A service business can automate eight jobs most owners do by hand: enquiry logging, quote drafts, booking confirmations, report cover notes, missed-call summaries, payment chasing, feedback requests and a weekly summary. For an illustrative 11-person testing laboratory, all eight ran on about 3,160 Zapier tasks a month: $89 billed annually, plus about $5 of AI and 35-45 hours to build.
The AI is the cheap part. The platform bill scales with steps and volume rather than intelligence, so the booking confirmations, which use no AI at all, are joint top of the task count. And most of the build time goes on tidying things the automations depend on, such as the price list, the quote template and duplicate client records. A business whose data is already tidy builds faster and cheaper.
The laboratory, its tools and its monthly volumes
The illustration is a commercial testing laboratory: clients send water, soil and material samples, the lab tests them and issues reports. Eleven people: an owner, a lab manager, six analysts, two client-services staff and an account manager. Its existing tools are Google Workspace, HubSpot's free CRM, a cloud accounting package with a Zapier connection, a sample log that the lab system updates in a Google Sheet, and a cloud phone system that emails voicemail transcripts.
| Activity | Per month |
|---|---|
| New enquiries (web form and email) | 150 |
| Quotes sent | 90 |
| Sample batches received | 350 |
| Reports issued | 350 |
| Missed calls with voicemail | 80 |
| Payment reminders needed | 120 (40 overdue invoices, up to 3 stages) |
| Feedback requests | 120 |
How the costs below are counted
Zapier bills per task: each successful action step counts, while triggers, filters, Paths and Formatter steps don't. An AI step run through your own OpenAI or Anthropic account counts as one task, and you pay the model provider for tokens; Zapier's built-in AI step instead uses one, three or five tasks per run depending on the model tier. The lab uses its own API keys. Zapier's Professional plan on its pricing page runs from $19.99 a month billed annually for 750 tasks, through $49 for 2,000, to $89 for 5,000 ($133.50 billed monthly).
AI costs use list API prices: OpenAI's gpt-5.6-luna at $0.20 per million input tokens and $1.20 per million output tokens for sorting and extraction, and Anthropic's Claude Sonnet 5 at $2 and $10 for text a client will read. Token counts per run are estimates, so treat each AI figure as an order of magnitude. Hours saved are the lab's before-and-after estimates in this illustration.
1. Enquiries logged and triaged in the CRM
What it does: a new web form or email to the enquiries inbox triggers the Zap. An AI step extracts company, contact, sample type, tests requested, number of samples, deadline and an urgency flag. Zapier finds or creates the contact in HubSpot, creates a deal with a three-line summary, and posts urgent enquiries (roughly 30%) to the team chat.
Numbers: three tasks per enquiry plus one for urgent ones: about 495 tasks a month, or $14 of the plan. AI cost: 150 runs of roughly 1,200 tokens in and 250 out on gpt-5.6-luna, about $0.08. Setup: 3 hours. Saving: client services spent about 4 minutes reading and logging each enquiry; a glance at the deal takes one. About 7.5 hours a month.
What broke in testing: clients who emailed from two addresses got two contacts. The fix was to search by email domain first and let a person merge near-matches. An illustrative extraction that passed: "Company: regional water supplier; tests: metals suite, pH, conductivity; samples: 24; deadline: 14 Oct; urgency: normal".
How the lab checks it: once a week, client services compares the number of new deals with the number of emails in the enquiries inbox. A gap means something failed or was misfiled; in the first month there was one, a supplier invoice wrongly logged as an enquiry.
2. Quote first drafts from the price list
What it does: when a deal moves to "quote requested", an AI step matches the tests the client asked for to codes on the lab's price list, flagging anything it can't match. Zapier writes the matched lines into a quote sheet where ordinary formulas calculate prices, discounts and turnaround surcharges. A second AI step writes the cover email, Zapier fills a Google Docs quote template and saves a Gmail draft for the account manager.
Numbers: five tasks per quote, 450 a month, $12.70 of the plan. AI cost: the matching step on gpt-5.6-luna is about $0.09 a month; the cover email on Claude Sonnet 5, at roughly 2,000 tokens in and 600 out, about $0.90. Setup: 8 hours, mostly tidying the price list so every test had one code and one name. Saving: quotes took about 25 minutes; checking a draft takes 8. About 25.5 hours a month, and quotes now go out the same day.
What broke in testing: "metals" was matched to the full heavy-metals suite when the client wanted three specific metals. The fix: the AI must flag any ambiguous request as "check with client" instead of choosing, and prices come only from formulas, never from the model. That rule is the reason this automation is safe.
How the lab checks it: the account manager records every edit made to a draft quote for the first month. By week four, edits were down to wording and the occasional flagged test, which is the signal the matching is trustworthy. A rising edit count would mean the price list has drifted from what clients ask for.
3. Sample-arrival confirmations, with no AI at all
What it does: when the lab system adds a received batch to the sample log, Zapier emails the client a confirmation (client reference, number of containers, expected report date from a formula) and marks the row confirmed.
Numbers: two tasks per batch, 700 a month, $19.70 of the plan. AI cost: none. Setup: 2 hours. Saving: client services used to send these by hand at about 2 minutes each when they remembered; about 11.7 hours a month, plus fewer "did you get our samples?" calls.
The lesson: this is the joint-largest task user and has nothing to do with AI. If your own booking or job system can send confirmations itself, use that feature and save the tasks. Not every automation in an AI project needs AI.
How the lab checks it: a formula in the sample log counts rows received today without a "confirmed" mark. Anything above zero at 5pm means a confirmation failed, usually because a client contact had no email address on file.
4. Report cover notes drafted for review
What it does: when a report is marked "authorised" in the sample log, an AI step drafts a short cover email in plain English: which results met the limits on the client's method sheet, which didn't, and anything the client asked to be told. Zapier saves it as a Gmail draft to the client contact with the report link. An analyst reads the draft against the report and sends it.
Numbers: two tasks per report, 700 a month, $19.70 of the plan. AI cost: about 2,500 tokens in and 400 out on Claude Sonnet 5, roughly $3.15 a month. Setup: 6 hours, plus testing on 30 past reports. Saving: analysts spent about 6 minutes per cover email; review takes 2. About 23 hours a month.
What broke in testing: on one report the model wrote "all results within limits" when a result sat exactly on a "less than" limit, which counts as a failure. The fix moved the comparison out of the AI: a formula in the sample log marks each result PASS or FAIL, and the model only writes up what the formula decided. An illustrative draft after the fix: "All 12 samples met your limits except sample W-07, where lead measured 10 µg/L against a limit of less than 10 µg/L. The full results are in the attached report." Every draft is still reviewed; this automation stays in "draft for a person" mode permanently.
How the lab checks it: the lab manager reviews ten sent cover notes a week against their reports, looking for any wording that overstates a result. Two in the first month needed softening, both where the client's method sheet had unusual limits.
5. Missed-call and voicemail summaries
What it does: the phone system emails each voicemail transcript; an AI step summarises it in two lines and pulls out the caller's name, company, reference numbers and what they want. Zapier finds the contact in HubSpot, logs the summary as a note and creates a callback task for the right person.
Numbers: three tasks per call, 240 a month, $6.76 of the plan. AI cost: about $0.05 a month on gpt-5.6-luna. Setup: 3 hours. Saving: listening, noting and logging took about 5 minutes; checking the note takes one. About 5.3 hours a month. The full pattern is covered in logging every phone enquiry in your CRM with AI call summaries.
What broke in testing: transcripts misheard sample references ("W-O-7" became "W07" or "double-you oh seven"). The summary now labels any extracted reference as "unconfirmed", and the callback task reminds staff to confirm it.
How the lab checks it: open callback tasks older than one working day show on a HubSpot view that client services looks at each morning.
6. Payment chasing in three stages
What it does: every working morning, Zapier searches the accounting package for invoices 7, 21 and 35 days overdue. For each, an AI step drafts a reminder that names the client's PO and the reports it covered; stages one and two are sent automatically, stage three is saved as a draft for the owner. Each reminder is logged against the client in HubSpot.
Numbers: one search a day (22) plus three tasks per reminder (360): about 380 tasks, $10.76 of the plan. Zapier's looping can add a few tasks, so check the task history in the first week. AI cost: about $0.06 a month. Setup: 5 hours. Saving: ad hoc chasing took about an hour and a half a week; approving stage-three drafts takes 20 minutes. About 4.5 hours a month, and invoices are chased on time every time. Wording that stays polite at each stage is covered in chasing late payments with AI reminders that sound human.
What broke in testing: a reminder was drafted for a client whose invoice was disputed after a retest. The fix: a "disputed" tag in the accounting package, and a filter step (no task cost) that skips tagged invoices.
How the lab checks it: the owner compares the overdue total in the Monday summary with the previous month's. After two months it had fallen, which is the only measure that matters for this one.
7. Feedback requests and review reply drafts
What it does: each morning a Zap finds reports sent five days earlier and emails the client a two-question feedback request with a review link. When a new review arrives, an AI step drafts a reply and posts it to the team chat for the account manager to edit and publish.
Numbers: 22 searches, 120 requests and about 15 review drafts at two tasks each: roughly 170 tasks, $4.84 of the plan. AI cost: under one cent. Setup: 3 hours. Saving: none, because nobody did this before. It's new work, about 30 minutes a month, done because the owner wanted more reviews and earlier warning of unhappy clients.
What broke in testing: the first request went to a client who had complained that week. The Zap now skips any client with an open support ticket in HubSpot.
How the lab checks it: response rate to the requests, and the number of new reviews a month. Both are visible in HubSpot and the review platform, so nothing extra needs building.
8. A Monday summary for the owner
What it does: at 7am on Mondays, Zapier pulls last week's rows from three sheets (enquiries and quotes, samples and turnaround, invoices) and an AI step writes a 200-word summary: volumes against the four-week average, turnaround against promise, quotes won and lost, overdue totals, and anything that moved more than 20%. Zapier emails it to the owner and lab manager.
Numbers: five tasks a run, about 22 a month, $0.62 of the plan. AI cost: around 20,000 tokens in and 800 out on Claude Sonnet 5, about $0.20 a month. Setup: 3 hours. Saving: the owner's Monday spreadsheet session of about 80 minutes became a 10-minute read, about 5.7 hours a month.
What broke in testing: the summary explained a drop in enquiries as "seasonal", with no evidence. The prompt now says: report changes and flag the big ones; don't explain them.
How the lab checks it: each Monday, the owner checks one figure in the summary against the sheet. If a figure is ever wrong, the summary stops until the cause is found, because a summary you can't trust is worse than none.
All eight on one page: tasks, costs and hours
| Automation | Tasks a month | Share of $89 plan | AI cost a month | Setup hours | Hours saved a month |
|---|---|---|---|---|---|
| 1. Enquiry logging | 495 | $14.00 | $0.08 | 3 | 7.5 |
| 2. Quote drafts | 450 | $12.70 | $1.00 | 8 | 25.5 |
| 3. Confirmations | 700 | $19.70 | $0 | 2 | 11.7 |
| 4. Report cover notes | 700 | $19.70 | $3.15 | 6 | 23.3 |
| 5. Call summaries | 240 | $6.76 | $0.05 | 3 | 5.3 |
| 6. Payment chasing | 382 | $10.76 | $0.06 | 5 | 4.5 |
| 7. Feedback and reviews | 172 | $4.84 | under $0.01 | 3 | New work |
| 8. Monday summary | 22 | $0.62 | $0.20 | 3 | 5.7 |
| Total | 3,161 | $89 | About $4.50 | 33, plus testing | About 83 |
So the eight cost about $94 a month on annual billing, or about $138 on monthly billing, and roughly 35-45 hours to build and test. On the lab's own estimates they give back about 80 hours a month across the team, most of it from quotes and report cover notes. Before copying the numbers, run your own: volumes and minutes per job decide everything, and calculating the ROI of an automation before you build it shows how to do that with your figures.
The order the lab built them in
Building all eight at once is how projects stall. The lab spread them over six weeks, starting with the lowest-risk ones so the team got used to the tools before anything touched a client-facing message:
- Week 1: confirmations (3) and call summaries (5). Simple, no client-facing AI text, quick wins everyone noticed.
- Week 2: enquiry logging (1) and the Monday summary (8). Internal only; good practice at writing extraction prompts.
- Weeks 3-4: quote drafts (2), including the price-list tidy-up, which took longer than the build.
- Week 5: payment chasing (6), with the "disputed" tag set up first.
- Week 6: report cover notes (4) after two weeks of testing on past reports, then feedback requests (7).
Each automation ran alongside the old manual method for its first week, so a failure meant a duplicate email at worst rather than a missed one.
Cutting the bill: fewer tasks, or a different platform
- Move no-AI jobs into existing software. If the lab system could send confirmations itself, 700 tasks would disappear. That alone wouldn't drop the lab below the 2,000-task tier, but combined with the next two changes it could.
- Merge steps. Writing the deal summary into the contact record instead of a separate note, or skipping the chat post for non-urgent enquiries, trims a task per run.
- Batch where timing doesn't matter. The feedback requests could go out as one weekly batch from the CRM's own email tool rather than one Zap run each.
- Consider Make. Make counts credits per module action. A rough count for these eight, using instant triggers, comes to about 4,400 credits a month, close to the 5,000 included in its entry plan at about $9 a month. Scheduled triggers that check every few minutes can push that up, so watch the usage page in week one. The build is fiddlier, and when Zapier gets too expensive helps you judge the switch.
Five fixes from the first month of running all eight
- Put the decisions in formulas, not prompts. Prices (automation 2) and pass/fail calls (automation 4) moved out of the AI after testing. The model writes; formulas decide.
- Every automation that writes to a client starts as a draft. Automations 2, 4 and stage three of 6 still do. The lab reviews whether any can send on their own after three months of clean drafts, using the approach in adding human approval steps to AI automations.
- Skip rules protect relationships. Disputed invoices and open complaints now stop reminders and feedback requests. Both rules came from near-misses in testing.
- Test runs cost tasks too. The first month used about 400 more tasks than expected because of testing and a loop that ran twice. Start a tier higher than you think, or keep testing on a copy with small volumes.
- Someone owns each automation. Client services owns 1, 3, 5 and 7; the account manager owns 2 and 6; the lab manager owns 4 and 8. The owner is who gets the alert when a step fails.
The same eight for a courier firm
The pattern carries across service businesses; only the nouns change. An illustrative courier firm running contract deliveries would map them like this: enquiries become contract enquiries logged with volumes and delivery areas; quotes become rate-card drafts where formulas apply the rates; confirmations become "collection booked" emails from the job system; report cover notes become proof-of-delivery summaries for account clients; call summaries, chasing, feedback and the Monday summary work as above. The volumes would differ, perhaps far more confirmations and far fewer quotes, which is exactly why the costing has to be redone with your own numbers rather than borrowed. If you plan to build on Zapier, adding AI steps to Zapier covers the classify, summarise and draft steps used in six of the eight.
Questions about costing service-business automations
Why does the platform cost so much more than the AI?
Because the platform charges for every action step, and most steps aren't AI: finding a contact, creating a record, sending an email. The AI steps themselves use cheap models on short texts, so a few hundred of them cost cents. That's why cutting steps, moving simple jobs into your existing software and choosing a platform whose pricing fits your workflow shape save more than picking a cheaper model.
Could one person build all eight?
Someone comfortable with Zapier or Make and patient with testing could, over four to six weeks part-time. The hard parts aren't technical: agreeing the quote template, tidying the price list, deciding which reminder stage needs approval. If nobody in the team has done automation before, build the two simplest first (confirmations and call summaries) and decide after those whether to continue in-house or get help.
What happens to the automations if we change software?
Each automation depends on its trigger and action apps, so switching CRM or accounting package means rebuilding the steps that touch it. Keep a one-page note per automation listing the apps, the fields it uses and the prompts, and export the prompts somewhere you control. With that note, a rebuild on new software takes a fraction of the original time, because the design decisions are already made.
Further reads
- AI Workflow Automation Examples: 20 Processes to Automate First — Twenty more processes, scored so you can pick your own first eight.
- How to Stop Zapier and Make Automations Breaking Silently — Add alerts so a failed step never goes unnoticed.
- Quote to Cash: Connect Quotes, Invoices and Payments With AI — Join automations 2 and 6 into one quote-to-payment flow.
- How to Get a Weekly Business Summary Emailed to You by AI — A fuller build of automation 8.
- How to Build Your First AI Automation in Make, Step by Step — Build on Make if you take the cheaper route described above.
- Automate Client Onboarding With AI: Forms, Contracts, and Emails — The natural ninth automation once new clients start arriving faster.
- Too Many Zaps? Signs Your Automations Need Consolidating — A 13-point automation sprawl checklist with a clinic's filled-in audit: 23 Zaps cut to 10 and monthly tasks more than halved, without losing a step.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Zapier pricing page (task tiers, annual and monthly prices) and help pages on task counting and AI by Zapier; Make pricing and credits help pages; facts sheet for OpenAI and Anthropic API token prices. The laboratory, volumes, timings and test results are illustrative.