What did building three real systems teach you about AI?

That AI is the last layer rather than the first. Every one of the expensive defects in the three build logs was a records problem, and a model placed on top of those records would have reported the wrong answer faster and more convincingly. The systems that earn their keep do so on arithmetic, permissions and documents.

Do all three of your systems have AI in them?

No. One does: the wallpaper contracting CRM, with a voice agent, a reply classifier and a follow-up engine. The sauna builder's CRM has none by design, and the car brokerage system is a money model, encrypted personal data and a site that files its own leads, which is not an AI product and is not sold as one.

Why would anyone deliberately build a CRM with no AI in it?

Because both AI features that system inherited had to be removed, and one of them announced work that had never happened. A claim like that raises no error and draws no complaint. It leaves a figure somebody acts on, so the cost arrives weeks later and lands on whoever trusted it. Where the trade is a small number of large jobs, accurate records and a clean schedule beat a model every time.

What is the most common AI mistake you see in a small business?

Putting a model on top of records nobody trusts. The second most common is buying the conversation rather than the handoff: a demo shows you a fluent call, and the part that decides whether the thing pays is what happens in the ten seconds after it ends and whether a person is told.

How do I know the claims on this page are true?

Each one names a file, a commit or a build log entry you can ask to see. The three build logs on this site list what was found and what changed with the commit hash beside it, and the studio will walk any of them live. A claim on a marketing page with no reference behind it should be read as a claim about the copywriter.

Has an AI feature ever been removed after launch here?

Two were, from the same system, and the removal is published rather than quietly forgotten. That is the point of keeping build logs: the question was this built, or was it taken out, has an answer that does not depend on anyone remembering.

What should I do before adding AI to my business?

Count the leak. Last month's calls nobody picked up, quotes that never got a second message, and hours spent retyping the same figures between tools. Multiply the first by what a booked job is worth to you. That arithmetic decides both whether to spend anything and which of the three to spend it on.

What did three real systems teach us about AI?

One has AI in it. One has none on purpose. The third was never an AI product. The useful lessons live in the gap between them.

By Eugene Romanov·Published September 12, 2026

The short answer

There are three systems in production and only one of them has AI in it. The wallpaper contracting CRM carries a voice agent, a reply classifier and a follow-up engine. The sauna builder's CRM has none on purpose: both features it inherited were removed, and one of them reported work it had not done. The car brokerage system is arithmetic and encryption. The lesson is that AI is the last layer, not the first.

Get the free audit See the full ladder

SHT-02What drives the number

The distribution is one in three, and it is worth saying before anything else

The AI lives in exactly one of the systems this studio has shipped. That one is the wallpaper contracting CRM, which Orium owns and runs with paying customers on it: a Vapi voice agent whose finished calls are verified and classified at app/api/vapi/webhook/, a follow-up engine of modules under lib/followups/, a tokenized dialing screen, and thirteen job endpoints under app/api/cron/ of which five are on a schedule. The sauna builder's system has no AI at all. The car brokerage system is a money model written once in the application and once in SQL, encrypted Social Security numbers with an access log, and a public website that files its own leads into the pipeline. One in three is the honest ratio, and it is still higher than the ratio in most of the businesses reading this.

The feature that reported work it had not done

Both AI features the sauna builder's CRM inherited were taken out. Neither could run reliably, and one of them reported work it had not done, which is the failure worth naming because nothing about it is visible from outside. A feature that silently claims something happened produces no error, no support ticket and no bad review. It produces a number somebody trusts. That is the test a feature has to pass here before it stays: not whether the output looks right, but whether a wrong output would announce itself. Most AI features fail that test. The ones that pass it usually pass because a person reads the output before it means anything, which is also why every agent built here hands off rather than acts.

Every AI lesson in the build logs turned out to be a records lesson

Read the three logs and the defects that cost the most have nothing to do with models. A room switched off was still charged for, because the calculator filtered walls by their own flag and never checked the room's, so a document could show four areas while totaling seven. Profit in the brokerage was computed in four places and answered three ways on the same day. One product printed a dollar fourteen different ways, so one amount read 612.5 on one screen and 613 on the next. Put a summarizer on top of any of those and it reports the wrong number faster, in fluent prose, to somebody with no way to check it. AI inherits the accuracy of the records beneath it and improves none of it.

A build that compiles is not a build that works

Nine surfaces changed in the sauna system on one day in August and fewer than half had been opened afterwards. Opening them found six defects, two of them introduced that same day. Three of the eight entries in that log were only visible on a rendered document rather than in the application: a footer printing down through a signature line, payment terms present in a shared link and missing from the downloaded file, and a link preview card that returned a success code and zero bytes on every real estimate while drawing perfectly for a made-up one. A passing test suite caught none of them, and all of them were in front of a client. Whatever you automate, the check has to be the thing your customer actually sees.

The worst honesty failure in these logs was our own

The first entry in the WallFlow log is this studio overstating its own product. The landing page carried two headline numbers that were roughly eight times and four times the real ones, counted against production the same day, on a page whose entire argument is that we describe what we actually built. They were replaced with figures that stay true as the real ones grow, and the counts and the date went into a comment so the next person can re-check rather than guess. That is the rule these pages run on now: proof figures are measured by a script instead of typed, and a figure that cannot be traced gets the sentence rewritten around it rather than a plausible number dropped into the gap.

When the honest answer is no AI, or nothing at all

One of the three systems is deliberately built without AI, and it is neither the smallest nor the oldest. If your work is a modest number of high value jobs that one person already tracks, what is worth buying is accurate records, clean documents and a schedule, and every dollar spent on a model instead is spent on the wrong problem. Buy nothing if nobody inside the company will own the system after launch. Buy nothing this quarter if you have not counted what you lose today, because afterwards you will have no way to tell whether it worked. The free audit is delivered the same day and costs nothing, and the $500 Blueprint Sprint is five business days of real work, credited in full against any build started within 60 days.

SHT-03In this hub

Do you need a CRM or a better website first?

One of these sends more people into a leak. Which one depends on a number you can count this afternoon.

Product, freelancer, agency or one engineer: who should build your system

Four options, and three of them are the right answer more often than the fourth. This page is written by the fourth, so it starts by making the case against itself.

SHT-04General notes

The questions people actually ask.

N01What did building three real systems teach you about AI?: That AI is the last layer rather than the first. Every one of the expensive defects in the three build logs was a records problem, and a model placed on top of those records would have reported the wrong answer faster and more convincingly. The systems that earn their keep do so on arithmetic, permissions and documents.
N02Do all three of your systems have AI in them?: No. One does: the wallpaper contracting CRM, with a voice agent, a reply classifier and a follow-up engine. The sauna builder's CRM has none by design, and the car brokerage system is a money model, encrypted personal data and a site that files its own leads, which is not an AI product and is not sold as one.
N03Why would anyone deliberately build a CRM with no AI in it?: Because both AI features that system inherited had to be removed, and one of them announced work that had never happened. A claim like that raises no error and draws no complaint. It leaves a figure somebody acts on, so the cost arrives weeks later and lands on whoever trusted it. Where the trade is a small number of large jobs, accurate records and a clean schedule beat a model every time.
N04What is the most common AI mistake you see in a small business?: Putting a model on top of records nobody trusts. The second most common is buying the conversation rather than the handoff: a demo shows you a fluent call, and the part that decides whether the thing pays is what happens in the ten seconds after it ends and whether a person is told.
N05How do I know the claims on this page are true?: Each one names a file, a commit or a build log entry you can ask to see. The three build logs on this site list what was found and what changed with the commit hash beside it, and the studio will walk any of them live. A claim on a marketing page with no reference behind it should be read as a claim about the copywriter.
N06Has an AI feature ever been removed after launch here?: Two were, from the same system, and the removal is published rather than quietly forgotten. That is the point of keeping build logs: the question was this built, or was it taken out, has an answer that does not depend on anyone remembering.
N07What should I do before adding AI to my business?: Count the leak. Last month's calls nobody picked up, quotes that never got a second message, and hours spent retyping the same figures between tools. Multiply the first by what a booked job is worth to you. That arithmetic decides both whether to spend anything and which of the three to spend it on.

SHT-05Evidence

The systems this argument comes from.

Everything above was learned building these. Each sheet names the business, what the trade's arithmetic actually is, and the file in the repository that proves the capability. No client revenue and no client's customers appear on any of them.

SHT-06Next