AI Policy
V1.0 availability notice — 19 October 2026: AI features are not available in V1.0. AI-specific provisions below apply only if a feature is subsequently enabled with the required notices, permissions and controls. Existing upload and content-rights obligations continue to apply to non-AI use.
Effective date: 6 November 2026
1. Scope and available features Lighthouse may offer AI-assisted search, catalogue processing, product matching, project descriptions, planning, quantity or budget estimates and visual inspiration where those features are enabled. An AI feature is not a professional appointment, certification or autonomous authority to procure. Each feature must explain its intended use, input requirements and important limitations before a user relies on it. Future functionality is not promised by this policy.
2. User control and input rights Use accurate inputs and upload only material you have permission to process for the requested purpose. Remove unnecessary identity documents, people’s details, exact addresses, access/security information and confidential third-party data. Public portfolio content and private plans have different visibility. An upload for analysis does not authorise publication, distribution to all suppliers or use in advertisements. A user must review an output before publishing it, including an AI-drafted story, brand mapping, product tag or quotation summary. Editing assistance must not invent factual project details.
3. Inference, retrieval and training Processing an input to answer a request, generating search embeddings, keeping requested project history and training a model are distinct activities. The feature notice must identify the applicable activities and retention.
Private plans, RFQs, quotations, personal information and enterprise material must not be used to train or fine-tune a general model without separate explicit lawful authorisation and sufficient intellectual-property rights. Optional training must be off by default and must not be required for unrelated core services. Before using a third-party model, Lighthouse must verify its contractual training, retention, safety-review and transfer terms and configure them consistently with the notice. A “no training” promise cannot be made if the selected provider settings contradict it.
4. Output limitations and human review Outputs may omit materials, misread scales, confuse units, invent product attributes, match the wrong SKU or use obsolete prices. Inspect original drawings, manufacturer specifications, local conditions and confirmed quotations. Qualified human validation is required before safety-sensitive design, engineering, installation, final quantities, procurement budgets or binding orders. Do not rely solely on AI for structural calculations, fire safety, electrical loading, statutory submissions or professional sign-off. Supplier rankings and moderation flags can also be wrong. Provide a review route for consequential disputed decisions. Do not make unsupported discrimination or quality claims about a person or firm.
5. Synthetic content and prohibited use Do not fabricate finished-project photographs, professional credentials, product approvals, test reports, delivery evidence or endorsements. Do not create unlawful impersonations, non-consensual intimate imagery, child exploitation, false records or other prohibited synthetic material. Where the law requires, synthetic media must carry a prominent disclosure and technically feasible provenance. Do not remove or suppress required labels. Lighthouse must implement appropriate prevention and labelling measures for the features it operates. Routine good-faith cropping, colour correction or accessibility processing that does not materially misrepresent content is distinguished from deceptive synthetic media. A generated concept still cannot be published as a completed project.
6. Governance and incident handling The internal AI Risk Classification Schedule assigns feature owners, validation requirements and release gates. Changes to models, prompts, data sources or providers require reassessment of accuracy, privacy, IP and safety risks. Users can report an inaccurate or harmful output through [CONFIRM in-feature reporting control] or the grievance channel. Record the output/request identifier, issue and context without republishing harmful material. We may restrict a feature, correct information, investigate and notify affected users where appropriate.
7. Commercial influence and rights Label sponsored recommendations and disclose material commercial influence. AI must not silently turn paid placement into an independent professional recommendation. Input ownership remains unchanged. Output rights depend on law and relevant terms; uniqueness, copyright protection and non-infringement are not guaranteed. The IP Policy applies.
Release checks Confirm live features, providers, training configurations, storage periods, safety filters, label/provenance handling and reporting controls. The policy is a proposed operating standard until these controls are implemented and approved.
8. Feature-by-feature AI governance Each enabled feature must identify its purpose and expected output: extracting catalogue fields, describing an approved project, matching materials, estimating quantities, translating text or assisting a procurement plan. A permission to run one feature is not permission to use every uploaded file for every AI purpose. A user must be told when a material task uses an external AI provider before the relevant upload or activation. Public catalogue generation requires verification against an authoritative manufacturer specification, including model code, dimensions, finish, units, compatibility and safety claims. Confidence scores, fluent language and visually convincing renders are not evidence. If a source is absent or contradictory, the result must say that it is unverified and must not invent certification, stock, prices, project credits or product performance.
9. Older content, retrieval and recommendations An older project may be resurfaced only within the rights and visibility originally granted. Preserve the actual completion year, creator credits and the distinction between original publication and later editorial updates. Do not imply the firm currently offers a service, or a supplier stocks a product, merely because they appeared in an old project. Retrieval systems must distinguish approved public content from private drawings, RFQs, quotations and enterprise workspaces. Access permissions must be enforced before retrieval, not merely hidden in the final answer. A removed or restricted item must be excluded from subsequent recommendations and retrieval indexes as the removal propagates. A synthetic summary must not become a loophole for continuing to display protected material.
10. Human authority and consequential decisions AI may propose quantities or alternatives; the authorised buyer or qualified professional must approve them before an order or site action. Approval must identify the input revision, assumptions, product variants, units and reviewer. An automated recommendation must not place an order, change a specification, spend a budget or contact a supplier without an explicitly authorised action and confirmation.
AI must not make a final adverse verification, account-ban, professional-competence or eligibility decision without the applicable human review. Users may challenge materially incorrect source data and ask for an explanation of the factors relevant to the decision, subject to lawful security and third-party protections. Human reviewers must have actual power to change the outcome.
11. Training, evaluation and service-provider controls Operational inference, safety monitoring, product evaluation, model training and fine-tuning are separate uses. Model improvement involving private or personal material requires an identified lawful basis and any necessary separate permission. An opt-in must explain data categories, recipients, retention, withdrawal effects and whether trained parameters can be individually edited; it must not promise technical erasure that cannot be delivered. Provider terms and settings must be verified for the actual account tier, endpoint, region and retention mode. A provider’s general marketing statement is insufficient. Where a provider retains prompts for abuse monitoring, the notice must describe the actual arrangement. Staff must not paste customer plans into personal AI accounts or use production records in demonstration datasets.
12. Safety, abuse and release control We may decline unsafe prompts, redact unnecessary personal data, rate-limit automated use and suspend a feature that creates material harm. A person must not bypass these controls, fabricate professional approvals, impersonate a real person, generate fraudulent reviews or remove required provenance labels. Before release or a material model change, evaluate representative construction cases, Indian languages, unit conversion, unsupported claims, bias, leakage and prompt injection. Keep versioned results, reviewer decisions, rollback arrangements and an incident contact. These are approval requirements for implementation, not claims that every proposed feature has already passed.
13. International AI requirements For an EU/EEA release, classify Lighthouse’s actual role as provider, deployer, importer or distributor under the applicable AI rules and assess the use case separately from the internal risk score. A general chatbot is not automatically a legally high-risk system; using AI as a regulated safety component can change the assessment. Apply required interaction and synthetic-content transparency and staff literacy obligations when applicable. For all markets, local privacy, consumer, discrimination, professional and product-safety rules continue to apply to AI-assisted results. A global launch must be approved for the specific feature, destination and use case; an “AI disclaimer” does not authorise otherwise prohibited activity.
14. Input quality and defined operating envelope Before using a construction-related AI feature, the user must identify the available input revision and material assumptions. Where scale, units, product identity, room dimensions or
completion status cannot be reliably established, the feature must request clarification or identify the output as incomplete. An estimated quantity must distinguish measured quantity, assumed wastage, optional accessories and excluded work rather than hide those items in a single apparently exact number. Lighthouse will limit each feature to the functions approved for release. A conversational interface must not silently expand a request to search unrelated private workspaces, publish content, communicate with dealers or spend money. Any tool that can affect an order or account must operate within the separately authorised transaction workflow.
15. Approval before consequential actions An approval screen must identify what action will occur, the affected product or record, the quantity or amount, the recipient and any irreversible consequence. Approval of an estimate is not approval of a purchase. Approval of a public story is not approval of an advertising campaign. A user may correct or cancel a proposed action before the binding confirmation step. Where approval is delegated to an organisational user, the recorded authority must cover the action and budget. A model-generated statement that “the customer approved” cannot replace authenticated approval. Repeated calls or retries must not create duplicate orders, duplicate messages or repeated charges.
16. Testing errors and correcting affected records An error report should identify the source file or output reference, the mistaken field and any consequential use known to the reporter. Lighthouse will distinguish a cosmetic error from a safety issue, privacy exposure, wrong product specification, incorrect commercial action or systemic model failure. Material errors may require disabling the affected feature, correcting published records and contacting affected users where justified or required. Corrections must address downstream catalogue fields, search indexes, recommendations and approved exports within Lighthouse's control. A corrected answer must not leave an erroneous product specification live elsewhere. Users who already downloaded an output should receive a correction notice where the impact and available records justify it. An investigation may preserve a restricted copy without continuing ordinary commercial use.
17. Feedback, human review and provider access Submitting a thumbs-up, thumbs-down or support complaint authorises handling that feedback for the disclosed support purpose; it does not automatically opt the underlying private files into model training. Where an optional evaluation programme uses user material, Lighthouse must separately describe the selection, reviewers, retention, permissions and withdrawal process. Authorised reviewers may access only material necessary for the stated task. Access to confidential plans must be logged where the approved security design requires it. Vendor troubleshooting must use redacted or synthetic examples where feasible, and customer files must not be supplied to a provider beyond the applicable contract and notice.
18. Output reuse, third-party claims and export Users must retain relevant AI labels and limitations when presenting outputs to clients, suppliers or a site team. Removing the label does not turn an estimate into a certified professional document. Where the user substantially revises an output, the resulting document should accurately identify the human contribution and any necessary professional approval. A report of copied or infringing output will be handled under the IP and moderation procedures. Lighthouse may ask for the output identifier, allegedly copied work and explanation of the conflict; a complainant need not publish private source files to the whole community. Output similarity, by itself, does not establish ownership or infringement.
19. Feature retirement and preservation of user choices If an AI feature is withdrawn, Lighthouse will explain the effect on saved work and available export where relevant to a paid commitment. Existing privacy choices and training restrictions continue to apply to retained data. Switching models, providers or regions does not permit broader use merely because an earlier feature was enabled.
20. Optional import and AI consent controls For media-import assistance, obtain the required explicit permission before personal data is sent to the identified AI provider. Explain what is shared and why at the point of activation. Keep that decision separate from publication and optional training. Manual use must remain available when optional AI consent or source access is declined. Apply P04 section 18 and the separate website/Instagram release gates in I05; an approved upload feature does not authorise a third-party import method.