Questions 6–12: Can you control what the system does?

6. Are booking, pricing and approval rules explicit?

• A — 0: We expect the system to work out what it can promise.
• B — 1: We have a script, but important exceptions remain undocumented.
• C — 2: Permitted actions and escalation rules cover service area, job type, customer authority, availability and pricing boundaries.
• D — 3: Those rules are implemented and tested; the system cannot confirm an appointment, price or change outside its authority. Where approval is required, it remains a request until a person approves it.

Evidence to collect: A permissions matrix: may collect / may draft / may request / may commit / must escalate.

For commercial HVAC, include site access, equipment, contractual priorities and engineer allocation. Airco’s published coordinator role illustrates these operational dependencies, including scheduling, logistics, quotations, asset records and client-specific requirements; it does not establish an industry-wide workload benchmark. (airco.co.uk)

7. Has the exact integration been authorised and demonstrated?

• A — 0: We only have a claim that the supplier “integrates with” our software.
• B — 1: A connector or API exists, but our required actions, permissions, costs or terms are unverified.
• C — 2: We have confirmed the required endpoints or supported actions, subscription access, vendor conditions and responsibility for failures.
• D — 3: The permitted workflow has passed an end-to-end non-production test, including failed requests, retries, duplicate prevention and reconciliation. If there is no integration, the controlled manual handoff has been tested instead.

Evidence to collect: A witnessed demonstration of your workflow—not a different calendar or a generic CRM demo.

Current procurement warning: Joblogic’s API terms, updated 22 September 2026, place approval and scope conditions on integrations, require disclosure of AI integrations and restrict automated record changes subject to specified conditions. Technical API availability is therefore not, by itself, permission for an agency’s proposed implementation. Confirm the terms applying to your account and workflow. (joblogic.com)

8. Is there a tested boundary around safety-sensitive and technical enquiries?

• A — 0: The system may diagnose faults, reassure callers about safety or improvise emergency advice.
• B — 1: It has been told to “escalate emergencies”, but the approved response and destination are unclear.
• C — 2: A competent technical lead has approved workflow boundaries, relevant local emergency instructions and escalation routes.
• D — 3: These controls have passed scenario tests. Urgent safety instructions are not delayed by ordinary booking or data collection; failed transfers have an approved fallback; technical judgement remains with appropriately qualified people.

Evidence to collect: An approved escalation procedure covering your geography, services and operating hours.

HSE guidance distinguishes immediate action for suspected gas escapes from investigation or repair by Gas Safe registered engineers. Our procurement rule is consequently to exclude autonomous safety diagnosis and improvised reassurance—not to suggest that a generic AI prompt constitutes a safety system. (hse.gov.uk)

9. Can you explain and control the personal-data lifecycle?

• A — 0: We cannot explain where customer information, recordings or transcripts go.
• B — 1: The supplier says it is “GDPR compliant”, but we have not reviewed the arrangement.
• C — 2: We have documented purposes, lawful bases, organisational roles, suppliers, access, retention, training use, transfers and individual-rights handling; required actions remain open.
• D — 3: The accountable lead has completed the review with appropriate specialist input, required assessments and contracts are in place, and relevant access, retention and deletion controls have been checked.

Evidence to collect: A data-flow diagram and supplier data-protection pack, including subprocessors and incident responsibilities. Determine whether a DPIA is required; do not assume that either every AI use automatically requires one or that a small business is exempt. ICO guidance makes likely high risk the general trigger. (ico.org.uk)

Choose and document the lawful basis before processing, and explain the processing transparently. These are separate from deciding whether the product is commercially useful. (ico.org.uk)

10. Have you separated service communications from marketing?

• A — 0: We intend to contact everyone in the database without reviewing purpose or permissions.
• B — 1: We assume any previous enquiry or customer relationship permits automated follow-up.
• C — 2: We have classified each message and channel, reviewed the applicable rules, and identified permission and suppression requirements.
• D — 3: Approved templates, recipient rules and opt-outs have been tested. Any unapproved outbound marketing—including voice campaigns—is disabled; if the pilot is inbound-only, that restriction is documented and enforced.

Evidence to collect: A message register covering appointment confirmations, quotation chasing, service reminders, renewal invitations and promotional offers.

The ICO distinguishes purely administrative service messages from promotional communications; adding marketing can change the classification. Do not classify every maintenance reminder or quotation follow-up by its label alone. (ico.org.uk)

Its automated-marketing-call guidance describes autodialled recorded messages and requires specific consent for that form of marketing. Do not assume an interactive AI outbound caller can simply be treated as a live human call: obtain an assessment of the proposed implementation before enabling it. (ico.org.uk)

11. Does the office team have ownership, time and a usable fallback?

• A — 0: Nobody internally will own the workflow after launch.
• B — 1: The owner or supplier is enthusiastic, but the people handling exceptions have not been involved.
• C — 2: A named owner and deputy have agreed review time, training, escalation responsibilities and a manual fallback.
• D — 3: Relevant staff have practised reviewing outputs, correcting records, handling exceptions and switching back; monitoring time is included in the plan.

Evidence to collect: A short operating procedure and a rota showing who responds when the system cannot complete an action.

Buyer recommendation: Treat the office manager’s objections as test cases. If nobody has time to supervise the pilot, do not describe it as supervised.

ICO governance guidance places responsibility with accountable management and emphasises appropriate roles, resources and training; purchasing a supplier service does not remove those governance questions. (ico.org.uk)

12. Has the proposed workflow survived realistic failure testing?

• A — 0: We have only seen a polished demonstration.
• B — 1: We have tested straightforward examples with cooperative users.
• C — 2: We have an agreed test pack covering exceptions, outages, privacy and unauthorised requests, but material issues remain unresolved or untested.
• D — 3: Representative tests have passed, launch-blocking defects are resolved, rollback has been demonstrated and monitoring responsibilities are agreed.

Evidence to collect: Test inputs, expected outcomes, actual results, severity, fixes and retest evidence.

Include attempts to make the system ignore restrictions, disclose another customer’s information or take an unauthorised action. The NCSC advises using deterministic safeguards outside the language model to constrain actions, rather than relying only on instructions to resist malicious input. Passing a finite test pack does not establish that all such attacks have been eliminated. (ncsc.gov.uk)