BBM-STD-92-AI_GOVERNANCE-V0.1 | Office standard | 4 November 2026 | V0.1 | Private controlled WIP draft for office review
| Version | Date | Change |
|---|---|---|
| V0.1 | 2026-11-04 | First issue: closes rows,,,,, research finding R5 and §1.3 "AI governance annex" + cross-doc pointer pass (BBM-STD-87 SensitivityClass/T1–T2 wording) |
1 Purpose, scope and owner#
This standard governs every use of AI and large language models (LLMs) on office and project information: chat assistants, code and script generation, model checking, image generation, document drafting and summarising. Tool admission, pinned versions, fixtures and rollback for automation stay in BBM-STD-73; its AIUse record carries the fields in §5. The office director owns this standard; the BIM Manager keeps the AI register; the data-protection lead rules on personal data (BBM-STD-90).
2 Management-system interface#
| Reference | Office use |
|---|---|
| ISO/IEC 42001:2023 (AI management system, published 2023-12) | Structure for the office AI policy, roles, risk and impact assessment, register, monitoring and improvement. The office aligns to it; no certification is claimed. Each AIUse record holds the risk and impact assessment that 42001 asks for (§5 fields "Risk" and "Impact") |
| PAS 1958:2026 (Built environment. Data and information standards landscape. Guide, BSI, 31 Mar 2026) | Landscape reference for which data and information standards make project data fit for AI use; used to check that inputs given to AI are structured to the office standards (classification, IDS, IFC) before relying on outputs |
| ISO/IEC 27001:2022 + Amd 1:2024 | Information-security controls for AI services (access, logging, supplier security) through BBM-STD-87 |
| ISO 9001:2026 | Nonconformity and corrective action for AI incidents through BBM-STD-68 |
| NIST AI RMF | Voluntary context only |
3 Approved uses#
| Use | Allowed data | Condition |
|---|---|---|
| Drafting and editing office text (standards, emails, reports) | Office-internal, non-client information | Author remains responsible; output reviewed before issue |
| Script and graph generation (Dynamo, Python, C#) | Synthetic fixtures, office templates | Admitted and tested per BBM-STD-73 before running on a project model |
| Model and code-compliance checking | Project data only on an approved service (§4) | Advisory only (§6) |
| Summarising and searching project documents | Project data only on an approved service (§4) | Summary is not a record; cite the source container |
| Visualisation and image generation | Office or client-cleared imagery | Rights and labelling per BBM-STD-80; marketing use needs client confirmation |
| Anything else | None | Not approved until a register entry exists |
4 Data that may not leave the office#
Unless the AIUse record shows an approved service with recorded terms, hosting region, retention and training settings for that data class, these may not be sent to any external AI service:
| Data class | Rule |
|---|---|
| Client data | Models, drawings, briefs, costs and correspondence of any appointment: only to a service approved in writing for that appointment, within the client's confidentiality terms |
| Personal data | Only with a recorded lawful basis under Federal Decree-Law 45/2021 (PDPL) and a permitted transfer route if hosted outside the UAE (BBM-STD-90 §4); DIFC/ADGM regime where applicable, confirmed at project start |
| Security-classified information | Never to an external AI service when its SensitivityClass is Confidential or Restricted, or when it falls under a T1/T2 security management plan (BBM-STD-87 §2–§3); Internal information only to a service approved for Internal data; local-only processing of Confidential/Restricted data needs approval under the BBM-STD-87 SMP |
| Credentials | Never, in any prompt or file (BBM-STD-94) |
A training opt-out toggle or a vendor label is not evidence of zero retention; the AIUse record must point to the actual service terms and configuration. An authority portal that offers AI or BIM checking is used only through the authority route recorded in BBM-STD-85 with a verified link; its existence is not permission to send project data to another AI service. Unknown terms hold the upload.
5 AI register (AIUse record fields)#
| Field | Content |
|---|---|
| AIUseID, owner | Stable ID; accountable person |
| Provider, service, account type, endpoint | Exact service and plan used |
| Model and version | Model name and version/date; change policy (pinned or provider-updated) |
| Purpose and approved use (§3) | One purpose per record |
| Data classes permitted (§4) | Client / personal / classified / office-internal / synthetic |
| Hosting region, sub-processors, retention, training use | From the service terms, with link and date read |
| Personal-data check | Lawful basis, transfer route outside the UAE, free-zone regime: yes/no + reference to the BBM-STD-90 register row |
| Risk and impact | ISO/IEC 42001-style assessment: what can go wrong, who is affected, controls |
| Human reviewer | Named competent person who signs outputs (§6) |
| Licence | BBM-STD-94 entry ID |
| Review date, state | Candidate / Approved / Suspended / Retired |
6 Checks are advisory#
AI and LLM results are advisory. A mandatory requirement passes only on a deterministic check (IDS, rule-based checker, schema validation) or a competent human check; an AI result is never the check of record. Research shows LLM-generated checking scripts work on small rule samples, so every AI-assisted check output records: the rule source checked (code, clause or requirement ID and edition), the AI tool and model version, the prompt or script hash, and the named human who signed the result. Outputs without a signatory are not issued.
7 Incident handling#
An AI incident is any of: restricted data sent to an unapproved service, a wrong AI output used in an issued record, a service change of terms, or a model change that breaks a qualified use.
- Stop the affected use; set the AIUse record to Suspended.
- Record within one working day: what, which data, which records, who found it.
- Personal data involved: follow the BBM-STD-90 §6 breach procedure. Security information involved: BBM-STD-87 §4 security information and breach management.
- Wrong output in an issued record: raise a nonconformity (BBM-STD-68), correct and reissue under a new revision.
- Close with corrective action and a recheck before the record returns to Approved.
References#
- ISO/IEC 42001:2023, Artificial intelligence management system: https://www.iso.org/standard/42001, checked 2026-11-04.
- BSI PAS 1958:2026, Built environment. Data and information standards landscape. Guide: https://knowledge.bsigroup.com/search?query=PAS%201958, checked 2026-11-04.
- ISO/IEC 27001:2022: https://committee.iso.org/standard/82875.html, checked 2026-11-04.
- UAE Federal Decree-Law No. 45 of 2021 on the Protection of Personal Data: https://uaelegislation.gov.ae/en/legislations/1972, checked 2026-11-04.
- Madireddy et al. (2025), LLM-driven code compliance checking in BIM, arXiv:2506.20551 (preprint): https://arxiv.org/abs/2506.20551, checked 2026-11-04.
- NIST AI Risk Management Framework resources: https://www.nist.gov/itl/ai-risk-management-framework/ai-risk-management-framework-resources, checked 2026-11-04.