The stack that got you here stopped fitting.

No — do it so the role is worth taking. A first ops hire who arrives to an undocumented stack spends two quarters reverse-engineering it before they can improve anything. One who arrives to a reporting layer, a set of runbooks and the manual jobs already automated spends that time running the business instead. We would rather build the thing they inherit than compete with them for it.

Sometimes, and we will tell you when it isn’t. Judgement work, supplier relationships and anything that needs a person deciding are not automation problems. The copy-paste between two dashboards is. The audit separates the two and puts a number on the second, so the hiring decision gets made against evidence rather than against how busy everybody feels.

Both, and the split is the interesting part. Most of what we build at this stage is deterministic: a rule that reconciles two exports, routes an exception, retries a declined card on a schedule the decline code chose, or posts the same number to Slack every morning. A rule is cheaper per run, testable, and incapable of inventing an answer. A model earns its place where the input is genuinely unstructured — a customer’s actual sentence, a supplier’s PDF, a product description that has to be written rather than looked up. Anything that decides money is drafted by the system and pressed by a person. If we cannot tell you which of those three a given job is, we have not finished the triage.

No, because the models are not the bottleneck at your size. What kills an AI project here is the layer underneath it: eight to twelve tools that do not agree on what a customer is, a shared inbox holding operating facts no system can read, and no reconciled profit number to judge any of it against. That is exactly why the reporting layer goes first on this page. An agent pointed at data nobody trusts produces confident answers nobody should act on — and it produces them faster than a person could check them.

Probably, and we would rather say so than take the money. The published floor is $3M+ annual revenue, on Shopify Plus or running a paid subscription platform. Below it, a five-figure build is a large share of a year’s profit at the margins this category runs on, and the honest advice is to spend it on the single leak costing you most rather than on a system. Ask us which one on a call and we will tell you without a proposal attached.

Rarely. Most of what breaks at this size is the absence of a layer underneath the tools rather than the tools themselves. We build the warehouse and the automation in your own accounts and leave the front ends alone. Where something genuinely has to go — usually the second attribution dashboard nobody trusts — we say so and show the arithmetic.

The audit is two weeks at $1,500–$3,000, fixed, and credited in full against any build. Builds run in four-week blocks at a fixed price per block. The reporting layer normally goes first, because every decision after it is easier to argue when the numbers agree.

Your team, which is the point. Everything is built in your accounts, on tooling you pay the vendor for directly, with a runbook per workflow. Self-hosted n8n is the automation layer precisely because it does not hold the business hostage to a per-task price, or to us.

No. Most of this is infrastructure rather than campaign marketing — billing, reporting, the manual ops jobs — and it is usually the piece nobody at a campaign agency owns. Where the flow work genuinely overlaps, we hand your agency the audit and the sequence map and let them build against it.

It is the most common answer we get, and it is the first thing the audit fixes. A brand running eight to twelve marketing tools has eight to twelve versions of revenue, each attributed by the tool that sold it to you, and none of them carrying cost of goods or the real shipping charge. Until that is reconciled, every other decision on this page is being made on instinct.