AI handles language. Rules and people keep control.

01 · Operating principles

Safety before conversion. Least authority by design. AI may understand ordinary language, retrieve approved information and draft a useful response. It does not receive open-ended authority over safety, consent, service area, availability, booking writes, prices, complaints or legal commitments.

• Transparency: people should know when a service is automated or AI-assisted, its limits and how to reach a person.
• Human accountability: neither Lead Systems Studio nor a client may blame an algorithm for a decision they authorised.
• Data minimisation: collect only what is reasonably needed to answer, route, hand over or book.
• Contestability: people can correct information, dispute a route and ask for human review.
• Auditability: material actions should retain their rule, model, decision reason, handover and provider receipt.

02 · Authority model

AI interprets: extract facts, classify intent, ask approved clarifying questions and word responses from approved knowledge.

Fixed code authorises: validate allowed values, check safety and suppression first, enforce service area, recheck live availability and make idempotent booking writes.

People remain accountable: correct exceptions, give technical or commercial judgement, handle complaints and vulnerability, and approve prices, safety wording and material changes.

03 · Approved and prohibited use

Subject to testing and client approval, AI may classify an enquiry, extract details, ask approved questions, retrieve approved content, draft channel-appropriate wording, identify a request for a person and produce a structured summary.

• No diagnosis of boiler, gas, electrical, plumbing, structural, roofing or other technical faults.
• No claim that a situation is safe, impersonation of emergency services or unauthorised attendance promise.
• No binding quotation, refund, warranty or legal commitment without an approved deterministic rule.
• No solely automated decision with legal or similarly significant effects.
• No invented reviews, scarcity, qualifications, integrations, availability or results.
• No collection of card details, passwords, authentication secrets or unnecessary sensitive data.
• No continued cadence after opt-out, booking, closure or a defined stop event.

04 · Safety and handover

A safety signal interrupts the ordinary conversation before AI responds. For immediate danger, call 999. These systems are not emergency services and do not diagnose faults or promise emergency attendance.

Production heating and plumbing rules screen approved patterns including gas smell, carbon-monoxide concerns, fire or burning smell, uncontrolled water, water near electricity, vulnerable occupants without heat in cold conditions and direct statements of danger.

A matched high-priority signal stops ordinary qualification and booking, uses only approved static wording, creates an urgent human alert and records the reason. Human handover also applies to requests for a person, complaints, vulnerability, technical or price discretion, disputed information and system failure.

05 · Data and model use

Client conversations do not become a shared training dataset by default. Client-controlled end-customer data is processed on the client's instructions.

Identifiable client data is not used to train a general-purpose or shared model without a separate written purpose, role assessment, lawful basis, privacy update, minimisation and retention controls, client opt-in and any required data-subject choice.

Only approved, dated knowledge should ground production answers. If reliable content does not answer the question, the assistant should say so and offer human help rather than improvise.

06 · Testing and incidents

Models, prompts, tools, knowledge and deterministic rulepacks are controlled production components. Material changes require representative normal, edge, adversarial, accessibility, consent, safety and failure scenarios before release.

• Safety recall on the approved release set targets 100%; a miss blocks release.
• Cross-tenant isolation, opt-out and duplicate-booking protections must pass completely.
• Invalid or uncertain output degrades to safe handover rather than a guessed action.
• Monitoring covers safety, delivery, invalid output, booking, integration, suppression, complaints, model changes and cross-tenant access.
• Critical safety or privacy incidents trigger containment, evidence preservation, notification assessment, regression tests and controlled reactivation.

07 · Related documents

Read this account alongside the Privacy policy and Client service terms. To ask for a human explanation, correct information or raise a concern, email hello@leadsystems.studio.

Questions

Email hello@leadsystems.studio with any question, correction or request.