Guide
Where to start with AI automation in a service business
A practical way to pick your first AI automation: map one enquiry workflow, score each step for effort, risk and value, and start where a person stays in control.
Short answer
Start with one enquiry workflow you already run every week. Write down each step as it happens today, score each step from 1 to 3 for effort, risk and value, and automate the step with high value and low risk first. For most service businesses that is acknowledging a new enquiry and collecting the details your team asks for anyway. Keep pricing, complaints and judgement calls with a person.
Most owners who ask where to start with AI already have a list: answer the phone, chase quotes, write the newsletter, sort the inbox. The list is the problem. Automation pays off when it takes over one job your team does the same way every week, and it causes trouble when it is spread thinly over work nobody has written down. This guide walks through a short exercise you can do in an afternoon with whoever handles your enquiries.
1. Pick one workflow you already run
Choose a workflow, not a tool. A good first candidate passes all four of these tests:
- It happens at least weekly, so you will see whether a change helped within a month.
- The same questions come up each time: service area, availability, what the job involves.
- A slow or missed response costs you something you can name, such as a lost booking or a frustrated repeat customer.
- One person on your team owns it today and can tell you how it really works.
For most service businesses this is the path from a new enquiry to a booked visit or call. Quoting, scheduling a crew and following up on unpaid invoices are common second choices.
2. Write the steps down as they happen today
Sit with the person who owns the workflow and list every step, including the awkward ones. Describe what happens, not what should happen. The worksheet below has the columns that matter later.
| Step | Who does it now | Where it happens | What they need to know | What goes wrong |
|---|---|---|---|---|
| Acknowledge a new enquiry | Owner, between jobs | Phone, web form, texts | That it arrived | Evening enquiries wait until morning |
| Collect address, home size, preferred day | Owner | Text thread | The standard questions | Two or three back-and-forth messages |
| Check the area is served | Owner | Memory | The service-area list | Edge-of-area jobs accepted by mistake |
| Give a price | Owner | Phone call | Pricing rules and judgement | Inconsistent quotes |
| Book the visit | Office manager | Shared calendar | Crew availability | Double bookings on busy weeks |
| Record the customer | Office manager | Spreadsheet | Name, address, notes | Often skipped |
3. Score each step for effort, risk and value
Give every step a score from 1 to 3 on three questions. Agree the meaning of each score before you start so the scores are comparable.
| Score | Effort to automate | Risk if it goes wrong | Value of doing it faster or more consistently |
|---|---|---|---|
| 1 | Fixed questions and rules already written down | A customer gets a slightly awkward message | Saves a few minutes a week |
| 2 | Rules exist but live in someone's head | A booking or detail needs correcting | Saves time daily or stops occasional losses |
| 3 | Needs judgement, negotiation or several systems | Money, safety, legal exposure or a lost customer | Directly affects whether enquiries become bookings |
Applied to the hypothetical cleaning company, the scores look like this:
| Step | Effort | Risk | Value | Verdict |
|---|---|---|---|---|
| Acknowledge a new enquiry | 1 | 1 | 3 | Start here |
| Collect the standard details | 1 | 1 | 2 | Start here, with the step above |
| Check the service area | 2 | 2 | 2 | Next, once the area list is written down |
| Book the visit | 2 | 2 | 2 | Later, after calendar rules are agreed |
| Record the customer | 1 | 1 | 2 | Bundle with the first step |
| Give a price | 3 | 3 | 3 | Keep with a person |
The rule of thumb: start where value is high and risk is low, and use effort to break ties. A step that scores 3 for risk stays with a person, however valuable it looks. If you want a fuller structure for thinking about AI risk, NIST publishes a voluntary AI Risk Management Framework; this exercise is a small, practical slice of the same idea.
4. Define what the first automation does, and what it never does
Write a short brief before anyone opens a tool. For the example above it might read:
- Reply to every new text, form and missed call within a few minutes, day or night.
- Ask for the address, home size and preferred day, one question at a time.
- Never quote a price, promise a date or discuss a complaint. Hand those to the owner with a short summary.
- Create or update the customer record with what was collected.
- Tell the office manager when a conversation is ready to book.
Decide how you will know it worked, using records you already keep: the time between an enquiry arriving and the first reply, and how many enquiries reach the booking step with every detail filled in. Measure both for two weeks before the change so you have something to compare against.
5. Run it beside your team before it runs alone
For the first week or two, have the system draft replies that a person approves, or send only the acknowledgement while a person handles the rest. Read every conversation. Note where it asked a question twice, misread a message or missed a handover. Fix those, then widen what it does. The guide on testing an AI agent before customers use it has a checklist for this stage.
What to leave for later
Pricing, complaints, refunds, cancellations with fees, and anything that touches health, legal or safety advice should stay with a person until your first workflow has run cleanly for a while. These are the steps where a confident but wrong answer costs the most.
Limits of this method
- The scores are judgements. Two people will score the same step differently; that discussion is useful, so do it together.
- It looks at one workflow at a time. If your real bottleneck is that work arrives in five places, connecting those channels may matter more than automating any single step.
- It does not cover your legal obligations. Texting customers, recording calls and storing their details are regulated. In Canada, CASL covers commercial electronic messages including texts, and privacy law sets rules on consent and how much you collect; in the US, federal rules on automated calls and texts depend on consent. Check the rules for your province or state with an adviser.
Sources
- AI Risk Management Framework, National Institute of Standards and Technology. Read October 21, 2026. A voluntary framework, released January 26, 2023, for managing AI risk across design, use and evaluation.
- Canada's anti-spam legislation, Innovation, Science and Economic Development Canada. Read October 21, 2026. CASL covers commercial electronic messages, including text messages.
- PIPEDA requirements in brief, Office of the Privacy Commissioner of Canada. Read October 21, 2026. The 10 fair information principles, including consent and limiting collection.
- 47 U.S. Code § 227: Restrictions on use of telephone equipment, Legal Information Institute, Cornell Law School. Read October 21, 2026. US federal rules on automated and prerecorded calls, which turn on the called party's prior express consent.
How this guide was checked
Every factual statement is listed with where it came from. "Assumption" marks examples and recommendations that are the author's own rather than sourced facts. This guide is general information, not legal or financial advice.
| Statement | Basis | Status |
|---|---|---|
| The cleaning company, its steps and its scores | Invented for illustration and labelled hypothetical | assumption |
| CASL covers commercial electronic messages, including text messages | ISED CASL page | verified |
| Canadian privacy law sets rules on consent and limiting collection | OPC, PIPEDA requirements in brief (10 principles) | verified |
| US federal rules on automated calls depend on prior express consent | 47 U.S.C. § 227 via Cornell LII | verified |
| Scoring by effort, risk and value; risk-first gating | Author's method; framed as a rule of thumb, not a standard | assumption |