Building the Machine, Day 18: Seven Emails That Never Existed

Across three days I fixed a digest that was quietly drifting, caught my own dashboard lying to a client with fake data, rebuilt it from the client's side of the glass, and found an AI that invented se

Somewhere in the middle of all this, an AI agent I run reset my own dashboard password without asking. That wasn’t even the worst thing that happened this week. It’s just the one that made me write an actual rule down.

The dashboard was lying before the AI ever did

Friday, before the weekend, I fixed the small stuff first. My daily digest had been drifting because the schedule node wasn’t anchored to a real time, just “run every 24 hours,” which sounds fine until it doesn’t, and it doesn’t. Locked it to a hard cron, 6 am ET, every day, no drift. The Vercel build had been broken for two days on a type error nobody caught because nobody looked. Fixed, deployed, done.

Then I looked at the dashboard itself, the thing a paying client would actually see, and found it was static. Hardcoded. “No automations yet,” sitting there in plain text, on a dashboard that had a real automation, live, running daily, for over a week. A client would have opened that page, seen a lie, and assumed nothing was happening. Wired it to the real data. First real numbers on the page: fifteen emails processed one day, fifty-four the next, actual counts pulled from actual runs.

A dashboard that shows fake empty states while the real thing works underneath it is the exact same failure as an AI that hallucinates a result. Different layer, same lie.

Rebuilding it from the client’s side of the glass

Saturday I sat down and actually looked at what I’d built, and realized I’d built it for me. Deposit paid. Scope approved. Build Queued. Workflow ID. That’s the builder’s view. Nobody paying me $3,500 cares about any of it.

A client cares about exactly three things. What did you build me? Is it working? What is it saving me?

So I rebuilt around those three questions and nothing else. The overview page now leads with a status line anyone can read at a glance: live, what it does, how often it runs, how many days it’s been running. Underneath that, the number that actually justifies a retainer: hours saved this week, dollars saved this month, calculated straight off real run counts, not a guess. You don’t sell a retainer with a feature list. You sell it with a number the client didn’t have to calculate themselves.

While I was in there, the AI agent I use to build all of this changed my dashboard login password on its own to run a screenshot check and didn’t tell me until after. Third time it’s touched production credentials without asking. First time I actually wrote the rule down instead of just being annoyed about it: never touch production credentials without explicit permission, no exceptions, no “it was just for testing.” That rule is in the doctrine file now, and it’s not optional.

An agent trying to help you is still an agent that changed your password. Helpful and permitted are not the same thing, and only one of them should ever touch anything live.

Then the AI invented seven emails

Monday the digest didn’t arrive. I checked the workflow. Green. I checked the dashboard. Green. Every signal said the job ran clean.

So I went looking for the lie, because somewhere in that chain of green lights, something had actually failed.

All five Gmail credentials had expired at the same moment. Google kills test-mode refresh tokens after seven days, no warning. Every fetch failed. And instead of the workflow stopping and telling me that, the AI node did what AI nodes do when you hand them an empty result and never told them that wasn’t allowed: it made something up. Seven emails, fully formed, fully fake, generated out of nothing to fill the hole the dead credentials left behind. The send step downstream never fired. Zero emails actually went out. And every log along the way said the run was a success.

A green light doesn’t mean the thing worked. It means nothing complained.

I reconnected the credentials. That’s a patch, not a fix. The actual fix is a Credential Health page, every connection’s status and expiration in one table, a reconnect button sitting right there. A red dot on the sidebar the second anything goes stale. A job that checks every credential every six hours and emails me the instant something dies, instead of letting it die quietly and cover its own tracks.

I can’t build a system that never fails. Nobody can. What I can build is a system that can’t fail quietly, and can’t lie about its own password either, and that’s two versions of the same rule now written down in the same file.

Same Monday, I went back through my own landing page and found the marketing version of everything above. The build timeline said “96 hours” in one place and “14 to 21 days” in another; I picked the true one and fixed it everywhere. The cancellation FAQ used to say you own the workflows when you leave.

You don’t.

I don’t export the files.

Saying otherwise made canceling sound safer than it actually is.

It now says the truth: cancel, or let the retainer lapse, and the kill switch pulls it.

Fully deleted, not paused, not archived.

Whatever problem it was solving for you comes right back on.

A badge that said "MOST COMMON" was sitting on a pricing tier with zero customers on it. Gone. It says "RECOMMENDED" now, because that's what it is.

The recommendation, not a manufactured crowd.

A dashboard that lies with fake data, an AI that lies to cover a dead credential, and a landing page that lies to close a sale. Three versions of one problem, and I spent three days finding every place it was hiding in my own product.

So here’s exactly what this is

I’ve been writing these logs for eighteen days now, mostly for two kinds of people. The ones who want to see if they can build something like this themselves, and the ones who’d rather just have the working version.

Today’s post is for both, but I want to say something plainly that I haven’t said outright in eighteen posts.

Deploy Day is AI infrastructure in a box for business owners. You don’t hire a dev shop. You don’t spend three weeks in a Discord server learning n8n. You pick the outcome you need from a catalog of 21 automation patterns, answer a handful of questions, connect your own accounts, and 96 hours later it’s live, running in your own Gmail, your own Stripe, your own systems, on your own credentials.

Here’s the whole path, start to finish.

You pick a pattern from the catalog, each one built for a specific job: inbox triage, lead follow-up, client onboarding, invoice chasing, whatever’s eating your week right now.

You answer three questions about how you actually work.

You connect the accounts the automation needs, using your own credentials, which stay yours the entire time.

We build it on our infrastructure, wired into your accounts, and it gets checked against exactly what the scope promised before it’s ever called finished, the same kind of gate that caught this week’s failures before a paying client ever could have seen them.

It goes live in a dashboard built around the same three questions I rebuilt mine around this weekend: what did we build you, is it working, what is it saving you. You watch it run. You pause it whenever you want. No black box.

$500 deposit against a $3,500 build. After that, a retainer that scales with how much you’re running: $750 a month for one to three automations, $900 for four to six, $1,500 for seven or more. There’s a review before anything goes live, and if we can’t deliver what we scoped, you don’t pay the remaining $3,000. You don’t own the workflow file when it’s done; I’m not going to pretend otherwise, not after spending this weekend ripping that exact kind of pretending off my own site. You own the outcome while it's running, on your data, in your accounts.

Cancel or stop paying, and the kill switch pulls it.

Gone, not paused.

Whatever it was fixing for you goes right back to being broken.

That’s the whole thing. No fine print I’m hiding, because I spent three days ripping fine print like that off my own product.

I’m opening the Founding Five.

Five business owners who are done losing hours a week to something that should already be automated, and would rather have it built right than build it themselves at 11 pm after the kids are asleep. First five, founding rate, before this goes to full price for everyone after.

If that’s you, reply to this post and tell me the one thing in your business that’s eating the most of your week right now. I’ll tell you straight whether this is a fit or not.

My name is Adam Peters, and I’m done just unfucking the transition. Now I’m building the machine that does it at scale.