AI Architecture

Four internal tools. Reporting, the creative pipeline, attribution, and a competitive read. All of them are in production and used by people other than me.

They exist because of arithmetic. Around $167K of live budget a month, nine brands across six categories, two platforms that both change their reporting surface without asking. Done by hand that is a few hours every week of moving numbers between two interfaces and a sheet, and those hours land on the same day the decisions have to be made.

None of it was built to be impressive. Each one replaced a specific task that was being done manually, badly, against a deadline.

A web app over the ad data, dashboards on top of that, and a chatbot on top of the dashboards. Weekly reporting went from a few hours to a few minutes.

The chatbot is the part that earns its place. A dashboard only answers the question someone thought to build a chart for, and most of the questions worth asking in a live account are shaped like why this account is costing more than it did on Thursday. Nobody builds that chart in advance.

The model was never the constraint. The data model underneath it was. An answer that arrives confidently over a badly joined table is worse than no answer, because budget gets moved against it the same afternoon.

Creative used to move through six manual steps. Rename to convention, push to storage, write the row into the repository, generate the unlisted link, load into Meta, record what it was. In a normal week at least one of those steps had a naming error in it, and a naming error is not a small thing. It is an asset that cannot be read afterwards.

Now it is one step, and every asset gets a unique promo ID at the moment it is created. That is the whole idea and it is not clever. It means creative is attributable at the asset level rather than the campaign level.

What that buys is memory. A cut that worked can be found again six weeks later, and spend moves toward it on evidence instead of recollection.

A scraper that pulls the Meta and Google ad libraries and competitor social into one place.

It does not tell me what to do. It tells me what the auction is about to feel like. When several advertisers in the same category go loud inside a fortnight, the price of the same audience moves, and it moves before platform reporting shows it moving.

That is worth a few days of reaction time and not much more. Treated as strategy it turns into copying.

I am not an engineer. These were specified by someone who knows the data model well enough to be precise about it, and built with tools that closed the rest of the gap. Anything that needs real uptime guarantees or a security review is not mine to own, and I say that early rather than after it breaks.

Automation is expensive while a process is still moving. Building a tool freezes a decision, and if the decision is still being argued over, the tool encodes the wrong version and then defends it, because now there is a script that says so. The ones that lasted were built after the manual version had run long enough to stop changing.

None of this decides anything. A tool can tell you what the spend did. It cannot tell you whether the spend was worth making, whether the ceiling has moved, or whether the number in front of you is real. That judgement stays with whoever holds the budget.