How Much Does Automation Maintenance Cost in 2026?

You already paid to build it. A freelancer stood up the Zaps. Someone pasted an OpenAI step on the lead form and called it done.

That invoice is not the bill that matters. The bill that matters is year two: the Zapier task tier, the OpenAI project nobody capped, the Tuesday the invoice Zap is off, and the person you page when it is.

This is not how to pick a first workflow. That is What Is AI Automation for Small Business?. It is not MVP build cost. That is How Much Does It Cost to Build an MVP in 2026?. It is not which platform to buy — see Zapier vs Make vs n8n. It is not the takeover how-to if you inherited a stack with no owner — see Inherited automations: the audit.

This is the cost of keeping automations running after they exist. Tools, tokens, people. Miss any one and the other two look cheap until a silent failure eats a client or a week of invoices.

The three line items

Buyers argue about the first cost. The third surprises them.

Tools: the subscription you can screenshot

Zapier sells tasks. Make sells credits. n8n sells executions. The unit is not interchangeable. A five-step flow that succeeds is five Zapier tasks, a handful of Make credits, or one n8n execution.

Last month’s invoice is the easy number. A 12-person agency at a few thousand successful actions is usually still on a mid Pro plan, not Enterprise.

The plan is a cap, not a forecast. Add one action step to eighteen Zaps and the same volume jumps a tier. Turn on a loop. Replay a week of failures. You used more of the unit it sells.

Official list prices change. Re-check before you renew: Zapier pricing, Make pricing, n8n pricing.

Tokens: the invoice from a different vendor

An AI classify, draft, or extract step is not “included” just because the module is on the canvas.

On Make paid plans, your own OpenAI or Anthropic key is 1 credit plus the provider invoice. Zapier still meters a successful AI action as a task; the model vendor bills tokens separately if you brought a key or used a usage-based feature. n8n executions do not include model spend.

Do not forecast that step from a blog’s per-token table. Rates move. A prompt that was cheap in March is a different invoice if someone swapped the model.

Set spend alerts on the company project. Current rates: openai.com/api/pricing. Guardrails: Production best practices. If the alert goes to a freelancer’s Gmail, you will meet the bill when finance forwards it.

People: the line that is not on the vendor PDF

Tools do not replay a held run. Tools do not notice a Zap that is still green and has not fired in three weeks. Tools do not write down which Google account owns the connection.

Someone has to own that: an hourly freelancer, an ops hire who context-switches for part of their week, a Care-style retainer, or nobody. Nobody is the default. Nobody is how you get the customer as the monitoring system.

People cost more than Pro 10K. They also prevent the failure that is more expensive than a year of Pro 10K. If you only budget the screenshot, you have budgeted hope.

What tool fees look like in 2026

Read the unit before you read the dollar. Pick the tier that covers a bad month, not an average Tuesday.

Zapier: you pay for successful actions

A task is a successful action step. Triggers do not count. Filters, Paths, Formatter, Delay, Looping, Sub-Zap, Digest, Storage, Tables, and Forms do not count. Error-handler path steps do count. Replay a run and the already-successful steps count again. Zapier published the definition here: How is task usage measured in Zapier.

Free is 100 tasks. Two-step Zaps. Fine for a personal experiment. Not a company inbox.

Pro yearly starts at $19.99/month. Team yearly starts at $69/month. List prices on zapier.com/pricing, Pro monthly / yearly:

Yearly is what a competent owner picks. Monthly is what you pay when nobody reviewed billing after the freelancer left their personal card on the account.

When the bill jumps: you added action steps, not users. You turned on error-handler paths. You replayed a backlog. You crossed 5,000 successful actions and the next published Pro row is 10,000 — or 10,000 and the next is 50,000. That gap is how a quiet extra step becomes a $200/month surprise.

Overage and hold live on the pricing page. If pay-per-task is off, you get held at the limit. Held is work that did not happen. A successful retry is still a task.

Make: you pay for credits

Credits replaced operations on 27 August 2025. Make documents the unit in Credits (updated 23 September 2026). Most modules cost 1 credit per operation. Router and error-handler modules cost 0. Bring your own OpenAI or Anthropic key on a paid plan and the AI module is 1 credit plus the provider’s tokens.

Free is 1,000 credits.

We confirmed Core list prices from the 4 September 2026 archive of Make’s official pricing page, because the live page often blocks bots. Re-check https://www.make.com/en/pricing before you buy. Archived Core monthly / yearly:

Pro and Teams sit above Core on the same credit rows. Do not copy a tweet for those numbers. Open the live page.

When the bill jumps: the unit is credits now, not operations, and an AI module on Make’s own provider is token-priced, not 1 credit. A retry storm burns the month in a few days. When you run out, scenarios stop. Webhooks queue until storage is used.

Make is often cheaper than Zapier at the same volume if someone reads the graph and you are not piping every lead through a hosted model. Cheap credits plus an uncapped Claude step is not a cheap stack.

n8n: you pay for executions

An execution is one full workflow run, any number of nodes. Unlimited users, workflows, and integrations sit on every plan. Prices on n8n.io/pricing:

Community self-host has no license fee. It is not free of a person. You lose paid features (SSO, environments, Git, sharing). Someone owns upgrades, secrets, and the box. If that someone is not on payroll, Community is a hidden retainer.

When the bill jumps: you scheduled every minute instead of every fifteen. You split one graph into five workflows and each trigger bills an execution. You hit Starter’s 2,500 and moved to Pro, or you outgrew 10,000 and paid for 50,000. History caps are not the execution cap. Hitting history trims logs. It does not always stop runs. You can be blind and still spending.

Business from €800/month is not Starter. Do not compare €20 to Zapier Pro 10K and call n8n “always cheaper.” Compare the unit, the person, and the failure mode.

When the tool bill is the wrong number to optimize

A 12-person agency at 8,000 Zapier tasks is on Pro 10K: $193.50 monthly or $129 yearly. That is $1,548 a year if you prepay. You will not save the company by migrating those eighteen Zaps to Make this quarter so Core 10K is $9/month. You will spend the quarter rebuilding, miss a week of invoices, and still need a human.

Optimize the tool bill when you are one published tier away from a jump you can avoid by deleting a wasteful action, or when volume is 10× this. Until then, the expensive line is people and incidents.

Token and AI-step cost

The AI step is where a $129/month stack becomes a second vendor.

OpenAI’s public rates live on openai.com/api/pricing. They change. This article will not publish a per-million-token number that is stale the week you read it.

Illustrative only: if a classify-and-draft step costs about $0.02–$0.08 per run depending on model and tokens, 2,000 runs is $40–$160. Confirm on the invoice. Extract-from-PDF with a long system prompt is not that band. A one-word category might be under it.

In the agency TCO later, we use $0.03 per classify on 1,500 runs: $45/month. That is an assumption, not a quote from OpenAI.

Spend alerts, loops, and model swaps

Turn alerts on before the first production week. A hard cap plus an email a human reads, on a company domain. OpenAI’s guide: Production best practices. If you cannot name who gets the 80% spend email, you have a card on file, not an AI automation.

Alerts do not replace a weekly look at runs. A loop can spend the month between Tuesday and Thursday.

The expensive failures are almost never “the model got 10% pricier.” They are a Zap that fires on every CRM update, including the update it just wrote. A Make scenario that retries a failed HTTP module and re-calls the model each time. An n8n cron that classifies the same ten rows all night.

Maintenance here is a person who can read the trigger. Not a new model. If you inherited the loop, do not add another model to “make it smarter.” Freeze, inventory, fix — inherited automations audit.

A model swap looks free. Tokens per run jump. The Zapier task count does not. The Make credit is still 1 if you brought your own key. The OpenAI invoice is the one that moves. Write the model name on the runbook. A swap that keeps the same job working can be in-scope on a retainer. A swap that turns a classifier into a long-form writer is a new automation.

Tokens also do not include Twilio, transactional email, enrichment, or OCR. In the TCO we put a small stand-in for “occasional Twilio/email.” Replace it. Do not let a token debate hide three other usage bills.

What people actually charge

No Upwork median. No Toptal graph. Those get invented and then quoted as fact. Here is what buyers typically see on calls — a range we see, not a published study.

Hourly freelance

Zapier and Make work is often quoted hourly, junior to specialist, in a wide band. On calls we see roughly $80–$150/hour for people who can own a production Zap without learning HubSpot on your dime. Below that you are often paying for someone who can click Publish and cannot explain a Paths branch. Above that you are usually buying a specialist, an agency blended rate, or a short-notice incident.

Four hours a month at that band is $320–$600/month before tools. That is the quote. They are not sitting on your Slack. You notice the failure, you write the brief, you wait. The week the invoice Zap is off, those four hours are used up on one incident or unavailable until next sprint.

Hourly is right when the stack is small, you can write a ticket, and you can wait.

Agency retainers and fixed projects

Agencies often sell retainers or fixed projects, not a raw hourly. A retainer buys a reply window, a cap (hours or incidents), and a standing look at the account. A fixed project buys a defined outcome: takeover, rebuild three flows, migrate nothing.

Spofylabs prices:

Takeover is $600–$1,200 fixed, 5–10 days. Inventory every live flow. Fix 1–3 broken or dangerous ones. Failure alerts. One-page runbook. Move ownership and billing to you. Not a platform migration. Not a new product. Details: AI Solutions.

Care is $1,000/month ($750 if you have 1–2 simple Zaps; we do not go below $600). Up to 6 hours or 4 incidents, whichever comes first. Weekly 15-minute health check: runs, errors, spend. Prompt and scenario changes that keep the same job working. Monthly one-pager. First reply in 1 business day, not 24/7. New automations are extra, $400–$900 each. You pay Zapier, Make, OpenAI, Twilio. We do not mark those up.

Other shops will quote higher or lower. If a quote has no cap, no “out of scope,” and “unlimited Zaps,” it is not cheaper. It is undefined.

In-house

In-house is a fraction of an ops or engineer salary, plus context-switching.

A $70,000 ops hire is the number we use in the TCO because it is a round assumption, not a salary survey. Salaries vary. 0.15 FTE of $70,000 is $10,500/year in salary, before benefits, before the Tuesday they were in client delivery and the Zap stayed red.

Engineer time is usually the wrong slice. You pay more per hour and they still will not want to own Zapier Autoreplay. In-house wins when “automation owner” is a real part of a job description, with admin. A name on a Slack channel is not 0.15 FTE.

Do not hire a retainer to find out what you have. If the builder left, buy a fixed takeover, then decide. DIY only if you have one to three simple Zaps, you already own the login, and the failure is a dead connection. How-to: audit post. After the stack is yours, the monthly question is hourly, in-house, Care, or hope. Hope is option 1 in the table below.

Worked 12-month TCO for the 12-person agency

Assumptions, labeled so you can replace every cell.

• 12-person agency. 18 live Zaps or scenarios. No in-house owner.
• 8,000 successful Zapier tasks per month. They need Pro 10,000: $193.50 monthly or $129 yearly (Zapier pricing).
• We use yearly in the table ($129 × 12 = $1,548). If nobody prepaid, use $193.50 × 12 = $2,322.
• One AI classify step on 1,500 of those runs. Illustrative $0.03/run = $45/month = $540/year. Confirm on the OpenAI invoice; do not treat $0.03 as a published rate (OpenAI API pricing).
• Occasional Twilio and transactional email. Stand-in $30/month = $360/year. Replace with last month’s invoice. We are not publishing a Twilio rate.
• No new platform. No migration project. Same 18 flows all year.
• Hourly column: 4 hours/month at the $80–$150/hour band we see on calls, not a survey.
• In-house column: 0.15 FTE of a $70,000 ops salary. Salaries vary. This is salary only, not loaded cost.
• Care column: $1,000/month. Client still pays vendors.

Tools subtotal we will reuse: $1,548 + $540 + $360 = $2,448/year.

Option 1 is cheaper until the first silent failure that loses a client or a week of invoices.

Walk it. The invoice Zap has been off for nine days. Nobody owns Alerts. The customer writes in asking where the bill is. You write off a week of send-outs, or you reconstruct them from Stripe. If that week is $15,000 or $25,000 of invoices — a labeled example; your number is on your ledger — option 1 just cost more than a year of Care. A lost client is worse. Eighteen Zaps usually include onboarding email, a Slack that starts delivery, or a form that creates the project. One of those going quiet for a month is not a $2,448 problem.

Hourly column. $6,288–$9,648 assumes the freelancer actually worked 48 hours on these flows, with a brief ready, and no emergency premium. It does not include your time to notice the failure. If you only called them twice all year, you got option 1 with two invoices. Same risk profile.

In-house column. $10,500 is 0.15 × $70,000. If your ops person is $55,000 or $95,000, redo the cell. If “0.15 FTE” is 20 minutes on Fridays plus panic, you have option 1 with a salary on a different line. Benefits make loaded cost higher — we are not inventing a multiplier; ask finance and multiply by 0.15. Context-switch does not appear in the grid. A person who can fix a Paths branch still has a client workshop at 2pm.

Care column. $14,448 is tools you already owe plus a human who looks every week, replies in one business day, and is capped at 6 hours or 4 incidents. New Zaps are $400–$900 extra. Four new flows in a year add $1,600–$3,600. Care is the expensive column on a year with no incidents. It is often the cheap column on a year with one bad week. This scenario has 18 live flows, not the $750 two-Zap price.

Left out on purpose: migration fees, “rebuild in n8n to save $1,200/year,” Enterprise Zapier, a Make token-credit surprise, an engineer’s loaded cost. Add those only if they are on your roadmap. A scoring sheet for new work is a different job: the AI Automation Audit.

What an automation support contract should include

If you hire anyone — us, a freelancer, another shop — write this down before work starts. A Slack handshake is how you get unlimited scope and no owner.

SLA. First reply is a number and a calendar. Ours is 1 business day, not 24/7. A one-person shop that claims 24/7 is selling a phone they will not answer at 2am, or they will answer and invoice a panic rate. Fix time is harder: a dead Google connection is an hour; a wrong-send into 400 CRM records is a project. Separate “we saw it” from “it is repaired.” Name timezone, holidays, and who covers the weekly check when the named person is out.

Hours versus incidents. A cap must be one or both, and you must say which comes first. Care is 6 hours or 4 incidents, whichever comes first. An incident is a live flow failing, drifting, or burning money — not “can you add LinkedIn.” Hours are the health check, the same-job prompt tweak, the replay, the runbook update. Cap only hours and one messy incident eats the month. Cap only incidents and a slow rebuild runs away. Unused hours should not roll forever. Forever rollups become guilt, then a cancel after a quiet quarter, right before the quarter that is not quiet.

In scope: keep existing flows doing the same job (field maps, filters, prompts, broken connections); weekly watch of runs, errors, and spend; alerts to a company inbox; a monthly one-pager; replay and backfill when history still exists and the fix is safe.

Out of scope: new automations (ours are $400–$900 each); migrations — that is a project, see Zapier vs Make vs n8n; 24/7; vendor bills; product/MVP work. Same-job versus new-job is the fight to have on paper. “Make the invoice email nicer” can be a prompt change. “Also post it to three new tools” is a build.

Who owns logins. Company email, password manager, card. Shared connections, not the contractor’s Google. If the builder keeps the only owner seat, you do not have a vendor. A takeover ends that: $600–$1,200, 5–10 days, billing on your email. Contractor gets a member seat you can revoke. How to remove them without deleting history: audit.

Who pays vendors. The client. Always. If the contractor’s card is on Zapier, you lose the account when they churn you. Pass-through is fine if it is itemized and rare. Markup on OpenAI is how trust dies.

Cancel terms. 30 days is normal. Same-month cancel with no notice is how you get a half-finished Paths rewrite and no runbook. Annual lock-in for a 6-hour retainer is usually their cash, not your stack. On the way out: latest one-pager, runbook, seat removed, 30-minute handoff. If that is not in the contract, you will buy another takeover.

Reporting. Weekly: 15 minutes, runs/errors/spend, not a slide deck. Monthly: one page — what to turn off, what is near a task or credit cap, what changed in a prompt. No written artifact, no way to tell work from vibes at renewal.

Copy-paste checklist

AUTOMATION SUPPORT CONTRACT — CHECKLIST

SCOPE
[ ] Named tools (Zapier / Make / n8n / OpenAI project)
[ ] Named live flows (or "all live in account X, inventoried on date Y")
[ ] Same-job changes in; new automations priced separately
[ ] Migrations / CRM rebuilds out unless a separate SOW

CAP
[ ] Hours per month: ____
[ ] Incidents per month: ____
[ ] Whichever comes first: yes / no
[ ] Unused hours: lapse / roll ____ months, cap ____

SLA
[ ] First reply: ____ business hours/days
[ ] Coverage window: timezone + days
[ ] 24/7: no (if yes, named backup person)
[ ] Escalation when the named owner is out

OWNERSHIP
[ ] Company email is owner on every tool
[ ] Company card on every vendor
[ ] Contractor is a revocable member
[ ] Password manager items listed
[ ] Shared connections, not personal Gmail

VENDORS
[ ] Client pays Zapier / Make / n8n / OpenAI / Twilio
[ ] Spend alerts to company inbox
[ ] No token markup

REPORTING
[ ] Weekly health check: duration and weekday
[ ] Monthly one-pager
[ ] Runbook exists and is updated on change

EXIT
[ ] Cancel notice: ____ days
[ ] Deliverables on exit: runbook, last report, seat removed, handoff call
[ ] History export before anyone is deleted

Pin it next to the runbook. If a proposal cannot fill the blanks, do not sign.

Red flags

Walk away, or rewrite the SOW.

The builder keeps the only login. Personal Gmail as Zapier owner. Make on their card. n8n on their laptop. You cannot export, remove them, or see the bill. Pay for a takeover before you pay for more Zaps.

Hourly with no cap. “We just work what you need.” You will need a lot the week it breaks, and you will argue about the invoice. Cap hours, cap incidents, or use a fixed project.

“Unlimited Zaps.” Unlimited is not a plan. It is a way to avoid saying no. Eighteen flows and a weekly check is a plan. Unlimited new ideas is closer to how teams spend 40 hours than to maintenance.

24/7 from a one-person shop. One person sleeps. If you need follow-the-sun, you need a roster and a price that is not $1,000. Our Care is 1 business day, on purpose.

No runbook. If the engagement ends and the only documentation is their head, you will buy the same work again. One page per dangerous flow is the minimum. The template is in the audit post.

Migrating the whole stack as the first move. New tool, same missing owner, same missing alerts. Migrations are how contractors generate a large invoice when the job was “turn Autoreplay on and fix the invoice Zap.” Migrate when the unit, the failure mode, or the person no longer matches — Zapier vs Make vs n8n — not because a thread said n8n is cheaper.

AI as the fix for a flow nobody can explain. A new classify step on a Zap you cannot name is a token bill plus the same silent failure. A green toggle is not evidence. A recent run that produced the side effect you expect is evidence. If the monthly note is only “all good,” ask for spend numbers.

When hourly is right vs when a retainer is cheaper

Hourly is right when you can write a ticket and wait. One to three simple Zaps. You own the login. The work is a field change, a dead connection, a held task limit. Someone internal notices failures within a day. You can stand a two-week wait.

Hourly is also right for a one-off: rebuild this scenario, add this filter, explain this bill. Buy it as a small fixed quote if you can. Ours for a new automation is $400–$900.

A retainer is cheaper when the calendar risk is the cost. More than a handful of flows that touch customers or money. Nobody internal will open Zapier on Monday. You already had one quiet month. You cannot brief a freelancer from scratch every incident because the last owner left.

Napkin math: four hours at $80–$150 is $320–$600. Care is $1,000 (or $750 for the tiny stack). The retainer is more cash in a quiet month. It is less cash in a month with two incidents, a spend spike, and no internal owner — because hourly is unavailable or blows past four hours.

If you are paying $1,000 and using one hour, you may be over-served. Drop to $750 if the stack shrank to 1–2 simple Zaps. If you are paying hourly and still being the monitoring system, you are not saving $400. You are donating ops time. In-house beats both when automation is a real slice of a job and the person has admin. It loses when 0.15 FTE is fictional.

How to budget the first 90 days after a freelancer leaves

Do not open a retainer on a pile you cannot name. Do not rebuild in a new Zapier account. Do not migrate.

Days 1–10: Takeover. Fixed $600–$1,200. Inventory. Fix 1–3 dangerous flows. Alerts. Runbook. Owner and billing on the company. You can follow the audit yourself if the stack is small and you already have admin. If you do not have admin, the waiting is the work; start that email today.

Budget the tool bills in this window too. Move the card. Set OpenAI spend alerts. Screenshot the current Zapier tier so you know whether you are on monthly $193.50 or yearly $129.

Days 11–40: first Care month, or a watched DIY month. If nobody internal will do the Monday check, start Care at $1,000 (or $750 if it really is 1–2 simple Zaps). Leftover failures show up here: the fourth-ranked flow, the Google Testing OAuth that dies on day 7, the Make scenario that still has incomplete executions off.

If you DIY this month, put 30 minutes a week on a calendar. Runs, errors, spend. Miss two weeks and you are back to hope.

Days 41–90: decide the steady state. You now have a map, a month of numbers, and a sense of incident rate.

• Quiet, 1–2 Zaps, you watch them: stop paying for Care, keep the runbook, stay hourly for the odd change.
• Quiet, 18 Zaps, nobody will watch: keep Care. $1,000 is the cheaper version of “we will notice in a month.”
• Noisy: you need more than 6 hours or 4 incidents. Either buy extra hours, retire flows, or staff in-house. Do not pretend the cap is unlimited.
• Wrong platform: now, and only now, look at migration. You have inventory. You can price a rebuild against a real task graph.

Cash for a typical 18-flow agency in those 90 days, labeled: Takeover $600–$1,200 + two months of Care $2,000 + two months of tools ($408 on the yearly mix, or $537 if still monthly). Call it $3,000–$3,800 plus whatever OpenAI already spent before you set the cap. That is the number for finance. Not “Zapier is $20.”

Score new work later with the AI Automation Audit. Do not mix the two.

FAQ

How much should I pay for Zapier help? A defined takeover from us is $600–$1,200 fixed. Ongoing: buyers see hourly in a wide band; we see $80–$150/hour on calls, and we sell Care at $1,000/month ($750 for 1–2 simple Zaps). A new Zap from us is $400–$900. “Unlimited” with no cap is not a price.

Is $1,000/month a lot? Against Pro 10K at $129 yearly, yes — larger than the tool. Against 0.15 FTE of a $70k hire ($10,500/year) it is similar. Against one silent week of invoices it is small. If you have two simple Zaps and you already watch them, use $750 or hourly.

Do I need 24/7? Almost no 12-person agency does. You need a same-business-day look when a money or customer flow fails, and a weekly check so silence does not last a month. 24/7 is an on-call roster. Our first reply is 1 business day.

Should I hire in-house? When “automation owner” is a real percentage of a job and that person has admin. A $70k ops hire at 0.15 FTE is $10,500 in salary in our example; salaries vary. Do not hire a full engineer to babysit 18 Zaps. Do not assign it as a 2% side quest and call it in-house.

What if I only have two Zaps? DIY if you own the login and they are simple. A retainer for that stack is $750/month from us, never below $600. Do not buy an 18-flow Care plan. Do not buy a migration.

Who pays OpenAI? You do. Same for Zapier, Make, n8n, Twilio. Spend alerts on the company project (Production best practices). Current rates: openai.com/api/pricing. If the contractor pays, you lose visibility and, later, the keys.

Can I pause? You can pause flows. You can pause a retainer if the contract says so — we work month-to-month after the work is scoped, with notice so you get the runbook and a removed seat. You cannot pause Zapier history retention or a silent failure. If you pause Care, keep the Monday check or you are back on option 1.

What is not included? Takeover: not a full-stack rebuild, not a platform migration, not unlimited new scenarios. Care: not 24/7, not new automations ($400–$900), not vendor invoices, not product/MVP work. First workflow: AI automation for small business, not here.

Keep it running, then keep it boring

The build invoice is a sunk cost. The operating cost is tools + tokens + a human who will notice. Budget the published tier you are actually on, the AI invoice with a cap, and a person with a cap and a reply window. Replace our TCO assumptions with your ledger.

If the freelancer is already gone, do not start with a new Zap. Take the stack back, then decide.

Spofylabs Takeover is $600–$1,200 over 5–10 days. Care is $1,000/month ($750 for 1–2 simple Zaps): weekly check, same-job fixes, monthly one-pager, first reply in one business day. You pay the vendors. Send the tool and what is failing. A short call is enough to tell whether you need a Takeover week, a Care month, or a flow to turn off.