Scorecard
Score three supplier plans against the same evidence and launch gates.
Compare the plan.
Keep the evidence beside it.
Score the supplier plus the proposed workflow. No named supplier has been rated. Gates override every total.
| Assessment criterion | Weight | |||
|---|---|---|---|---|
| Launch gates · every applicable gate must pass | ||||
G1. Safety and authorityCompetent review of emergency handling; approved scripts; explicit limits on diagnosis, prices and technical commitments Stop condition: The system invents safety advice, delays the approved emergency route or makes an unauthorised commitment | Gate | |||
G2. Booking and record integrityDemonstrated permissions, customer/site matching, availability validation and duplicate prevention appropriate to the permitted actions Stop condition: The system confirms an appointment it has not successfully recorded, changes the wrong record or bypasses booking restrictions | Gate | |||
G3. Personal-data governanceDocumented purposes and lawful bases; processing-role assessment; appropriate contracts; data-flow map; retention rules; DPIA decision and transfer assessment Stop condition: Real data would enter an unexplained processor, unauthorised training process or unresolved high-risk processing arrangement | Gate | |||
G4. Human control and continuityWorking diversion or shutdown control; named exception owner; tested failure routes; clear customer messaging when humans are unavailable Stop condition: Customers become trapped in the AI interaction, failures go unnoticed or staff cannot stop the workflow | Gate | |||
G5. Spend and exit controlWritten charges, usage limits, alerts and fallback behaviour; access and export arrangements; a bounded pilot commitment Stop condition: Uncapped exposure, undisclosed material charges or no workable route to recover operations and records | Gate | |||
G6. Outbound communicationsApproved purpose, audience, channel, permission rules and suppression logic for each message sequence Stop condition: The supplier treats an inbound enquiry as unrestricted permission for future marketing | Gate | |||
| Weighted criteria · unknown remains visible | ||||
| 18 | ||||
| 16 | ||||
| 15 | ||||
| 12 | ||||
| 10 | ||||
| 9 | ||||
| 7 | ||||
| 5 | ||||
| 5 | ||||
| 3 | ||||
A canned demo cannot earn more than 2 for operational performance. A sandbox test alone cannot earn more than 3 for live reliability or economic results. A material control failure also fails its gate and caps the affected row at 1 until remediation and retesting.
Below 60 · 10 criteria unknown
Evidence references: 0 / 10 criteria.
Unresolved gates: G1, G2, G3, G4, G5, G6.
Below 60 · 10 criteria unknown
Evidence references: 0 / 10 criteria.
Unresolved gates: G1, G2, G3, G4, G5, G6.
Below 60 · 10 criteria unknown
Evidence references: 0 / 10 criteria.
Unresolved gates: G1, G2, G3, G4, G5, G6.
Weights and thresholds are editorial purchasing controls, not industry benchmarks or certification. Keep unknown evidence visible. Read the decision rules →
Inside this section.
How should a heating or air-conditioning company judge an AI agency or automation proposal?
Choose the supplier that can demonstrate a controlled improvement in your business—not the most convincing conversation with a chatbot.
What exactly are you scoring?
Score the supplier-plus-plan, not the supplier’s brand.
Why must the comparison include existing software and human answering?
Do not commission an agency until it has explained what your existing platform cannot do.
Which failures should block a live pilot?
A failed or unresolved gate means no live customer traffic or personal-data processing within the affected scope.
How does the 100-point score work?
Rate each criterion from 0 to 5, then multiply by its weight divided by five.
What should good and bad look like for each criterion?
Business value, safety, integration and data governance carry 61 of the 100 points.
What evidence should an AI agency provide?
Request evidence that can be checked against your systems and operating rules.
What must the privacy and security review actually check?
Review the complete processing arrangement, not a compliance badge.
When does quotation follow-up or a maintenance reminder become marketing?
Classify the actual message and audience before enabling the sequence.
Which stress tests should an HVAC AI receptionist pass?
Test operational behaviour, not just whether the voice sounds natural.
How should the pilot prove an operational improvement?
Use buyer-owned records and a comparison that preserves normal customer service.
How should you calculate full-year cost and break-even?
Build the calculation from your enquiry funnel and job contribution, not a generic missed-call revenue claim.
What should the contract and handover protect?
Negotiate the ability to operate, verify and leave—not just the right to cancel a subscription.
How should the final score change the purchasing decision?
Use stage-specific decisions rather than a single “approved supplier” label.
What are the research limits?
The research covered official product documentation, regulators, safety and engineering bodies, trade reporting, recruitment material, supplier case studies, podcasts and YouTube descriptions.
Read →Sources
Review the source material cited in this section.
Read →Research quality & limitations
Editorial self-assessment from the research pack; these are not independent ratings or verified outcomes.