs.
Defined Classification and handling rules include AI inputs and outputs for in-scope uses.
Managed Controls are applied and tested, outputs and logs inherit classification, and exceptions are recorded.
Adaptive Rules and controls adapt to new tools, observed leakage, and attacker behavior, and the organization detects misuse across the AI information flow.

Review and exit triggers

  • sensitive data appears in an unapproved place;
  • a tool changes its handling of inputs or outputs;
  • a new class of data or artifact is introduced;
  • a control test fails; or
  • exceptions rise.

4.4 Privacy-preserving design

Purpose

Prevent AI uses from collecting, exposing, inferring, or retaining more personal data than their purpose justifies.

Intended outcome

Personal data use is minimized and protected according to purpose, consequence, consent, and obligation.

Decisions enabled

  • whether personal data is needed at all;
  • what protections apply;
  • what consent or other basis supports use; and
  • whether residual privacy risk is acceptable.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • identify personal data in inputs, outputs, training, retrieval, logs, and memory;
  • assess privacy risk in proportion to purpose and consequence;
  • minimize data and apply protections such as de-identification, access limits, or processing controls;
  • record the lawful basis, consent, and notices where required;
  • test whether protections work, including whether personal data can be recovered or inferred; and
  • handle access, correction, and deletion requests.

Required evidence

  • privacy assessment;
  • minimization decisions;
  • selected protections and validation results;
  • basis and notice records; and
  • request handling records.

Common failure patterns

  • collecting data because it might be useful;
  • assuming de-identified data cannot be re-identified;
  • personal data hidden in prompts, logs, and memory;
  • assessment done at the start and not after change; and
  • requests for deletion that cannot reach the AI artifacts.

Capability levels

Level Observable condition
Emerging Privacy is considered when a concern is raised, and protections vary by team.
Defined In-scope uses have a privacy assessment and chosen protections.
Managed Protections are validated, requests are handled across AI artifacts, and assessments are repeated when use changes.
Adaptive Privacy design evolves with techniques and threats, and the organization reduces personal data in use without losing the outcomes it needs.

Review and exit triggers

  • a new personal data type or purpose is added;
  • a validation test fails;
  • a request cannot be completed;
  • law or consent basis changes; or
  • a privacy incident occurs.

4.5 Residency and sovereignty

Purpose

Prevent the organization from losing track of where data and AI artifacts are processed, stored, replicated, and accessed, and therefore from meeting location and transfer obligations.

Intended outcome

The organization knows where relevant data and artifacts are processed, stored, replicated, and accessed.

Decisions enabled

  • where processing and storage are permitted;
  • whether a transfer or access arrangement is acceptable;
  • which providers and locations to use or exclude; and
  • how to respond to a change of location or legal exposure.

Applicability

Workloads: All
Delivery modes: All
Typical state: Shared

Key activities

  • map flows of data and artifacts, including inference, logging, backup, support access, and subprocessors;
  • record locations and who can access them;
  • compare locations with legal, contractual, and customer requirements;
  • decide and record transfers and exceptions;
  • require provider evidence and notice of change; and
  • test or verify claims where consequence justifies it.

Required evidence

  • data-flow map;
  • locations and access records;
  • transfer decisions;
  • provider attestations and notices; and
  • verification results.

Provider evidence

Ask the provider for:

  • provider statements of processing, storage, support-access, and subprocessor locations;
  • notice of changes to locations or subprocessors; and
  • audit or assurance reports that address residency claims.

Common failure patterns

  • assuming the service location equals the processing location;
  • ignoring support access and backups;
  • location claims not verified;
  • subprocessors unknown; and
  • sovereignty requirements discovered after contract.

Capability levels

Level Observable condition
Emerging Locations are assumed from the provider's marketing or the contract summary.
Defined In-scope uses have a flow map and recorded locations, compared with requirements.
Managed Transfers are decided and recorded, provider evidence is obtained, and changes are tracked.
Adaptive Residency and sovereignty are managed as a design choice, and the organization can move processing when requirements or exposure change.

Review and exit triggers

  • a provider changes location, subprocessor, or support arrangement;
  • law or a customer requirement changes;
  • a flow is found that was not mapped;
  • a verification test contradicts a claim; or
  • a geopolitical or legal event raises exposure.

4.6 Artifact retention and disposition

Purpose

Prevent AI artifacts from being kept forever, deleted too early, or disposed of without verification, which leads to exposure, legal breach, or loss of evidence.

Intended outcome

Prompts, outputs, logs, evaluations, decisions, model artifacts, and derived data are retained or deleted deliberately.

Decisions enabled

  • how long each type of artifact is kept;
  • when legal or investigative holds apply;
  • how deletion is carried out and verified; and
  • where retention conflicts with other needs, such as evidence versus privacy.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • list artifact types such as prompts, outputs, logs, evaluations, decisions, model artifacts, and derived data;
  • set retention periods and reasons for each;
  • apply legal holds and releases;
  • delete or archive at the end of the period using methods suited to where the artifacts live;
  • verify deletion, including copies, backups, and provider-held data; and
  • resolve conflicts between retention rules.

Required evidence

  • retention schedule;
  • legal hold records;
  • deletion and archive records;
  • verification results; and
  • conflict decisions.

Common failure patterns

  • retention set for convenience;
  • logs kept without a purpose;
  • deletion requested but copies persist;
  • evidence deleted before an inquiry; and
  • provider-held artifacts excluded.

Capability levels

Level Observable condition
Emerging Retention depends on default settings, and nobody can say what is kept.
Defined A retention schedule covers the main artifact types for in-scope uses.
Managed Holds, deletions, and verification are carried out and recorded, including provider-held data.
Adaptive Retention adapts to obligations and risk, and the organization removes unneeded artifacts while retaining the evidence it needs.

Review and exit triggers

  • a hold or inquiry is received;
  • a verification test finds data remaining;
  • law or contract changes retention requirements;
  • new artifact types appear; or
  • storage cost or exposure increases.

4.7 Feedback and evaluation-data quality

Purpose

Prevent feedback and evaluation data from misleading improvement, because it is biased, incomplete, manipulated, or collected without protection.

Intended outcome

Feedback used to judge or improve AI is collected, interpreted, protected, and tested for bias and usefulness.

Decisions enabled

  • which feedback to rely on;
  • how to interpret feedback from different groups;
  • whether evaluation data remains valid; and
  • what changes to make based on it.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • design how feedback is collected, including from those affected by outputs;
  • protect feedback and evaluation data and respect privacy;
  • test for bias in who provides feedback and in how it is labeled;
  • check usefulness against outcomes rather than satisfaction alone;
  • keep evaluation datasets current and separate from training data; and
  • record how feedback changed the AI or its use.

Required evidence

  • feedback design;
  • evaluation datasets and their upkeep;
  • bias and quality review;
  • protection controls; and
  • records of changes made.

Common failure patterns

  • ratings from a small, vocal group;
  • labels that reflect reviewer habits;
  • feedback used for training without notice;
  • stale evaluation data; and
  • no evidence that feedback was acted on.

Capability levels

Level Observable condition
Emerging Feedback is gathered informally, and its use is not recorded.
Defined Feedback and evaluation data are collected by design and have named owners.
Managed Bias and quality of feedback are examined, data is protected, and changes are traced to feedback.
Adaptive Feedback design adapts to what it fails to reveal, and the organization combines it with outcome evidence to improve safely.

Review and exit triggers

  • feedback comes from a narrow group;
  • ratings and outcomes diverge;
  • evaluation data becomes stale;
  • feedback is found to be manipulated; or
  • privacy or consent limits change.

4.8 Knowledge and retrieval management

Purpose

Prevent AI answers from resting on knowledge that is unauthorized, outdated, inaccessible, or unattributable.

Intended outcome

Knowledge sources are authorized, current, accessible, attributable, and evaluated for retrieval and answer quality.

Decisions enabled

  • which sources an AI use may draw upon;
  • who may retrieve what;
  • how freshness and quality are maintained; and
  • whether retrieval and answer quality are adequate.

Applicability

Workloads: All
Delivery modes: A, B, D
Typical state: Required

Key activities

  • inventory the sources and their owners and permissions;
  • preserve source access rights in retrieval so users see only what they are entitled to;
  • set freshness rules and remove or mark outdated content;
  • attribute answers to sources people can check;
  • test retrieval and answer quality with realistic questions and adversarial content; and
  • handle corrections and source retirement.

Required evidence

  • source inventory;
  • access model and tests;
  • freshness rules and results;
  • retrieval and answer tests; and
  • correction records.

Common failure patterns

  • indexing everything available;
  • retrieval that bypasses source permissions;
  • outdated documents ranked above current ones;
  • answers without citations; and
  • quality tested on easy questions.

Capability levels

Level Observable condition
Emerging Content is loaded into tools as convenient, and access and freshness are not checked.
Defined Sources are inventoried and approved with owners and basic access rules.
Managed Permissions carry through retrieval, freshness is enforced, and retrieval and answers are tested.
Adaptive Source management adapts to usage, errors, and content change, and the organization detects degraded retrieval before users do.

Review and exit triggers

  • a source owner or permission changes;
  • a retrieval test fails or an unauthorized result appears;
  • content ages past its freshness rule;
  • answers cite retired material; or
  • a new source class is added.

4.9 Memory and state governance

Purpose

Prevent persistent context and memory from accumulating personal, sensitive, or wrong information that no one can see, correct, or remove.

Intended outcome

Persistent context and memory have a defined purpose, owner, access model, correction process, retention period, and retirement treatment.

Decisions enabled

  • what the AI may remember and for how long;
  • who can see or change memory;
  • how users correct or delete it; and
  • what happens to memory when the AI is changed or retired.

Applicability

Workloads: All
Delivery modes: All
Typical state: Required

Key activities

  • define the purpose of each memory or persistent state;
  • assign an owner and an access model;
  • set what is stored, retention, and the sensitivity limits;
  • give users and owners a way to inspect, correct, and delete;
  • test that correction and deletion take effect; and
  • handle memory in migration, rollback, and retirement.

Required evidence

  • memory inventory and purposes;
  • permissions and ownership;
  • retention settings;
  • correction and deletion tests; and
  • retirement treatment.

Common failure patterns

  • memory enabled by default with no owner;
  • remembered errors presented as facts;
  • memory shared across users unintentionally;
  • deletion removes the display but not the stored state; and
  • memory carried into a replacement without review.

Capability levels

Level Observable condition
Emerging Memory features are used as provided, and nobody can say what is remembered.
Defined Persistent memory is inventoried with purpose, owner, and retention for in-scope uses.
Managed Access, correction, and deletion are tested and effective, and memory is handled at change and retirement.
Adaptive Memory design adapts to observed errors and misuse, and the organization limits persistence to what the outcome requires.

Review and exit triggers

  • memory holds data beyond its purpose;
  • a correction or deletion test fails;
  • memory leaks between users or contexts;
  • the AI, provider, or retention rule changes; or
  • an owner leaves.