AI writes code that runs. It does not choose the cheap way to do things, it chooses the common way. Two senior architects read your app and tell you what is quietly costing you every month, what it would cost instead, and what happens when you have ten times the users.
A worked example
One example, assembled from things we have actually found, not a specific client. Every number here is one we have seen on code nobody would call broken. Your numbers will differ; the pattern rarely does.
What it cost, as found
€610a month
€85 after the fixes, for the same product, same features, same users
What the report says
Every query returns the right answer, every page renders, every test passes. The app simply asks the database for the same thing eight hundred times instead of once, because nobody told it not to. You pay for that twice: once a month on the bill, and once again on the day the traffic doubles and it stops responding. The wrong foundation costs almost nothing while you are small. It is the most expensive thing you own the month you start to scale.
Where the money goes
These are not bugs. Nothing crashes, no alert fires, and no tool reports them, because the code is correct. They are choices, made by something that had no idea what your bill looks like. These six show up in almost every app we read.
01
One list of twenty items becomes twenty one separate queries instead of one. It is invisible at ten rows and it is your entire hosting bill at ten thousand. The usual fix is a few lines.
02
A managed queue for forty jobs a day. A vector database for two hundred records. A search service for a table you could filter in the database you already pay for. AI reaches for whatever is popular, not whatever fits.
03
No token limit, no timeout, the whole document sent when a paragraph would do, and the expensive model used for work a cheap one does just as well. This is usually the single largest line on the bill and the easiest to halve.
04
The same exchange rate, address lookup or embedding fetched from a paid API on every request, when the answer has not changed since the last one. Each call is correct and each call is billed. A cache with a sensible lifetime turns thousands of calls a day into a handful, and it is one of the cheapest fixes on the list.
05
An index is right or wrong only in relation to what is in the table and how it is queried, and the agent saw neither. So it adds the obvious ones, misses the ones the real queries need, and indexes columns nothing filters by, which slows every write for no gain. It looks fine at a thousand rows and shows up as a full table scan at a million.
06
No pagination, no index, everything loaded to filter it in code. This one is quiet until the month you get traction, and then it arrives as an outage rather than an invoice.
The larger number
Overpaying by a few hundred a month is annoying. It is not what this is really about. The same choices that inflate the bill are the ones that decide whether the app survives its first good month, and those cost differently.
01
A data model that cannot represent what you now sell is cheap to change in week two and a migration with downtime in month nine. Firms that do this work charge between €10,000 and €35,000 for the rebuild. The version where you catch it early costs a morning.
02
Your first serious customer sends a security questionnaire. If the honest answers are no, that deal stalls until they are yes, and the work takes as long as it takes.
03
Everything holds at the traffic you tested with. The interesting question is what happens at ten times that, and it is answerable now, cheaply, rather than live.
Pricing
Every price here is what you pay. Each one buys a fixed block of our time rather than an open scope, so you know the number before we start and there is no hourly bill to watch. For comparison, firms that do the rebuild instead charge between €10,000 and €35,000.
If we find nothing worth fixing, you pay nothing and the report is still yours. We invoice on delivery either way, so you read it before you pay. The whole fee comes off the Sprint if you book within 30 days.
We agree the list on day one and write it down. If we deliver less than the list, we invoice less, in proportion. The list can change along the way if you need it to; anything added is quoted before we do it and billed on top, never slipped in. 30 day warranty: if we broke something that worked before, we fix it free.
An architect on your side while the team and its agents keep building on the codebase: we see changes before they ship, keep the foundation from drifting, and are the people you call when something needs a decision. Available after a Cost and Risk Map. No minimum term; either side can end it with 30 days notice.
The obvious questions
Probably not, and we will tell you that on a twenty minute call before you pay anything. The honest rule: if you are spending under about a hundred a month and nobody outside your team uses the app yet, wait. Come back when the bill moves, when a real customer arrives, or when someone is about to look at the code. Those are the moments this pays for itself.
You can ask, and it will change something. The difficulty is that it does not know what your bill is, which parts of the app matter, or what you are about to do next, so it optimises whatever you happened to paste in. We use models heavily ourselves, to read the whole repository and map what calls what. That is a large part of why this costs what it does rather than several times more, and we would rather say so than let you assume we read every line by hand. Deciding which of the findings is worth your money is the part you are paying two people for.
You are not handing it over. Read only access, under a mutual NDA signed before we open anything. We ask for a read only repository invite or a zip. We do not ask for, and will refuse, production credentials, your database, or anything with a real customer in it. We work on our own machines and delete our copy within 14 days.
Then you pay nothing and you keep the report, which is a written second opinion that the foundation is sound, worth having before a funding round or a first enterprise customer. We would also have told you on the twenty minute call, before you paid anything, if we did not think you needed us. Above that, our fee is the same whether we find one thing or thirty, and it does not go up with the severity we assign, so nobody here earns more by exaggerating.
Between us we have architected systems and led engineering teams for large enterprises in jewellery retail, banking, automotive and other regulated industries, the kind of places where an architectural decision is expensive enough that somebody counts it. We are bound by those contracts and cannot name them, so do not take our word for any of it. Book twenty minutes instead and judge us on the answers. Bring the part of your app you are least sure about, and we will tell you what we would look at and why, before you have paid us anything. We will also tell you if we think you do not need us.
Talk to us