Custom workflow automation
Custom workflow automation that keeps running when a step fails.
I map the process, pick the tool your volume justifies (n8n, Zapier, Make or plain code), and build it in your accounts with retries, a dead letter queue and alerts that reach a named person. You get the run log, the runbook, and before and after numbers measured on your own work.
- Built in your accounts
- Alerts to a named person
- Measured before and after
Run 1041
Succeeded03:06.4 Run 1041 finished. 7 steps, 0 retries.
The problem
Automations rarely break on demo day. They break quietly, months later.
A provider starts rate limiting, a field gets renamed, a call times out halfway through a write. Most of the time the tool is fine. What is missing is a plan for the failed step: does it retry, does it write twice, who hears about it, and how does the run get finished?
Two more costs hide in the design. The platform meter counts runs, actions or module credits depending on the tool, so the same workflow can cost very different amounts. And an AI step is right most of the time, not all of the time. I design for all three before I build anything.
Nearly 1 in 3
respondents to McKinsey’s global AI survey reported consequences from AI inaccuracy. An AI step needs a check behind it.
McKinsey, state of AI, 2025~20%
of respondents say AI operating costs, token costs included, have constrained their AI use. The meter matters.
McKinsey, state of AI, 20262 tasks
for one MCP tool call on Zapier, whose pricing now says AI model tier, code runtime and connector type change what a step uses.
Zapier pricing, 2026See it fail
Break a step on purpose and watch what happens
A lead workflow of the kind I build: a webhook, enrichment, an AI fit score, a branch, a human approval, the CRM and Slack. Run a lead, then make a step fail. Most automation pages promise retries and alerts. Here they are, running.
Retry with backoff
Dead letter queue
Alert to a person
Recovered
Run log
simulated clock
- 00:00.0INFORun 1041 started: webhook from the signup form, lead 4821
- 00:00.0INFOEnrich: looking up the company
- 00:00.4OKEnrich: company found, 48 staff, software
- 00:00.4INFOAI classify: scoring fit against the written rules
- 00:01.7OKAI classify: fit high (0.86). Output matches the schema.
- 00:01.7INFOBranch: score above 0.70, so a person approves it
- 00:01.7INFOApproval card sent to the sales lead in Slack
- 03:05.7OKApproved by the sales lead
- 03:05.8INFOCRM: creating the deal, idempotency key 1041:crm
- 03:06.2OKCRM: deal created
- 03:06.4OKSlack: #sales told a qualified lead is in
- 03:06.4OKRun 1041 finished. 7 steps, 0 retries.
What stood out was that everything worked in production. Not just a demo. Monitoring, error handling, fallbacks. He thinks about the stuff that breaks at 2am on a Saturday.
Worth it?
What the workflow is worth, and the most it should cost
Put in your own numbers. You get the hours back, the break even build budget at 12 months, and what each platform’s meter reads at your volume. No price of mine is in here, on purpose.
The task today
Times a week
Minutes each
Loaded cost per hour
Example: the BLS median software developer wage (May 2025) times the employer cost load for professional occupations (June 2026), about $95 per hour worked. Most operations roles cost less. Use your own.
Error rate
Minutes to fix one
After automation
Share the workflow handles alone
70% automated30% to a person
Minutes a person still spends per automated task
Approvals and spot checks. The canvas above has one approval step; it is not free.
Run cost per month
Platform plan, hosting and AI tokens. The meter panel shows what each platform counts at your volume.
The workflow and your budget
Steps, with trigger
AI steps
Build budget to test
Your number, or any quote you have. It is not my price; I quote after a call.
On your numbers, over 12 months
About 6.0 hours a week back, worth $29,739 a year.
After $720 a year of running cost, a build can cost up to $29,019 and still pay for itself within a year. Hold any quote against that number, mine included.
Hours a year
313
Break even
$29,019
Run cost
$720/yr
Payback
3.3 months
Of the 313 hours, 291 come from the task itself and 22 from errors nobody has to fix by hand. 4,368 of 6,240 tasks a year run without a person.
Cumulative saving against your build budget
Net of running cost, from the day it goes live. Where the lines cross, the build has paid for itself.
Assumes a flat saving from launch. In practice the first weeks run alongside the manual process, so the line starts a little later than this.
What each platform’s meter reads at 520 runs a month
Same workflow, three different meters. This is the cost that grows with volume, so it belongs in the decision before anything gets built.
n8n
520 executions
One execution per whole run, however many steps it has. $20 a month on Starter: up to 2,500 executions, hosted by n8n, billed annually.
The free Community Edition is self hosted: no execution fee, but you pay for the server and its upkeep. n8n pricing, read Oct 10 2026
Zapier
3,640 tasks, at least
One task per action that succeeds. The trigger is free. From $19.99 a month. The tier that covers 3,640 tasks is set on Zapier’s pricing selector, so I do not print a number for it.
Your 1 AI step can use more than one task per run: AI model tier, code and connector type change the rate, and one MCP tool call uses 2 tasks. Zapier pricing, read Oct 10 2026
Make
4,160 credits, at least
One credit per module run, the trigger included. Some modules use 2 or 10. Price not shown: I could not verify Make’s plan prices from its own pricing page.
Extra credits cost 25% more than the ones in your plan. Make’s own AI bills by tokens; with your own AI provider connected, you pay the provider directly. Make help: credits, read Oct 10 2026
AI steps also cost model tokens, billed by your AI provider on top of n8n or Zapier. The AI agent cost calculator estimates them per run. Make’s billing rules are from its help centre, including the November 2025 plan changes.
Exactly how the numbers are worked out
Tasks a year = times a week x 52
Automated tasks = tasks a year x share the workflow handles
Hours back a year = automated tasks x (minutes each minus review minutes) / 60
+ automated tasks x error rate x minutes to fix one / 60
Labour value = hours back x loaded cost per hour
Net saving a year = labour value minus run cost per month x 12
Break even budget = net saving a year: the most a build can cost
and still pay back within 12 months
Payback = your build budget / (net saving a year / 12)
Runs a month = times a week x 52 / 12
n8n executions = runs a month
Zapier tasks = runs a month x (steps minus the trigger), a floor
Make credits = runs a month x steps, a floorEvery starting value is an example, not a benchmark. The hourly default is $65.38 (the BLS May 2025 median software developer wage of $135,980 over 2,080 hours) times 1.4455, the BLS employer cost load for professional occupations, June 2026. Platform counting rules come from each vendor’s own pages, read Oct 10 2026. Replace every number with your own before you decide anything.
How it runs
About three weeks to live, then a month of measuring.
Every phase ends with something you can open or sign off. The failure plan is written before the build, not bolted on after the first incident.
- 01
Days 1 to 3
Map the task and measure it
I watch the work being done, collect about twenty real examples, and write down the baseline: how often it happens, minutes per task, how often it goes wrong.
You getA workflow map with the failure paths drawn, plus the baseline numbers
- 02
Days 4 to 6
Pick the tool and plan the failures
n8n, Zapier, Make or plain code, chosen on your volume, steps per run and who will maintain it. For each step: how it retries, what its idempotency key is, where a failed run parks, and who gets the alert.
You getA one page design with the monthly platform bill at your volume and the failure plan
- 03
Days 7 to 15
Build in your accounts against real examples
Built in your platform account under service accounts you own. The recorded examples are replayed through it, including the ugly ones, until every one lands where it should.
You getA staging workflow that passes the recorded examples, with a run log you can read
- 04
Days 16 to 21
Shadow run, then switch over
It runs next to the manual process for a few days and the outputs are compared. Then it goes live with the alert channel wired to a named person.
You getThe live workflow, alerts tested by forcing a real failure
- 05
30 days after launch
Measure again and hand over
The same numbers as day one, measured again, so the saving is a figure rather than a feeling. Then the runbook, the exported workflow definitions and a recorded walkthrough.
You getA before and after report, the runbook and the walkthrough recording
Workflow, AI step or agent
Most jobs need less AI than they are sold
I pick the simplest shape that does the job, because every step of judgement you add is a step you have to check. Here is the line I draw.
A plain workflow
The path fits a flowchart and the inputs are structured: form fields, webhooks, rows in a sheet.
Cheapest to run and easiest to test. A lot of what gets sold as AI automation should be exactly this.
New order, invoice in accounting, row in the sheet, message in Slack.
A workflow with one AI step
One step needs judgement on messy input: sort an email, read a PDF, score a lead.
The model’s answer is checked against a schema, and anything it is unsure of goes to a person. The canvas above is this shape.
Inbound lead, enrich, AI fit score, approval when it matters, CRM.
An agent
The path itself changes from case to case, and the system has to decide what to do next.
That is a different build, with evaluation sets, approval levels and spend limits. It costs more to run and to trust.
Researching a prospect across several tools, then drafting the next step.
AI agent development“Many use cases positioned as agentic today don’t require agentic implementations.”
Stanford’s AI Index finds productivity gains concentrate in structured, measurable work with clear quality checks, and shrink on work that needs deeper judgement.
Time saved, honestly
What the research says about saving time, and what I build in
No independent study measures the time saved by workflow automation itself, and vendor surveys do not count. The credible numbers are about AI assistance, and they are modest. So I measure your task before and after, and build every workflow to fail safely.
What the research says
2.2%
of all work hours in the United States were saved by generative AI assistance by Q2 2026, up from 1.6% in Q3 2024. Self reported, economy wide, and approximate by the authors’ own account.
Federal Reserve Bank of St. Louis, 202614%
more issues resolved per hour by 5,179 support agents given an AI assistant: 34% for novices, minimal for the most experienced.
NBER, Brynjolfsson, Li and Raymond, 202326 min
a day of self reported time saved across 20,000 UK civil servants using Copilot. The trial could not see how that time was spent.
UK Government Digital Service, 2025Unproven
A UK Department for Business and Trade evaluation found no evidence that reported Copilot time savings became better productivity, and some tasks took longer with it.
UK Department for Business and Trade, 2025What I do about it
Retries with backoff
Every external call retries with growing waits and a cap, so a short outage does not turn into a lost record.
Idempotency keys
Each write carries a key, so a retry or a replay can never create a second invoice, deal or email.
Dead letter handling
A run that still fails parks with its full payload and replays in one click. Nothing silently disappears.
Alerts to a named person
Not a shared inbox nobody reads. One owner gets the failing step, the error and the replay link.
Run logs a person can read
Each run says what it did in plain words, so “what happened to this lead?” has an answer without an engineer.
A handover runbook
What each alert means, what to do about it, and which changes are safe to make yourself.
What ships
What you have at the end
The workflow, in your account
On the platform that fits, or in plain code. Definitions exported to your repository, so nothing lives only in my head or my login.
Failure handling you can see
Retries with backoff, idempotency keys so nothing is written twice, a dead letter queue, and a replay button for the runs that land there.
Alerts and a readable run log
A failed run reaches a named person with the step, the error and a replay link. The log says what happened in plain words.
Credentials you hold
Service accounts you own, secrets in the platform credential store or a secrets manager. Never pasted into a step, never one person’s login.
Before and after numbers
The baseline from day one and the same measures 30 days after launch, so you know what it actually saved.
A runbook and a walkthrough
What to do when an alert fires, which changes are safe without an engineer, and a recording of the whole thing.
The stack
Tools picked per job, not per partnership
I am not a reseller for any of these. The right one depends on your volume, your team and where your data has to stay.
n8n
Cloud or self hosted, billed per whole run
Zapier
Low volume jobs your team already runs there
Make
Visual scenarios, billed per module credit
Python or TypeScript
When logic, volume or testing outgrow a tool
Claude and OpenAI APIs
The AI step, with a schema check on every answer
Postgres or a queue
Dead letter queue and idempotency records
Slack or email
Alerts to the named owner, with a replay link
Your CRM and apps
HubSpot, Salesforce, Shopify, Sheets and more, through their APIs
Your options
Four ways to automate a workflow, compared honestly
Off the shelf apps are often the right answer. Here is where each option wins and where it costs you.
| Factor | Off the shelf apps, set up yourself | Custom, with me | Agency | In house hire |
|---|---|---|---|---|
| Who builds it | You or a colleague, often inside one person’s login. | Me, start to finish, in your own accounts. | A team behind an account manager; seniority varies. | Your hire, once you have recruited one. |
| When a step fails | Whatever the platform does by default, unless someone designs more. | Retries with backoff, idempotency keys, a dead letter queue and an alert to a named person. The demo above shows it. | Usually promised in the proposal. Ask to see it run. | Depends on who builds it and how much time they get. |
| Monthly platform bill | Grows with volume, rarely estimated up front. | Estimated at your volume, on each platform, before the build. | Ask for it at your volume before you sign. | The same bill, plus a salary. |
| Ownership and keys | Yours, though often tied to whoever set it up. | Your accounts and service accounts from day one, exported to your repository. | Check the contract: some build in their own instance. | Yours. |
| Time to the first workflow | Hours, for simple two step jobs. | About three weeks, including a shadow run. | Depends on their queue and process. | Recruiting comes first. |
| After launch | Whoever notices it stopped. | A runbook and an alert owner. Support is optional, scoped in hours. | A retainer or a new statement of work. | Continuous, at full salary whether or not there is work. |
Who builds it
- Off the shelf apps, set up yourself
- You or a colleague, often inside one person’s login.
- Custom, with me
- Me, start to finish, in your own accounts.
- Agency
- A team behind an account manager; seniority varies.
- In house hire
- Your hire, once you have recruited one.
When a step fails
- Off the shelf apps, set up yourself
- Whatever the platform does by default, unless someone designs more.
- Custom, with me
- Retries with backoff, idempotency keys, a dead letter queue and an alert to a named person. The demo above shows it.
- Agency
- Usually promised in the proposal. Ask to see it run.
- In house hire
- Depends on who builds it and how much time they get.
Monthly platform bill
- Off the shelf apps, set up yourself
- Grows with volume, rarely estimated up front.
- Custom, with me
- Estimated at your volume, on each platform, before the build.
- Agency
- Ask for it at your volume before you sign.
- In house hire
- The same bill, plus a salary.
Ownership and keys
- Off the shelf apps, set up yourself
- Yours, though often tied to whoever set it up.
- Custom, with me
- Your accounts and service accounts from day one, exported to your repository.
- Agency
- Check the contract: some build in their own instance.
- In house hire
- Yours.
Time to the first workflow
- Off the shelf apps, set up yourself
- Hours, for simple two step jobs.
- Custom, with me
- About three weeks, including a shadow run.
- Agency
- Depends on their queue and process.
- In house hire
- Recruiting comes first.
After launch
- Off the shelf apps, set up yourself
- Whoever notices it stopped.
- Custom, with me
- A runbook and an alert owner. Support is optional, scoped in hours.
- Agency
- A retainer or a new statement of work.
- In house hire
- Continuous, at full salary whether or not there is work.
Shipped, not demoed
Automations I have shipped
Six people were doing nothing but data entry. He automated the whole thing. Took about two weeks. Those six people now do work that actually matters and we process orders faster than before.
He plugged AI into our dispatch system without breaking anything. Orders route themselves now. Same team, 3x the volume. Wish I had found him a year earlier.
Straight answer
Is your task worth automating?
A good fit when
- It happens dozens of times a week or more and follows rules you can write down.
- The systems involved have APIs: your CRM, forms, inbox, accounting, Slack.
- You can name the person who should get the alert when it fails.
- You want it in your own accounts, documented, and measured against a baseline.
Not the right fit when
- It happens a few times a month. A checklist is cheaper than any automation.
- Every case needs real judgement, like legal review or medical triage. Keep a person on it.
- Your current Zaps work and cost little. Keep them; I will tell you so.
- You want a promised hours saved figure before anything has been measured.
Questions
What people ask before a build starts
I quote a fixed scope after a short call, because the price depends on how many systems are involved, how messy the inputs are and what a failure costs you. Before that call, use the calculator on this page: it gives the most a build can cost and still pay back within 12 months on your own numbers. Hold any quote against that figure, mine included.
Related
Operations Autopilot
The fixed scope version: one workflow, about three weeks.
AI agent development
When the path changes per case and needs an agent.
n8n consultant
When the workflow should live in n8n, cloud or self hosted.
The AI Audit
Not sure which process to automate first? Start here.
Build vs buy calculator
A custom build against a subscription, over time.
n8n vs Zapier vs Make cost
What each platform bills for your workflows and volume.
n8n vs Zapier
The two most common platforms, side by side.
Flowise vs Dify vs n8n
Which builder suits AI heavy workflows.
n8n AI agent workflows
AI steps inside n8n, from practice.
AI agent cost calculator
The tokens an AI step burns per run.
Show me the task. I will tell you if it is worth automating.
A short call, your numbers from the calculator, and a straight answer, including when the answer is to stay on Zapier or leave it manual.
Book Meeting