After go-live you pay for three things: the automation platform (Make from about $9 a month, Zapier from $19.99), AI usage (often a few dollars a month or less at API rates), and people: an hour or two a month of checking, plus paid fixes when a connected app or AI model changes. The people cost is usually the biggest.
Automations don't wear out, but everything around them changes. A booking system renames a field, a vendor retires the model your prompt was tuned on, a busy month runs through your plan's task limit, or the prices in your AI fact sheet go out of date. Each fix is small. The expensive part is not noticing for a week, which is why the monitoring hour matters more than the subscription.
Every running cost, itemised
Here's the full list for a typical small-business automation built on Zapier or Make with one or two AI steps. Not every line applies to every setup, but it's worth ruling each one in or out before you budget.
| Cost | How it's charged | Typical size for a small setup | Where to check it |
|---|---|---|---|
| Automation platform | Monthly or annual plan, sized by tasks (Zapier) or credits (Make) | Zapier Professional: 750 tasks for $29.99 monthly or $19.99 a month billed annually; 2,000 tasks for $73.50 monthly or $49 a month annually. Make: from about $9 a month for 5,000 credits | The platform's usage page, monthly |
| AI usage | Per token through your own API key, or as extra tasks | A few hundred drafts a month on a low-cost model is usually under $5. On Zapier, each AI step uses 1, 3 or 5 tasks per run depending on the model tier, or 1 with your own key | The AI provider's billing page, with a spending cap set |
| Messaging fees | Per message | On the WhatsApp Business Platform, from 1 October 2026, service replies beyond the first 1,000 a month per number are charged at rates that vary by recipient market | Meta's pricing page |
| Monitoring time | Internal hours | 15 to 30 minutes a week | An upkeep log |
| Content upkeep | Internal hours per change | 1 to 2 hours whenever prices, timetables or policies change | The "last updated" date on your AI fact sheet |
| Fixes after app changes | Freelancer or agency hours | Fiverr's cost guide for AI automation experts (April 2026) puts integration updates at $100 to $500 per major system update | Written quotes |
| Support retainer (optional) | Monthly fee | The same Fiverr guide lists maintenance agreements from $50 to $500 a month | The contract's scope and response times |
| Migrations | One-off, when a vendor retires a model or feature | A few hours of testing and rework each time | Vendor emails and deprecation pages |
For hourly work, Upwork's 2026 rates guide lists marketing automation consultants at $40 to $90 an hour, which is a reasonable yardstick when a freelancer quotes for a fix. None of these figures is a price anyone owes you; they're the ranges the marketplaces themselves publish.
A day-tour operator's first year of running costs
Take an illustrative day-tour company: an owner, two office staff and three automations built in the spring.
- Enquiry drafts. Each new enquiry email goes to an AI step (using the company's own API key), which drafts a reply from a two-page fact sheet; the draft lands in Gmail for staff to check, and a row is logged in a sheet. That's 3 Zapier tasks per enquiry.
- Review requests. The morning after each tour, the lead booker gets a thank-you email with a feedback link. 1 task per booking.
- Weekly feedback summary. Every Monday the week's feedback rows are pulled, summarised by AI and emailed to the owner. 3 tasks a week, about 12 a month.
In the nine quieter months there are about 140 enquiries and 160 bookings a month, which is 420 + 160 + 12 = 592 tasks, inside Zapier's 750-task tier. In the three peak months there are about 320 enquiries and 420 bookings: 960 + 420 + 12 = 1,392 tasks, which needs the 1,500-task tier. The owner had two choices:
- Pay for the 1,500-task tier all year on annual billing: $468.
- Stay on 750 tasks at $29.99 monthly for nine months and switch to 1,500 tasks at $58.50 monthly for the peak: $269.91 + $175.50 = $445.41.
The second saves about $23 a year, provided somebody remembers to switch up before the first busy week and back down afterwards. Forget once and the Zaps can pause in the busiest fortnight of the year. The owner chose the annual plan.
AI usage is almost a rounding error. Each enquiry draft uses roughly 2,500 input tokens (the enquiry plus the fact sheet) and 350 output tokens. On Claude Haiku 4.5, at $1 per million input tokens and $5 per million output, that's about $0.004 a draft: around $0.60 in a quiet month and $1.36 in a peak month. The weekly summaries add about $0.08 a month. The year comes to roughly $10.
People time is where the money goes:
| Upkeep job | How often | Hours a year |
|---|---|---|
| Weekly check of run history and a few AI drafts | 20 minutes, 52 weeks | 17 |
| Updating the fact sheet for new prices and timetables, then testing | 3 times a year, 2 hours each | 6 |
| Small breakages fixed in-house | About 3 a year, 1.5 hours each | 4.5 |
| Testing after a model or vendor change | About once a year | 2 |
| Total | about 30 |
At an illustrative $30 an hour, 30 hours is $900. Add one outside fix a year when the booking system changed how it sends data, which a freelancer did in about 4 hours for $240 plus platform fees, and the first-year running cost is roughly $468 + $10 + $900 + $260, or about $1,640. Against that, the automations save around 20 staff hours in a quiet month and 40 in a peak month, about 300 hours a year, or $9,000 at the same rate. Upkeep comes to just under a fifth of the value, which is healthy. If it crept past a third, I'd look at simplifying before adding anything new.
Three ways to cover upkeep, priced side by side
The same three automations can be maintained in different ways. Yearly figures, using the costs above:
| Line | Owner does everything | Owner checks, freelancer fixes | Monthly retainer |
|---|---|---|---|
| Zapier, 1,500 tasks, annual | $468 | $468 | $468 |
| AI usage | about $10 | about $10 | about $10 |
| Staff time at $30 an hour | 36 hours: $1,080 (fixes take longer when you're learning) | 23 hours: $690 | 8 hours: $240 |
| Outside help | $0 | about 8 hours at $60: $480 | $250 a month: $3,000 |
| Total | about $1,560 | about $1,650 | about $3,720 |
The first two cost about the same; the difference is who carries the stress and how quickly a peak-season breakage gets fixed. The retainer costs more than twice as much and only earns its keep when downtime is expensive, for example a customer-facing chatbot, or when there are many automations and nobody inside to watch them. For how to judge a retainer offer, see whether an AI consulting retainer is worth it.
When a busy month pauses your automations
Usage-based plans have a failure mode that has nothing to do with bugs. Zapier's pricing page explains what happens at the limit: it emails you as you approach and reach your plan's task limit; if you've switched on pay-per-task billing, Zaps keep running at a higher per-task rate up to a maximum; without it, or once that maximum is reached, Zaps pause until the next usage period starts.
Here's how that plays out. A campsite on the 750-task tier, with pay-per-task billing switched off, hits its limit on the 19th of its busiest month. The booking-confirmation Zap pauses. For twelve days new bookings get no confirmation email, and the first anyone knows of it is when guests start phoning to check their pitch is booked. The fix takes five minutes (moving to the 1,500-task tier), but the phone calls took hours and a few guests booked elsewhere. The warning emails had gone to an inbox nobody read in peak season.
Make has its own version. Make uses a credit for every scheduled check, even when nothing new has arrived, so a scenario polling every 15 minutes uses about 2,880 credits a month before it does any work. Three such scenarios would use more than the 5,000 credits of the entry plan on checking alone. Instant triggers (webhooks) avoid most of that, where the app supports them. If you're weighing a bigger Zapier plan against moving platform, when Zapier gets too expensive works through the tipping point.
Failures that cost more than any subscription
The costliest breakages are the ones that don't look like breakages. The platforms help with loud failures: Zapier pauses a Zap automatically when 95% of its runs have errored over seven days, and Power Automate turns off a flow after 14 days of continuous failure. But a run that "succeeds" with the wrong data raises no alarm at all, and on Zapier, no error email is sent when an error handler you've built catches the problem, so the handler itself has to notify someone.
The tour operator had one of these. A booking-system update renamed the "Tour date" field to "Activity date". The review-request Zap didn't fail; it carried on sending, with an empty date, so 54 customers over nine days received "Thanks for joining us on ." The first sign was a customer replying to ask which tour they'd been on. The fix took an office manager 90 minutes: remap the field, add a filter step that stops the run when the date is empty (Zapier's filter steps don't use tasks), and send an alert whenever the filter stops a run. The direct cost was about $45 of time. The indirect cost was 54 slightly baffled customers at the moment the business wanted a review from them.
Two habits catch most of these early: a weekly look at a handful of real outputs, not just the error count, and a filter or check on every field that a message can't do without. Stopping Zapier and Make automations breaking silently covers the alert set-up in detail.
Model retirements and other changes you don't control
AI vendors retire models and features on their own schedule, and every retirement is a small maintenance job for someone. Recent examples from the vendors' own pages:
- Anthropic gives at least 60 days' notice before retiring a publicly released model. It notified developers on 5 June 2026 that Claude Opus 4.1 would be retired on 5 August 2026. An automation that names a retired model in its settings simply stops working on the date. Its model deprecations page lists each retirement with a recommended replacement.
- OpenAI removed its Assistants API on 26 August 2026, a year after announcing it, and pointed developers to the Responses API. Its deprecations page is worth a glance every quarter.
- Custom GPTs stop running on 11 December 2026 across ChatGPT plans. A guest house whose part-time staff answer questions using a custom "house rules" GPT needs to move that setup into a ChatGPT Project, or use OpenAI's migration to a plugin, before then. Expect an hour or two to move the instructions and files and re-test the common questions.
- Excel's =COPILOT() worksheet function was retired on 14 September 2026. Existing results stay in the cells, but no new formulas can be created, so any spreadsheet built around it needs reworking with the Copilot side pane.
Model changes can also break requests in subtler ways. Anthropic has deprecated the temperature setting on its Claude 4.7 and later models, and OpenAI's reasoning models don't support it, so an older automation that sets it may need its request edited when you switch models. A replacement model can also write differently: longer, more formal, or more willing to fill gaps.
The cheapest protection is a saved test set: ten to twenty real inputs (anonymised) with outputs you've approved. When a model changes, run the set through the new model and compare. You can use AI to do the first comparison:
You are checking an automation after a model change. For each of
the 12 test enquiries below you have the OLD approved reply and the
NEW reply. Using only the fact sheet, flag any NEW reply that:
- states a price, time or pickup point not in the fact sheet
- drops the booking link
- is over 120 words
- changes tone from friendly to formal
Output one line per test: number, PASS or FLAG, reason in under
15 words.
[fact sheet]
[12 enquiries, each with OLD and NEW replies]In an illustrative run, the check flagged test 3 ("new reply gives pickup at 8:15; fact sheet says 8:30") and passed the rest. But test 7's new reply had quietly dropped the cancellation policy that the old one included, and nothing in the criteria asked about it. Two fixes: add "omits the cancellation policy" as a criterion, and always read three results yourself, picked at random, however clean the summary looks.
Retainer or pay-as-you-go: the break-even sum
A support retainer is insurance, so price it like insurance. Take a $250-a-month retainer against pay-as-you-go fixes at $60 an hour. The retainer costs $3,000 a year, which buys 50 hours of pay-as-you-go work, or a little over 4 hours a month. Even with a freelancer handling every fix, the tour operator needed about 8 hours of outside help in a year, so pay-as-you-go wins easily.
A retainer starts to make sense when one or more of these is true: a customer-facing automation where a day of downtime loses bookings; more than a handful of automations touching money or customers; no one inside the business able to spend 20 minutes a week checking; or regular small changes, such as new tours each season, that you'd otherwise quote for one by one. If you do take one, the contract should state a response time for breakages, what the monthly health check covers, how many change hours are included and whether unused hours roll over, and that documentation is updated after every change.
Trimming the bill without skimping on checks
- Bill annually once volumes are stable. Zapier's 750-task tier is $359.88 a year on monthly billing and $239.88 on annual billing, a saving of $120.
- Size the plan for the busiest month and treat pay-per-task billing as a safety net rather than a plan, because the overflow rate is higher.
- Put filters first. Zapier's Filter, Paths and Formatter steps use no tasks, and on Make a bundle stopped by a filter uses no credits, so stop unwanted runs before the paid steps.
- Prefer instant triggers to frequent polling where the app offers them, especially on Make.
- Use your own API key for AI steps in Zapier (1 task per run instead of 3 or 5) and a low-cost model for drafting and sorting.
- Merge near-duplicates. Two Zaps doing almost the same job double the upkeep. An automation audit usually finds a few.
- Keep the documentation and the test set current. A fix that takes a freelancer one hour with notes can take three without them.
Some of these costs don't show up at all when a project is quoted; the hidden costs of AI implementation covers the ones that appear before go-live, and the cost to automate one workflow shows how build and running costs fit together.
A 20-minute monthly upkeep routine, filled in
On top of the weekly glance, a monthly routine catches the slower problems. Here's the tour operator's for a July, the busiest month:
- Task use against the plan: 1,180 of 1,500 used by the 24th; projected 1,470. Tight, so the pay-per-task safety net stays on this month.
- Run history: 6 errored runs, all from one Gmail sending limit on the 12th. Re-run by hand; no action needed.
- Five AI drafts read in full: one quoted last season's pickup time for the coastal tour. Fact sheet corrected and dated.
- Alerts: sent a test booking with no date; the filter stopped it and the alert arrived.
- Vendor notices: none affecting the models in use. Deprecation pages checked.
- AI spend: $1.31 against a $5 monthly cap.
- Time logged this month: 1 hour 40 minutes, plus 45 minutes on the fact sheet.
Kept up for a year, that log is also your best evidence at renewal time: it shows what the automations really cost to run, which plans you actually need, and whether a retainer would have been worth it.
Budgeting for AI upkeep: questions owners ask
Is there a standard percentage of the build cost to budget for maintenance?
Not a reliable one. A cheap build can carry high running costs if it polls constantly or touches many apps, and an expensive build can be cheap to run. Estimate from the items instead: the platform plan sized for your busiest month, AI usage, 15 to 30 minutes a week of checks, a few hours of fixes a year, and a day or two for any model or feature retirement.
Who should do the weekly checks: me or the person who built it?
Someone inside the business, ideally the person who benefits from the automation, because they notice when outputs look wrong. The builder is better used for fixes and changes. If nobody inside has 20 minutes a week, that's a strong sign to simplify the automation or pay for a monitored retainer.
What should I do when a vendor announces a model retirement?
Note the date, then list every automation that uses that model, including AI steps inside Zapier or Make. Two to four weeks before the date, run your saved test examples on the recommended replacement, fix any prompt differences, switch over, and read the outputs closely for a week. Anthropic gives at least 60 days' notice for its publicly released models.
Further reads
- AI Consultant Handover Checklist: What You Need Before They Leave — What you need in hand so anyone can maintain the automation.
- What If Your AI Vendor Shuts Down? Checks Before You Commit — Checks for the bigger version of a vendor change.
- What Does the ChatGPT API Cost for a Business Automation? — Work out AI usage costs for your own volumes.
- AI Vendor Lock-In: How to Keep Your Data and Prompts Portable — Keep prompts and data portable so migrations stay cheap.
- How to Read AI Software Pricing: Seats, Credits, and Usage Fees — Seats, credits and usage fees decoded before you renew.
- How to Review an AI Tool After 90 Days: Keep, Fix or Cancel — A structured keep, fix or cancel review for each tool.
- How to Evaluate an AI Implementation Proposal or Quote — A 22-point checklist for any AI implementation quote, with the phrases to pin down, a scoring sheet and two quotes compared over three years.
- How Much Should a Small Business Budget for AI? — Itemised AI budget lines, three budgets by team size, a ceiling formula based on hours saved, and a kitchen fitter's first year month by month.
- How Much Does AI Visual Inspection Cost a Small Manufacturer? — Real price bands for one AI inspection station, the line items quotes leave out, and a worked payback sum for a small moulding shop.
- Predictive Maintenance With AI: Realistic for a Small Workshop? — When AI predictive maintenance pays in a small workshop, the five cheaper rungs below it, and a CNC shop costed end to end.
- n8n Self-Hosted vs Cloud: Real Costs for a Small Business — Both price lists side by side, the upkeep hours self-hosting really takes, and a roofing contractor's year costed both ways.
- How to Calculate the ROI of an AI Automation Before You Build It — A nine-step pre-build ROI method, a copyable worksheet, a roofing contractor's quote follow-ups costed, and a removals firm's idea that failed the test.
- How Much Does a Custom AI Assistant Cost a Small Business? — Three ways to get an AI assistant that knows your business, what each costs up front and every month, and the upkeep that most quotes leave out.
- No-Code AI Tools vs a Custom Build: When Templates Stop Working — Six signs a Zapier or Make workflow has outgrown no-code, the middle options to try first, and a worked cost comparison for a care agency's shift offers.
- How Much Does an AI Consultant Cost for a Small Business? — Published AI consultant rates from three marketplaces, the costs that sit outside the fee, and three first-year budgets for a small brewery.
- Done-for-You vs Done-With-You AI Implementation: Which Suits You? — What done-for-you and done-with-you AI implementation leave you holding, how upkeep decides it, and a four-shop butcher that uses both.
- 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 and task-limit FAQ; Make pricing and help pages; Anthropic model deprecations page; OpenAI API deprecations page; Anthropic and OpenAI API pricing; OpenAI help article on custom GPT retirement; Microsoft Support on the COPILOT function; Meta WhatsApp Business Platform pricing notice; Fiverr cost guide for AI automation experts (April 2026); Upwork hourly rates guide (2026).