CENACE defines the nodal price as the sum of its energy, losses and congestion components. That gives you a useful arithmetic check. It does not, by itself, answer questions about the observations included in your file or the evidence supporting your analysis. CENACE data dictionary, page 2.

Here is an independent review method you can use with your own records.

Define the review before opening the spreadsheet

Write down the intended use, system, market, node, operating dates and the physical rows you expect to review. Record the conditions that would prevent that use. Preserve the original export unchanged, alongside its retrieval date and source reference.

CENACE’s SW-PML documentation identifies fields for system, market process, node, operating date, operating hour and price components. Map those meanings against the columns in your actual export; a third-party export may rename or add fields. SW-PML technical manual, pages 4–5.

Keep node identifiers as text. Record the original hour label separately from any derived timestamp. If you cannot substantiate a time conversion or publication version, record that uncertainty before using the data for a purpose that depends on it.

Keep population and arithmetic checks separate

Declare whether you are reviewing the whole export, a selected sample or identified batches. Account for exclusions and rows outside the selected sample. Retain the reasons for missing observations and duplicates instead of silently dropping them.

For a row with valid numeric inputs, calculate:

residual = PML − (energy + losses + congestion)

Compare the absolute residual with a tolerance that you have justified and documented. Keep the signed residual so that an exception can be investigated. Missing or invalid inputs need an explicit NOT CHECKED outcome; replacing them with zero can hide an unresolved problem.

Choose the tolerance for the intended review and available source evidence. Changing it merely to remove red cells does not resolve the underlying question.

What a bounded real-data check showed

Validation of the kit used 24 cached nodal PML observations for one SIN/MDA node and operating date. They were selected from a 960-record cached run; the other 936 records were outside the declared review scope.

The diagnostic comparison used a tolerance of 0.00000001 MXN/MWh to absorb binary floating-point noise. That threshold was not a source-rounding allowance or an acceptance standard for other work. The five exceptions remained visible, and source publication/version evidence remained unknown.

LibreOffice and Excel for the web returned 19 PASS and five EXCEPTION results. Excel recalculated the test copy after its cached formula results were cleared. A separate blank-tolerance copy returned 24 NOT CHECKED results. The calculation tests did not establish that the selected observations were suitable for the intended analysis. Published kit and tested scope.

Leave a decision another reviewer can understand

Keep a short record with these fields:

A HOLD should say what would resolve it. If you accept a limited use, identify that use explicitly so the decision is not mistaken for approval of the entire dataset.

Optional review templates

I offer the independent CENACE Export Review Kit for US$19: an English-language nine-page PDF guide, an eight-sheet XLSX workbook and fictional practice data. The supplied workbook supports up to 100 selected observations. It is a manual review toolkit; it does not retrieve data or certify an export.

The workbook has bounded LibreOffice and Excel-web calculation evidence. Desktop Excel, interactive Excel-web editing, dropdown/protection enforcement and edited-file save/reopen remain unverified. Read the full compatibility notes before choosing it for your workflow. The product and this review method are not affiliated with or endorsed by CENACE.

The danger isn’t that automations break. It’s that they break quietly. A booking confirmation stops sending, nobody notices for three weeks, and the first sign is a customer who didn’t turn up because they never got the details.

Here’s why it happens, how to catch it, and what I include in ongoing care.

Why automations break

1. An app changes something. Software companies update their products all the time. A field gets renamed, a feature moves, a connection gets retired. Usually nothing happens. Occasionally the automation that relied on it stops.

2. A login expires or a password changes. Automations connect to your apps with saved access. Change a password, remove a user or let a token expire, and the connection drops.

3. Your business changes. You add a service, change your prices, hire someone, switch from one booking tool to another. The automation still follows the old rules.

4. Unusual input. A customer types their phone number with letters in it, attaches a huge file, or writes in another language. Every automation meets something it wasn’t built for eventually.

5. Limits and bills. You go over your automation platform’s monthly allowance, a card on file expires, or an AI service hits a usage limit.

6. AI drift. When AI models are updated, their answers can change slightly. A prompt that sorted inquiries perfectly might start putting a few in the wrong category.

How to catch problems early

Even without anyone’s help, you can do a lot:

• Turn on error emails in your automation platform and make sure they go to an inbox someone reads.
• Do a monthly test. Fill in your own form, book a test appointment and send yourself a test invoice. Check that everything that should happen does.
• Watch the numbers. If you normally get ten inquiries a week and suddenly get zero, assume something is broken before assuming business is slow.
• Keep a simple list of your automations, what each does, and which apps and logins it depends on.

What my care plan covers

The care plan on my services page is $450 CAD a month and covers:

• Monitoring and fixes. I watch for errors and fix broken automations, usually before you notice.
• Up to 5 hours of improvements each month. Adjusting to a new service, tweaking AI prompts, adding a step, or building a small new automation.
• A reply within 1 business day.
• Cancel any month.

You don’t need it for every automation. Every build already includes 30 days of fixes after hand-over, and simple automations often run happily for a long time with just the monthly self-check above.

A care plan makes sense when:

• an automation is part of how you make money (bookings, inquiries, payments),
• you have several automations depending on each other,
• you’d rather not be the person who has to notice and fix things,
• you expect to keep improving your setup month by month.

The cheapest fix is the one you don’t need

The best protection is building things simply in the first place: fewer steps, fewer tools, clear error alerts, and a written guide that says what each automation does. That’s how I hand over every build, care plan or not.

If you have automations that someone else built and nobody is looking after, or you’re not sure what you have running, book a free 20-minute discovery call and we’ll go through them. For more on choosing tools that are cheap to run and look after, see Zapier, Make, n8n or custom.

This is a good fit for automation, as long as it’s set up so that emergencies always reach a person straight away.

This post walks through a typical setup. It’s a sample flow, not a specific client’s.

The flow

1. One way in

Tenants report issues through a single form or a dedicated number they can text. The form asks for unit, a description, photos, and the best time to access the unit. Texts and emails that arrive elsewhere get forwarded to the same place.

One way in is half the battle. It means nothing gets lost in someone’s personal phone.

2. AI sorts the request

An AI model reads the description (and can describe the photos) and sorts each request:

• Emergency: water leak, no heat in winter, gas smell, electrical hazard, a lock-out, a security problem.
• Urgent: something that affects daily living but isn’t dangerous, like a broken fridge or one toilet out of two not working.
• Routine: a dripping tap, a sticky door, a burnt-out bulb in a common area.

It also picks the likely trade (plumber, electrician, HVAC, appliance, general handyman) and pulls out the unit and access details.

3. Emergencies go to a person, immediately

This is the most important rule. Anything the AI marks as an emergency, or isn’t sure about, triggers a phone call or urgent text to whoever is on call. The tenant also gets an immediate message with the emergency contact number and basic safety steps you’ve approved in advance, such as how to shut off the water.

AI never decides that something is not an emergency on its own if the tenant says it is. In Calgary, no heat in January is not a routine ticket.

4. Routine and urgent requests get routed

For non-emergencies, the automation:

• creates a work order with all the details,
• sends it to your preferred trade for that kind of job, or to you for approval first,
• tells the tenant it’s been received and roughly what happens next.

5. Everyone stays updated

When the trade books a time, the tenant is told. When the job is marked done, the tenant gets a short “is everything working?” message. If they say no, the work order reopens.

6. The record keeps itself

Every request, photo, message and invoice ends up attached to the unit. That’s useful at move-out, for insurance, for spotting units with repeat problems, and for knowing what each property actually costs to maintain.

Tools

Property management software like Buildium, AppFolio or Rentec has maintenance request features built in, and if you already use one, start there. The automation layer adds the AI sorting, the on-call alerts and the tenant updates in plain language.

If you manage a smaller portfolio and use email and spreadsheets, the whole flow can be built with a form, an automation platform, an AI model and a shared spreadsheet or Airtable base.

A note on the rules

Alberta’s residential tenancy rules set out landlord and tenant responsibilities for repairs and entering a unit. An automation should support those rules, not shortcut them. For example, entry notices should still go out the way the rules require. If you’re not sure what applies to you, check the Government of Alberta’s residential tenancy information or get advice.

Tenant information is personal information too, so the privacy checklist in my PIPA post applies here.

Is it worth it?

For a landlord with a few units, a simple form and a good spreadsheet may be enough. Once you’re handling more than a few requests a week, or you have staff on call, the time saved and the emergencies caught faster usually justify a build.

If you manage rentals in Calgary or anywhere in Alberta and want to talk it through, book a free 20-minute discovery call.

Here’s how I think about choosing between them, and when none of them is the right answer.

A note on prices: these tools change their plans often and mostly charge in US dollars, so I haven’t listed exact prices. Check each tool’s pricing page, and remember the exchange rate when you compare them with Canadian costs.

How they charge

• Zapier charges by the number of tasks, which are the individual actions your automations perform. Triggers aren’t counted, so an automation with a trigger and four actions that runs 100 times a month uses about 400 tasks.
• Make charges by operations or credits, which are similar: each module that runs counts. It’s usually cheaper per action than Zapier, but harder to estimate up front.
• n8n has a free Community edition you can run on your own server, so you pay for hosting rather than per action, or you can use n8n’s paid cloud service. It isn’t open source in the strict sense: its Sustainable Use License allows self-hosting for your own business’s internal use but restricts things like reselling hosted access, which is fine for most small businesses.

The practical upshot: a low-volume automation is cheap on any of them. A high-volume one (hundreds of runs a day) can get expensive on per-task pricing, and that’s where Make or self-hosted n8n start to make sense.

Which one when

Zapier: easiest to start with

• Connects to the most apps out of the box.
• Simplest to set up and understand.
• Good for simple, low- to medium-volume automations.
• Gets expensive as volume and complexity grow.

Pick it if you want to build simple automations yourself and you value ease of use over cost.

Make: more power per dollar

• Visual builder that handles branching, loops and data reshaping well.
• Usually cheaper than Zapier at the same volume.
• Steeper learning curve.

Pick it if your automations have several branches or you run them a lot.

n8n: most control

• Can be self-hosted, so you control where data lives. That’s useful for privacy, including Alberta’s PIPA requirements around storing data outside Canada.
• Strong AI features for building workflows that use language models.
• Self-hosting means someone has to keep the server updated and backed up.

Pick it if you have higher volumes or data you’d rather keep on a server you control, and you have someone to look after it.

Custom code: when the platforms don’t fit

A small custom script or app is worth it when:

• the volume is high enough that per-task pricing is a real monthly cost,
• you need something the platforms can’t do cleanly,
• the logic is complicated enough that a visual builder becomes a tangle.

The trade-off is that someone needs to maintain it. For most small businesses, a platform is the right starting point.

What to skip

• Paying for a big plan “in case”. Start on the smallest plan that fits, watch your usage for a month, then adjust.
• Automating every small thing. Each automation is something that can break. Automate the tasks that cost real time or money.
• Tools you don’t own. Whoever builds your automations, make sure they run in accounts you own and you can remove outside access at any time. That’s how I set up every build: software subscriptions are billed to you directly, so you keep ownership.

The cost that matters most

The platform bill is usually the smallest part. The bigger costs are the time it takes to build automations properly, and the cost of ones that break quietly and nobody notices. I wrote about that in why automations break.

If you’re stuck choosing, or you’ve built something that’s getting expensive, book a free 20-minute discovery call. I work with all three platforms and custom code, and I’ll recommend whichever is cheapest to run for what you actually need.

This checklist takes about five minutes. Answer each question yes or no, honestly. It’ll tell you whether you’re ready to automate now, and if not, what to sort out first.

Want it on paper? There’s a printable version of this checklist you can fill in by hand or save as a PDF.

The questions

1. Can you name the one task that eats the most of your week? Not “admin” in general, but one specific job: answering inquiries, writing quotes, chasing invoices, scheduling.

2. Do you do that task the same way most of the time? Automation needs a pattern. If every case is different, it may be a job for a person, or for better templates first.

3. Could you explain it in ten steps or fewer to a new employee? If you can write it down, it can probably be automated. If you can’t, write it down first. That alone often saves time.

4. Does the task happen at least a few times a week? Automating something you do twice a year rarely pays off. Automating something you do twenty times a week usually does.

5. Is the information it needs already digital? Forms, emails, spreadsheets, apps. If it’s on paper or sticky notes, the first step is getting it into a tool.

6. Do your main tools have logins you control? Your email, calendar, invoicing and booking tools should be in accounts you own, not a former employee’s personal account.

7. Do you know roughly what the task costs you? In hours, or in lost customers. Even a rough number (“about five hours a week”, “we miss maybe ten calls a week”) lets you judge whether a fix is worth paying for.

8. Would a mistake be easy to catch and fix? Start with tasks where an error is a minor annoyance (a reminder sent twice), not a serious problem (a wrong price on a signed contract). Build trust before automating high-stakes steps.

9. Is someone available to check the results for the first few weeks? New automations need a little attention while they settle in. Ten minutes a day for a couple of weeks is usually enough.

10. Are you willing to change the process a little? The best automations often come from simplifying the process at the same time. If the current way is untouchable, the gains will be smaller.

Your score

• 8 to 10 yes: You’re ready. Pick the task from question 1 and automate it.
• 5 to 7 yes: You’re close. Fix the “no” answers that are quick (usually 3, 5 and 6), then start.
• 0 to 4 yes: Do the groundwork first. Write the process down, move it into digital tools, and track how much time it takes for a couple of weeks. You’ll get value from that before any automation.

The most common gaps

From the questions above, the three that hold most businesses back are:

• No written process (question 3). Fix it by writing the steps down the next time you do the task.
• Information on paper (question 5). Fix it with a simple form or spreadsheet.
• No idea of the cost (question 7). Fix it by counting for one week. My post on the cost of missed calls shows one way to do the sum.

Next step

If you scored 5 or more and want help picking what to automate first, that’s exactly what the free discovery call is for. Book a free 20-minute call and bring your answers. If you’d like a written plan with tools, costs and build order, the $149 strategy session covers that, and the fee comes off any build you order within 30 days.

AI won’t cook or serve. But it’s good at that layer of admin, and in a business where margins are thin, every hour back matters.

Here’s where I’d start for a restaurant, cafe or food truck in Calgary.

1. Catering and group booking inquiries

Catering inquiries are high-value and slow to answer. Each one needs the same details (date, headcount, budget, dietary needs, pickup or delivery) and usually a few emails back and forth before you can quote.

What helps:

• A short catering form on your website that asks for all of that up front.
• An automatic reply that thanks them and confirms what happens next.
• AI that reads the request and drafts a quote from your catering menu and pricing, for you to check and send.
• A follow-up if they haven’t confirmed a few days later.

Calgary has busy catering seasons: Stampede in July, holiday parties in December, and graduation and wedding season in spring and summer. Those are the weeks a fast, complete first reply wins the booking.

2. Keeping the menu the same everywhere

Your menu probably lives in five places: your website, Google Business Profile, delivery apps, social media and a printed copy. Change a price or sell out of a dish and four of them are wrong.

What helps: keep one master menu (a spreadsheet works fine) and have changes flow out from there to the places that support it. Where a platform can’t be updated automatically, the automation at least produces a checklist of what to update by hand.

AI can also help write menu descriptions, translate the menu for customers who read another language, and flag likely allergens for you to double-check. Allergen information must always be confirmed by a person, never left to AI.

3. Answering the same questions

“Are you open on the holiday Monday?” “Do you have gluten-free options?” “Can I bring a cake?” “Is there parking?”

What helps: a simple FAQ that an AI assistant can use to answer messages on your website, with anything it isn’t sure about passed to you. Keep the FAQ short and current. When the answer changes, change it in one place.

4. Staff schedules and shift swaps

Scheduling tools like 7shifts, Homebase and Deputy already handle a lot of this, and if you’re still doing schedules in a group chat, moving to one of them is the first step.

What helps on top: automations that remind staff of shifts, collect availability in one place, and notify you only when a swap needs approval. AI can draft a first version of next week’s schedule from availability and expected busy times, which you then adjust.

5. Reviews

Restaurants get more reviews than most businesses, and replies matter. Use AI to draft replies in your voice, and reply to negative ones yourself. I wrote a whole post on replying to Google reviews in your own voice.

What to leave alone

• Allergen and food safety decisions. A person checks every time.
• Refunds and complaints about illness. Take these offline and handle them personally.
• Your social media voice. AI can help with drafts and scheduling, but regulars can tell when a restaurant’s posts stop sounding like the people who run it.

Where to start

If catering is part of your business, start there. It’s the highest-value inquiry you get and the one most often left waiting. If not, start with the FAQ answers, because they save time every single day.

If you run a restaurant, cafe or food truck in Calgary and want help setting any of this up, book a free 20-minute discovery call. I’ll tell you honestly which of these is worth paying for and which you can do yourself with tools you already have.

Automating the nudge takes the awkwardness out. It goes out on time, every time, in a friendly tone, and you only step in when someone replies.

Part 1: Quotes

Most quotes that go unanswered aren’t a no. The customer got busy, forgot, or has a question they didn’t ask. A short follow-up brings a lot of them back.

A follow-up schedule that isn’t pushy

• Day 2 or 3: a check-in that offers help.
• Day 7: a short reminder, with an easy way to say no.
• Day 14: a final note that closes the loop.

Then stop. If someone replies at any point, the automation stops and the conversation comes to you.

Templates you can copy

Day 3

Hi Jordan, just checking you got the quote for the fence repair. Happy to answer any questions or adjust anything. If it’s easier, reply with a good time and I’ll give you a call.

Day 7

Hi Jordan, following up on the fence repair quote. If the timing or the price doesn’t work, no problem, just let me know and I’ll close it off. If you’d like to go ahead, reply here and we’ll get you booked in.

Day 14

Hi Jordan, I’ll close off the fence repair quote for now so it doesn’t clutter your inbox. If you want to pick it up again later, just reply to this email.

The “it’s fine to say no” line matters. It gets you answers, and an answer, even a no, is worth more than silence. Over time those answers tell you why quotes are lost.

Make the quotes themselves faster

Follow-ups only help if the quote went out quickly in the first place. If writing quotes is the slow part, a good template saves the most time. My Service Quote Kit is a set of editable proposal documents and a matching slide deck built for exactly that. You can also have AI turn your job notes into a first draft, as I described in five jobs a small business can hand to AI.

Part 2: Invoices

Start with what you already have

QuickBooks, Wave, FreshBooks, Xero and Square all have some form of automatic payment reminder. Check your settings today and turn it on. It’s free and takes five minutes.

What those built-in reminders usually can’t do is change the tone as an invoice gets older, skip a customer you’ve already spoken to, or tell you which invoices need a phone call. That’s where a small custom automation helps.

A reminder schedule

• 3 days before due: a friendly heads-up with the payment link.
• Due date: a short “this is due today”.
• 7 days overdue: a polite reminder, asking if anything’s wrong with the invoice.
• 21 days overdue: a firmer note, and a flag to you to pick up the phone.

Templates

7 days overdue

Hi Priya, a quick reminder that invoice 1042 for $640 was due last week. Here’s the payment link: [link]. If anything on the invoice doesn’t look right, just reply and I’ll sort it out.

21 days overdue

Hi Priya, invoice 1042 for $640 is now three weeks overdue. Could you let me know when I can expect payment? If there’s a problem, I’d rather hear about it so we can work it out.

Always include the payment link. The easier it is to pay, the faster people pay.

Where AI fits

AI isn’t needed for the reminders themselves. Where it helps is reading the replies. When a customer writes back “we’re waiting on our boss to approve it” or “can we split it into two payments?”, AI can sort the reply, pause the reminders, and send you a one-line summary so you know exactly what’s going on.

Want it set up for you?

A quote and invoice follow-up system that connects your quoting tool, your invoicing app and your inbox is one of the builds on my services page. If you’re not sure it’s worth it for your volume, book a free 20-minute discovery call and we’ll work it out together.

This post walks through a typical automation that takes a website inquiry and turns it into a booked appointment in about two minutes, without you touching it. It’s a sample setup, not a specific client’s, and the details change from business to business. The shape stays the same.

Who this is for

Any business where the next step after an inquiry is “book a time”: physiotherapists, massage therapists, dentists, salons, tutors, accountants, consultants, photographers, home-service estimates.

The flow at a glance

1. Someone fills in your website form.
2. AI reads the message and works out what they need.
3. They get a personal reply with the right booking link within a minute or two.
4. If they book, everything else updates on its own.
5. If they don’t, a gentle follow-up goes out.
6. You see a short summary instead of a pile of emails.

Step 1: The form asks the right questions

The form should ask only what you need to point someone to the right appointment: name, email or phone, what they need help with, and maybe how soon. Every extra field means fewer people finish the form.

A free-text box (“Tell us a bit about what’s going on”) is fine. That’s what the AI is for.

Step 2: AI sorts the request

When the form is submitted, the automation sends the message to an AI model with a short brief about your business: your services, who each one is for, your hours and what you don’t do.

The AI returns a few simple answers:

• Type: new client, existing client, question, or not a fit.
• Best service: for example “initial assessment, 60 minutes” or “colour consultation”.
• Urgency: normal, soon, or urgent.
• Anything you need to know: a one-line summary.

It doesn’t make medical, legal or financial judgements. If the message looks like an emergency, or it isn’t sure, it flags it for you instead of guessing.

Step 3: A personal reply goes out

Within a minute or two, the person gets an email or text written in your voice:

Hi Sam, thanks for getting in touch about your shoulder. For a first visit we’d start with a 60-minute initial assessment. You can pick a time that suits you here: [booking link]. If none of those times work, just reply and we’ll sort something out.

The booking link goes straight to the right appointment type in your booking tool (Calendly, Cal.com, Jane, Square Appointments, Acuity or whatever you already use), so they don’t have to choose from a long menu.

Step 4: Booking updates everything

When they book, the automation:

• adds them to your client list or CRM, with the AI’s summary attached,
• sends any intake form you use,
• sets reminders by text or email, for example the day before,
• tells you, in one short notification.

Step 5: The follow-up for people who don’t book

Plenty of people mean to book and forget. If nobody has booked a day or two after the first reply, one friendly follow-up goes out. If they still don’t book, the automation stops. Nobody likes being chased five times.

Step 6: Your summary

At the end of each day, or each morning, you get a short list: new inquiries, who booked, who didn’t, and anything flagged for you. That’s the only part you have to read.

What it costs to run

The ongoing cost is mostly the tools you already pay for, plus a small amount for the automation platform and AI usage. For most small businesses that comes to a few dollars to a few tens of dollars a month, depending on volume and which tools you use.

The setup is the bigger cost. A build like this falls under the automation builds on my services page, which start at $1,200 CAD.

Doing it yourself

You can build a simple version yourself with your booking tool’s own features: an auto-reply that includes your booking link, built-in reminders, and a weekly look at who enquired but didn’t book. That gets you a lot of the benefit for free.

The AI sorting, the personal reply and the follow-up are where it starts to need a proper build.

If you want to talk through what this would look like for your business, book a free 20-minute discovery call. And if most of your inquiries come in by phone rather than through a form, start with the missed call post.

This isn’t a reason to avoid AI. It’s a reason to set things up properly, which is easier than it sounds.

This is general information, not legal advice. If you handle sensitive information (health, finances, children, legal matters), talk to a lawyer or the Office of the Information and Privacy Commissioner of Alberta before you automate anything.

What PIPA covers

PIPA applies to private-sector organizations in Alberta, including small businesses, and to the personal information they collect about customers and employees. In plain terms, it expects you to:

• collect only the personal information you reasonably need, and for reasons a customer would expect,
• protect it with reasonable security,
• keep it only as long as you need it, then destroy it or make it anonymous,
• let people see what you hold about them if they ask,
• report a privacy breach to the Commissioner when it poses a real risk of significant harm to someone.

None of that is new because of AI. What AI changes is how easily information gets copied into new places.

The part that catches people out: service providers outside Canada

If you use a service provider outside Canada to collect, use or store personal information, the Alberta government’s guidance says your written policies need to include which country that happens in and what the provider is allowed to do with the information. You also need to tell customers how they can see those policies and who at your business can answer their questions about it.

Most AI chat tools, automation platforms (Zapier, Make and others) and many booking and CRM apps run on servers in the United States or elsewhere. So if customer details flow through them, that rule likely applies to you. You can read the government’s summary on alberta.ca and the Act itself on open.alberta.ca. Laws get updated, so check the current version.

A practical checklist for using AI with customer data

1. Make a list of your tools. Write down every app that touches customer information: email, booking, invoicing, CRM, AI chat tools, automation platforms. Note where each one stores data, which is usually on its privacy or security page.
2. Send less. Most automations don’t need everything. An AI that sorts inquiries needs the message, not the customer’s address and date of birth. Strip out what isn’t needed before it goes to an AI model.
3. Check what the AI provider does with your data. Business and API plans from the major AI companies usually don’t train on your data by default, but free consumer plans may. Read the data-use settings and pick the option that keeps your data out of training.
4. Update your privacy policy. Add a short section listing the outside-Canada service providers you use, the countries involved and why. Name the person customers can contact with questions.
5. Tell customers. A line on your website forms and booking pages is usually enough: what you collect, why, and a link to your privacy policy.
6. Turn on two-factor login for every tool on your list. Most breaches at small businesses start with a stolen password.
7. Set a clean-up rule. Decide how long you keep inquiry and customer records in each tool and delete old ones on a schedule. Many tools can do this automatically.
8. Know what you’d do in a breach. If a laptop is stolen or an account is taken over, you need to work out whether there’s a real risk of significant harm. If there is, you report it to the Commissioner. The OIPC website has guidance on how.

What this looks like in a real automation

Take the missed-call setup from my earlier post. A privacy-minded version would:

• send only the caller’s message text to the AI model, not their full contact history,
• use an AI plan that doesn’t train on your data,
• keep the conversation log in one place you control, with a clean-up rule,
• be listed in your privacy policy, with the country it runs in.

That’s a few extra minutes of setup. It’s also the difference between “we use AI” being something you’re comfortable saying to customers and something you’d rather they didn’t ask about.

A privacy-policy paragraph you can adapt

Here’s a starting point for the outside-Canada part of your privacy policy. Replace the parts in square brackets, delete any tools you don’t use, and have it checked if your situation is unusual.

Service providers outside Canada

We use a small number of service providers to run our business, and some of them store or process personal information outside Canada. These are:

• [Tool name], used for [booking appointments / sending email / sorting inquiries], which stores information in [country].
• [Tool name], used for [purpose], which stores information in [country].

These providers may only use your information for the purposes listed above, and information stored outside Canada may be accessible to authorities in those countries under their laws. If you have questions about how we use service providers outside Canada, or want a copy of our policies, contact [name or job title] at [email or phone].

Keep the list current. When you add or drop a tool, update this paragraph the same day.

Want a second pair of eyes?

When I build automations for clients, this list is part of the job: what data flows where, what can be left out, and what needs a line in the privacy policy. If you already use AI tools and want to check your setup, or you’re planning something new, book a free 20-minute discovery call.

The problem is volume. Writing a fresh, thoughtful reply to every “great service, thanks!” gets old fast, so most owners either stop replying or paste the same line under every review. Customers notice both.

Here’s how to use AI to draft replies that still sound like you, and where to draw the line.

What the difference looks like

Here’s a made-up review for a made-up hair salon, with the kind of reply most owners paste in and the kind this approach produces.

Review (5 stars): “Booked last minute for my sister’s wedding and Mei fixed the colour disaster I walked in with. Will be back!”

Generic reply: “Thank you for your review! We appreciate your business.”

AI draft in the owner’s voice: “So glad Mei could rescue the colour in time for the wedding! Hope the day was wonderful. See you next time. – Lena”

The AI draft takes no longer to approve than the generic line takes to paste, and it tells future customers that a real person reads what they write.

The rule: AI drafts, you decide

AI is good at the first draft. It’s fast, it doesn’t get bored, and it can pick out the detail in a review that’s worth mentioning. It is not good at judgement: whether to apologise, whether to take something offline, whether a review is fake.

So split the work:

• Positive reviews (4 and 5 stars): AI drafts, you glance at it, it goes out.
• Mixed reviews (3 stars): AI drafts, you edit before posting.
• Negative reviews (1 and 2 stars): AI can suggest a calm starting point, but you write the reply yourself or rewrite the draft completely. Never auto-post these.

Step 1: Teach it your voice

Find five or six review replies you’ve written that you’re happy with. If you don’t have any, write them now, for a mix of positive and mixed reviews. These are your examples.

Then write down a few rules about how you talk to customers. For example:

• Use the customer’s first name if they gave it.
• Mention one specific thing from their review.
• Keep it to two or three sentences.
• Sign off with your first name, not “The Management”.
• Never offer discounts or refunds in a public reply.

Step 2: Draft replies with any AI chat tool

You can do this today with ChatGPT, Claude, Gemini or similar. Paste in your examples and rules once, then paste a new review and ask for a reply in the same style.

A prompt that works:

Here are examples of how I reply to Google reviews for my business, and my rules. Write a reply to the new review below in the same voice. Mention one specific thing the customer said. Two or three sentences.

Read the draft out loud. If it sounds like something you’d never say, tell it why and ask again. After a few rounds you’ll have a set of examples that produce good drafts most of the time.

Step 3: Make it automatic (optional)

Copy and paste works fine for a few reviews a week. If you get more than that, or you have several locations, it’s worth automating:

1. A new review arrives on your Google Business Profile.
2. The automation sends it, along with your examples and rules, to an AI model.
3. The draft lands somewhere you already look: your email, a Slack channel, or a simple approval list.
4. You tap approve (or edit) and the reply is posted.

Negative reviews skip the queue and come straight to you with a notification, so you can reply quickly and carefully.

What not to automate

• Anything that admits fault or makes promises. That’s a business decision, not a writing task.
• Reviews that might be fake or from a competitor. Flag them to Google instead of replying.
• Anything involving health, legal or personal details. If a review mentions a medical visit, a legal matter or a private situation, reply in general terms or take it offline. Don’t confirm someone was your client.

Why it’s worth doing

Replying consistently shows future customers how you treat people when things go right and when they don’t. It takes most owners a few minutes a day once the drafts are good, instead of an hour every few weeks of catching up.

If you’d rather have the whole thing set up, connected to your Google Business Profile with an approval step that fits how you work, that’s the kind of build I do. Book a free 20-minute discovery call and we’ll see whether it’s worth it for the number of reviews you get.

For more small jobs like this one, see five jobs a small business can hand to AI this month.