← Training & SpeakingStirling AI · John Kirkwood

AI for Home-Service Teams

Less office rework. Clearer customer communication. Practical course resources to use back at the office and in the field.

For owners, dispatchers, office managers and field leads.

Course resources · Edition 3.1 · Updated November 16, 2026

Your reference after class

Return to these worked examples, prompts, review checklists and implementation steps as you apply the course to your business. Check the update record for the latest published revisions.

Topics in this guide: reliable inputs and privacy, lead intake and dispatch handoffs, technician and inspection documentation, customer communication, and evaluating a small AI pilot.

What you should be able to do

  • Separate an observed fact from an AI inference or a customer request.
  • Produce a useful handoff without inventing an appointment, diagnosis or price.
  • Review a customer draft against the original job or inspection notes.
  • Choose a task with a named reviewer and a measurable result.

One repeated task is enough to begin

Choose a task whose answer a person can check quickly. Job-note cleanup, estimate follow-up drafts and approved FAQ explanations are useful starting points. Keep dispatch decisions, prices, safety assessments and warranty commitments with the people authorized to make them.

Your job-management system remains the source of truth for schedule, scope, price and ownership. If a message conflicts with that record, flag the difference instead of letting the AI choose.

All examples below are fictional. The teaching recommends a process, not a particular software purchase or a guaranteed business result.

Two demonstrations from a service business

A missed-call inquiry

“Water stain got bigger after last night. We are outside your usual area. Can someone come tomorrow? I can text photos. Call after three.”

A useful handoff records the requested evaluation and callback preference. It flags the missing address, number and schedule confirmation. It does not diagnose storm damage, promise an appointment or predict insurance coverage. Outside-area requests go to the person who can decide under your company's policy.

Technician notes into a customer draft

“Filter dirty. Replaced customer-supplied filter. Water at pan. Drain may be blocked. Did not test full drain run. Cooling running when left. Need drain evaluation.”

Sample draft: We replaced the filter you supplied. We observed water at the pan and recommend a drain evaluation. We did not test the full drain run. Cooling was operating when we left. Contact the office to arrange the recommended evaluation.

The technician verifies that draft before it leaves the business. It does not certify a repair or replace a physical inspection.

Inspection-report language

Source note: “Moisture staining at garage ceiling. Area dry at inspection. Source not determined. Attic above inaccessible.”
Unsafe AI rewrite: “The roof leaks and needs replacement.”
Bounded draft: “Moisture staining was visible at the garage ceiling. The area was dry during the inspection. The source was not determined, and the attic above was inaccessible. Further evaluation is recommended.”

The inspector must verify the wording, limitations and recommendation against the actual inspection and applicable standards. AI does not determine a defect's cause from a brief note or replace professional judgment.

A scheduling conflict

The customer's message says Tuesday. The approved job record says Wednesday. The assistant should flag the conflict for the office. It should not silently change either record or reassure the customer that Tuesday is confirmed.

A short list, not a shopping spree

Choose a tool for the job you already have

These are starting points to evaluate, not a requirement to buy four subscriptions. My recommendation: check the software you already use, then pilot one approved general assistant. Verify the current plan, permissions and total cost before connecting customer records.

Drafting and analysis

ChatGPT Business

Try it for: turning approved field notes into customer drafts, comparing scope documents, and analyzing an anonymized export for repeated office rework.

Boundary: verify every price, appointment and technical conclusion. OpenAI says Business data is not used to train models by default. That does not remove your duties around access, retention or customer consent.

Official Business data guidance
Long documents and reusable drafts

Claude for Work

Try it for: comparing an approved service checklist with a draft SOP, finding conflicting statements in customer instructions, or producing several reading-level versions of an FAQ.

Boundary: commercial and consumer data policies differ. Anthropic's commercial policy says it does not train on inputs and outputs by default, with exceptions such as opt-in feedback. Review the actual account policy.

Official commercial data policy
A reference desk for your documents

Google Gemini Notebook / NotebookLM

Try it for: a source-grounded staff reference using approved manuals, service FAQs and company procedures. Ask for the supporting passage and open the citation.

Boundary: a citation makes checking easier, not optional. Use the correct equipment model and document revision. Check Workspace edition and source-copy behavior before uploading.

Official product overview · Workspace privacy guidance
Your existing field-service system

Native AI before extra plumbing

Example: Jobber documents AI rewrite, voice interactions and a Receptionist feature. If your team already uses it, evaluate those capabilities before wiring several outside services together.

Boundary: voice actions may create or alter records. A receptionist may book work. Test permissions, duplicate prevention and escalation with a human owner before turning those actions on. Features and availability vary.

Jobber's official capability overview · Voice and Chat documentation

Tool facts checked November 16, 2026. Suggested uses are workshop recommendations, not vendor guarantees. Keep AccuLynx or your existing job system authoritative if that is where your team schedules and manages jobs. A new assistant does not justify replacing it.

The advantage beyond a better prompt

Useful AI can live inside a repeatable process

The practical edge is a shorter path from real information to a checked action. Here are workflows to prototype with fictional data first.

01

A voice note becomes a reviewable job record

The technician dictates after parking safely. Transcription separates work completed, observations and limitations. A draft maps those facts to the job template. The technician checks it, then the approved record goes to the office. The assistant never fills a blank measurement from imagination.

Proof: compare the approved note with the original audio, correct job ID and actual work.

02

A manual becomes a cited staff reference

Create a curated reference set for one equipment family. Ask what a supplied procedure says and require page or section citations. Compare model number, revision and source passage before relying on it. Route hazardous or uncertain work to the qualified professional. Remove superseded manuals.

Proof: every answer links to the right source and admits when the answer is absent.

03

A photo set becomes a documentation check

Ask an approved visual tool to label and organize supplied job photos or flag missing checklist views. A person confirms what each image actually shows. Keep original images and distinguish observations from guesses. Do not use an AI image assessment as a structural, electrical or safety clearance.

Proof: the report preserves the originals and does not claim a hidden condition is safe.

04

An exception queue replaces silent failures

Join each incoming event to its immutable event ID, job or contact, intended owner and message delivery record. Send incomplete or conflicting cases to a review list. Prevent repeat events from making duplicate messages. Let a human decide any resend when delivery is uncertain.

Proof: inspect actual records and provider outcomes. “Workflow succeeded” alone is insufficient.

05

Customer questions become a useful content library

Group de-identified questions from your office. Match each to an approved answer. Turn the answer into a short video outline, FAQ and appointment-preparation checklist. A trade expert checks the content. Link it from the next relevant customer conversation instead of answering from scratch.

Proof: no invented customer story, guarantee or project photograph enters the library.

Different trades, different first projects

Roofing

Organize leak inquiries and roof-access notes. Preserve “caller reports” language. Flag out-of-area requests for the authorized manager.

Keep human: diagnosis, insurance representation and scope.

HVAC

Turn service notes into customer summaries and draft approved maintenance explanations.

Keep human: combustion safety, measurements and repair decisions.

Plumbing

Prepare a dispatch checklist from a fixture complaint and draft access/preparation questions.

Keep human: emergency triage, diagnosis and cost commitments.

Electrical

Index approved documentation and organize a job's recorded observations.

Keep human: energized-work decisions and code compliance judgments.

Home inspection

Check whether narrative drafts preserve inaccessible areas and inspection limitations.

Keep human: defect conclusions, standards and signed reports.

Cleaning and maintenance

Draft scope-specific arrival instructions and identify missing items on a supplied checklist.

Keep human: chemical compatibility, access security and damage claims.

Customer trust is part of the job

Security and ethics that fit daily work

Before customer information enters a tool

  • Minimize the data. A writing exercise rarely needs a name, exact address, door code, face, license plate or full contract.
  • Use approved business accounts. Verify who can access the workspace, connected apps, uploads and shared links. Enable MFA and remove former employees' access.
  • Check where copies go. Retention, model training, third-party connectors and deletion are different questions. A “not used for training” promise answers only one of them.
  • Keep source content separate from instructions. An email or PDF can contain text telling an assistant to ignore rules or send data elsewhere. Treat those instructions as untrusted. Restrict tools and require approval for consequential actions.

Before the business acts

  • Disclose appropriately. Do not impersonate a customer, trade professional or employee. Review applicable call-recording, AI disclosure and messaging requirements before a voice or outbound deployment.
  • Preserve human accountability. Never give the assistant final authority over hazards, code determinations, contracts, financing or insurance outcomes.
  • Avoid unfair shortcuts. Do not use a customer's name, accent or neighborhood to guess creditworthiness, willingness to pay or service priority. Use legitimate operational criteria and review exceptions.
  • Have a stop plan. If wrong messages or duplicate alerts appear, stop the affected automation, preserve evidence and check who actually received what before resuming.

These are risk-management practices, not legal advice. NIST's framework offers a structured approach. FTC guidance explains fake and misleading review risks.

Copyable prompts

Use fictional examples while learning. Before adding real customer information, confirm your company's approved tool, account, access and data-handling rules.

Office handoff
Use only the message below. Extract a handoff for our office. Separate reported facts, customer requests and missing information. Do not diagnose, quote a price or promise a visit. If outside our usual service area, flag it for manager review. Output: Request, Requested timing, Callback preference, Missing information, Next human decision.

Message: [paste a fictional or authorized, minimized message]
Technician summary
Using only these technician notes, draft a customer summary under 100 words. Include work completed, observations, limitations and the recommended next step. Preserve uncertainty. Do not invent measurements, code requirements, cause, price, warranty, parts or completed repairs. List anything the technician must confirm separately.

Notes: [paste approved notes]
Estimate follow-up
Draft a short, friendly follow-up about this approved estimate. Ask whether the customer has questions. Use only the supplied scope and permitted next step. Do not invent discounts, urgency, availability, financing or insurance coverage. Do not send.

Approved facts: [insert]
FAQ content
Use the approved FAQ below to draft a 150-word explainer for homeowners. Explain the service and its limitations in plain language. Mark anything unsupported as a question for our team. Do not add guarantees, certifications or customer stories.

Approved FAQ: [insert]

The review before sending

Quick self-check: A caller asks for tomorrow and the calendar has no confirmed booking. Can the draft say “See you tomorrow”? No. It may record the request and say the office will confirm availability.

Customer-facing drafts

  • Does every factual statement come from the approved source?
  • Did it preserve uncertainty and inspection limitations?
  • Are price, scope, timing and warranty terms exactly authorized?
  • Does the next step belong to the right person?
  • Did it reveal information the recipient should not see?

Connected workflows

  • Verify the actual source event and correct job or contact.
  • Check the recipient and message content.
  • Inspect provider delivery evidence rather than a green run.
  • Prevent duplicate events from sending again.
  • Keep an exception queue for missing or ambiguous evidence.

Provider delivery does not prove that a person read the message or called the customer. Define where the automation's responsibility ends.

Customer trust

Never invent reviews, customer experiences, job photos or certifications. AI may help draft your response to a real review. Use actual completed-job information only with the appropriate permission. Avoid making unsupported guarantees.

Your two-week pilot

WhenWork
Days 1–3Choose one task, name its reviewer and measure the current time.
Days 4–7Draft with AI. Review every output and record corrections.
Days 8–10Test missing facts, conflicting schedules, repeated messages and unavailable systems.
Days 11–14Compare net time saved and errors. Continue, narrow the scope or stop.

Net minutes saved = old task time − preparation − review − corrections. Example only: 12 − 3 − 4 − 1 = 4 minutes. Measure your own results.

Track wrong facts, missing next steps and duplicate sends alongside time. Set your pass/fail threshold before starting. Keep outbound sends manual until the required checks pass.

A readiness check you can reuse

Checklist selections stay only in this open page and are not submitted.

A worksheet to print or copy

Task:
Current source of truth:
Reviewer:
Information allowed in the tool:
Human decisions the tool must not make:
Baseline time:
Acceptance threshold:
Stop condition:
Review date:

Choosing tools

Check what your existing job software already supports. Trial one approved assistant for drafting before adding integrations. Compare permission controls, review-before-send, audit history, exportability, support and total cost. Product features vary by plan and change over time.

Sources and maintenance

Workshop recommendations are educational, not legal, safety or trade advice. Check company policy and applicable requirements before recording calls, transmitting customer data or automating outbound communications.

Update record

November 16, 2026, edition 3.1: Organized the tool shortlist, trade-specific workflows, security and ethics checks, and readiness checklist as a reusable course resource. Added navigation back to Stirling training. Earlier resource versions are preserved. Source links, tool capabilities and company examples are reviewed before new revisions are published.