01Workflow automation · the whole system, not just the sheet
I build the automated version of the parts of your week that are pure remembering — the job that finished and never got invoiced, the invoice nobody chased, the Monday numbers nobody has time to assemble. Formulas do the arithmetic. AI drafts the words. You make every call that leaves the building.
02Try it — right here
This is a working miniature of the follow-up system I build. Overdue invoices get a drafted email in a tone that matches their age — and one that’s too fresh to chase gets deliberate silence. Approve the ones you’d actually send. Sample data — every number and business name invented.
Every workflow on this page runs this same five-step shape — the Money Leak Autopilot, the Lead Concierge, the invoice chase queue, the review flywheel, the Monday Pulse.
This queue is interactive — with JavaScript on you can approve each draft and watch what would leave. The four drafts, their tone tiers and the rule behind each one are all below either way.
No draft. Nine days is too fresh — a nudge now reads as twitchy, not attentive. The system waits on purpose.
Under fifteen days the invoice is still inside a normal payment habit, so the rule is silence: no draft is written at all.
“Hi — quick nudge on invoice 2123 from last month. No rush if it’s already queued up on your end…”
Twenty-two days is one cycle late, so the draft assumes good faith and offers an out.
“Hi — invoice 2118 is now a month past due. Could you let me know when payment is scheduled?…”
Thirty-seven days is past a full billing cycle, so the draft asks for a date rather than offering one.
“Hi — invoice 2101 is 68 days outstanding. I’d like to resolve this week; here are the options…”
Past sixty days the money is at real risk, so the draft names a deadline and puts options on the table. It is still a draft.
Run ledger
Layer 05, live: every run leaves a receipt.
…and that’s where this demo stops, by design. In the real system that approve click is the moment they actually go out — and the drafts you didn’t approve simply wait for the next queue. Nothing on this page sent anything anywhere.
The tone tiers aren’t decoration. A nudge at nine days reads as twitchy; the same words at sixty days read as overdue. The system picks the tier. You still pick whether.
03The problem
The job itself usually goes fine. What quietly eats the week is everything around it: retyping the same customer into three places, remembering which quote never got an answer, invoicing the job that finished last Tuesday, chasing the invoice from six weeks ago, and assembling the numbers into something you can act on.
Every one of those is a handoff. Every handoff depends on somebody remembering. That’s the leak.
Automation done right doesn’t replace your judgment — it replaces the remembering. The system watches the handoffs, does the drafting and the arithmetic, and brings you a short list of decisions instead of a long list of chores.
The system decides who and what tone. You decide whether.
Here’s the shape every one of my builds takes.
04How every build is wired
Anyone can connect two apps together. What makes one of these hold up on week forty is the shape it was built in. Every system I build has the same five layers, and each one has a rule I don’t break — including on the days it would be faster to.
Something has to notice. A schedule that wakes up Monday morning, a form submission, a new row in a sheet, an email landing in an inbox. The rule: the sensor is read-only. It looks at your data and never writes back to it. If you unplugged my system tomorrow, your spreadsheet would be exactly the file you’d have had anyway.
Every total, every days-overdue count, every “this is up 12% from last month” is computed by an ordinary formula in your own sheet, before any AI is involved. The rule: the AI never does arithmetic. This isn’t caution for its own sake — during an early build the model added up a column itself and got it wrong. The fix wasn’t a better prompt. It was removing the model’s opportunity to do arithmetic at all. Every number the AI writes is a number that already existed in a cell you can click on.
Now the AI gets the verified numbers and does the thing it is genuinely excellent at: turning them into three paragraphs a human would actually read, and drafting the follow-up emails in a tone that fits the situation. The rule: narrative only. Every figure in what it writes traces back to a cell.
Anything that would leave your business with your name on it stops here and waits. The drafts stack up. You read them, you delete the one that’s wrong, you press send on the rest. The rule: no exceptions — not for “obvious” ones, not for the third reminder, not at 2am. This is the layer everyone else skips, and it’s the reason you can hand this thing your customer list without lying awake. You are the only one allowed to press send.
try it ↑Every time the system runs, it writes a row: what it read, what it computed, what it drafted, what you approved, what went out. The rule: the audit trail is a deliverable, not an afterthought. Six months from now, when someone asks why a customer got that email, the answer is a row, not a memory.
Read-only in. Formulas for the math. AI for the words. A human at the door. A receipt for every run.
05Free · nothing uploads
Every one of these runs entirely inside your browser. There is no server to send your file to. You don’t have to take my word for it — open your browser’s network tab, drop a file in, and watch: after the page loads, zero requests. That’s architecture, not a privacy policy. No signup, no email, no account.
Paste a customer or vendor list — any number of lines — and this page will tell you how many real customers are in it. Same normalization as the full tool: case, punctuation and legal suffixes ignored, then near-miss spellings grouped for a human glance.
sample lines — clear the box and paste your own
third-party requests: measured on load
This box needs JavaScript. The full tool is at /duplicate-customer-finder/ and works the same way.
Same engine as the full tool — that one takes a whole CSV export and works out which column holds the names.
Drop any CSV and get an honest report: columns that carry no information, status fields that never vary, mixed date formats, numbers stored as text that SUM() silently skips, duplicate rows. Then download the file with the judgment-free problems already fixed.
On its own sample it finds six problems and the automatic fix takes it to two — and the two that remain are exactly the ones a script must not touch.
Two exports every business already has — the work log and the invoices — lined up so only the mismatches show: finished jobs nobody billed, invoices sitting unpaid past 60 days.
It matches on exact ID only, and says so on the report. Nothing fuzzy, nothing guessed.
Acme Roofing. Acme Roofing LLC. acme roofing inc. Three rows, one customer — inflating your count, splitting your revenue history, and sending the mail merge three times.
Every group states why it grouped: identical after ignoring case, punctuation and legal suffixes, or a near-miss typo worth a human glance.
For anyone who runs on a roster — senior living, clinics, restaurants, facilities. Finds days a role is uncovered, staff crossing 40 hours, roles only one person can fill, and turnarounds too short to be safe.
It reads your columns itself and shows you which ones it used. I built this one because I schedule people for a living.
All four, plus everything else I’ve built, are indexed at /builds/ — and there’s a single drop zone at /check/ that works out which check your file needs and runs it.
Why the tools stop where they stop. Every one of them ends at the point where a script would have to guess — which duplicate is the real customer, what a blank cell means, who should cover Thursday. Those decisions need a person who can ask you a question. That’s what the $300 cleanup and the free demo are. A tool that guessed there would be faster and worse.
06The $650 rung, running
/demo/ is a real Google Sheets dashboard built as an Apps Script web app — revenue, jobs, overdue invoices, tabs you can click, a chart you can tap. No login, no signup. It’s the $650 Automated Dashboard doing its actual job.
It also refreshes itself. A weekly trigger re-runs the whole build every Monday morning, which is why the numbers you’re looking at aren’t from whenever I last remembered to update the page. The data in it is invented sample data — the machinery is not.
There’s a second one at /staffing-commission-dashboard/: a live commission dashboard for a staffing desk that settles splits, draws and fall-off clawbacks by rule instead of by argument. Built for a vertical where the math is the whole disagreement.
07The reporting half
Revenue grew 47% from week 1 to week 12, and new customers rose from 38 to 61 over the same stretch. The growth is broad rather than lumpy: no single week carried it. Worth watching: acquisition is outpacing revenue per customer, so the next lever is ticket size, not traffic.
The rule on display: every figure above was computed by formulas before the AI saw a word of it. The model only writes the sentences — it is never trusted with the arithmetic.
08The flagship, step by step
Four steps, in order, on the same invented sample the free tool ships with. Nothing here is a screenshot — the arithmetic below is the arithmetic the tool actually does.
Cedar Field Services is an invented company. Every job, invoice and dollar in this section is made up — the arithmetic is real, the business isn’t.
09The part nobody shows you
Most people show you finished work. Here’s the machinery that keeps it honest.
/build-log/ walks through the multi-tenant reporting pipeline end to end — including the dashboard that displayed MONEY LOST ON OVER-BUDGET JOBS: $0 directly above two jobs flagged over budget at −$500 and −$600, because five summary formulas were off by one column. It survived review because my own build log computed the totals independently and printed the right answer while the sheet showed the wrong one. Two sources of truth that never met. Each of the four left a permanent check behind.
When a loop setting looked like it might collapse a multi-client run into one send, I stopped reading documentation and ran it live: two sends, six seconds apart, distinct narratives per row, in the execution record. Now it’s a fact I read off a log instead of a belief I took from a doc.
A script scores every workbook and every page for defects and runs on a git pre-push hook — if the score goes up, the push is refused. On the last recorded run the score was zero.
Twelve sample workbooks, each generated by its own Python script rather than assembled by hand, all reading brand colours from a single source of truth. Nothing drifts, because nothing is typed twice. This page is built the same way — the file you are reading is generated, and the build refuses to write it if any of its own checks fail.
flatline finds signals that carry no information and jobs that produce nothing — a check that fires on 100% of rows, a scheduled task that reports success while writing nothing. I built it after a real scheduled job of mine discarded four days of data while reporting success every run. MIT licensed, 108 tests. hearth is a national crisis-resource directory built on federal data and refreshed automatically, with a public health endpoint that reports freshness per source so a stale feed can’t hide behind a fresh one.
The interesting work isn’t wiring the happy path — a model does that before lunch. It’s the four times I didn’t trust it yet.
10What I’d build you
Each of these is the same architecture wearing a different job. One is being built right now; the other four are designs I’ll build the first time somebody needs one. None of them is running for a customer yet, and I’ll always tell you which is which.
Every week the system reads your work log and your invoice export, the sheet’s own formulas flag the jobs that finished and never got billed plus everything aging past 30 and 60 days, an AI writes you a short digest, and the chase emails draft themselves in a tone that matches each invoice’s age. Then everything stops and waits for you.
You’ve already used the miniature version — the approve panel above is this system’s follow-up half, running on invented data.
An inquiry arrives, by form or by inbox. Within minutes it’s a tracked row, a notification, and a drafted personal reply waiting for your approval — and if it’s still unanswered in 24 hours, the system tells on itself. Most lost leads aren’t lost to competitors. They’re lost to Tuesday.
I’m wiring this to my own contact form first, because a demo I can’t run on myself isn’t a demo.
The chase spreadsheet with a live layer on top: the sheet that chases for you. Overdue invoices move through 15 / 30 / 60-day tone tiers, drafts stage in your Gmail, you approve.
A job gets marked complete. The system waits the right number of days, drafts the review ask in your voice, and holds it for your approval. The ask that actually gets sent is the one that doesn’t depend on you remembering to send it.
Monday morning, the system reads your live numbers, the formulas compute everything, and an AI writes the three paragraphs you’d actually read: what moved, what needs attention, what to do next. That’s the panel further up, pointed at your business.
Four of these are designs, not builds. They get built for the first person who wants one — I’d rather label them honestly than let a status chip imply otherwise.
All of it runs on tools you already have — Google Sheets, Gmail, your existing forms. There’s no new software for your people to learn, which is the reason the spreadsheet won in the first place.
11Free templates
Three working Google Sheets templates, no cost and no email required: an executive KPI dashboard, an expense tracker, and a staffing commission tracker. Download, open, use. If one of them solves your problem and you never contact me, the template did its job.
12What it costs
Send me one messy file — a job log, an invoice export, an intake sheet, a schedule. I’ll send back a working automated version within a day. Yours to keep whether you hire me or not. This is how every engagement starts, and it’s the whole reason the tools on this page are free too: I’d rather you judge the work than the pitch.
Start here — free →Your messiest sheet, rebuilt into one you can trust. Formulas that don’t silently skip rows, dates in one format, statuses that mean one thing. 24–48 hours.
Starts with the free demo →I build one workflow in your own account — your Zapier, your Google, your Gmail — document it, and hand it over. You own it completely. If we never speak again it keeps running. No subscription, no lock-in, no dependency on me.
Starts with the free demo →Your numbers on one live page that updates itself. The /demo/ you can click on this site is this rung, running.
Starts with the free demo →The Monday report that writes itself, plus approve-gated follow-ups on money you’re owed. Setup, then I run it, watch it, and tune it.
Starts with the free demo →Intake to invoice to follow-up, built end to end around how your business already works.
Starts with the free demo →Two honest notes. First: the free demo is not a trick — you keep the artifact either way, and if the answer is “you don’t need me for this,” I’ll say that in the same email. Second: I’m early enough that the first few engagements are priced to earn an honest review rather than to maximize the invoice. That’s a real advantage to being first, and it doesn’t last.
13Who’s building it
I’m Colin McCarthy. I build and run Automated Workflow out of Gainesville, Florida, and I work with small businesses anywhere, because all of this is remote.
Before I wrote systems for operations, I ran them. I still do — I manage dining operations at a continuing-care community affiliated with the University of Florida: scheduling, inventory, vendors, and a staff who did not sign up to learn new software. That’s why the Shift Coverage Check exists, and it’s why I’m allergic to automation that adds a tool instead of removing a chore. I’ve held the clipboard the software is supposed to help.
I’m certified through Anthropic Academy — six credentials — and more usefully, I use the same AI tooling on my own business every day before I point it at anyone else’s.
I have zero paying customers. I’m early, and that’s why the first engagements are priced to earn a review rather than a margin. What I have instead is work you can inspect without asking me anything — and the first thing I’d build you costs nothing.
14The questions I actually get
15The whole pitch, honestly
Tell me in one sentence. I’ll tell you in one email whether it’s automatable, what it would cost, and what I’d build first — and if the answer is that you don’t need me, I’ll tell you that too. The first artifact is free either way.
Or just press a tool. Nothing uploads, nothing signs you up, and you’ll learn more about how I work in ninety seconds than this page can tell you.
colin@automatedworkflowllc.com · (703) 939-1174 · Gainesville, FL