Own the layer that connects your whole stack.

A cold-pressed juice brand runs two co-packers and one 3PL. Every morning somebody exports yesterday’s orders out of Shopify, opens the 3PL portal in a second tab and reconciles the two by eye — because the inventory feed lags a day and the subscription app has no idea which facility a given box ships from. Forty minutes, every morning, for two years. It has never been written down, and on the mornings that person is away it simply does not happen.

That is not one broken process. It is dozens, and they all live in the same place: the space between two systems that each work perfectly well on their own. A supplement brand answers the same month-three question — should I keep taking this past week eight — by assembling a reply out of a FAQ page, a product description and three old tickets. A roaster changes green lots every eight weeks, and every change means tasting notes, grind options, marketplace attributes and reorder points all have to move together, by hand, across four systems. A snack brand drafts every purchase order off a stock figure somebody eyeballs on a Tuesday.

Each of those systems has an app store, and every app in it owns one box. Almost nothing is sold for the space between the boxes, so the space between the boxes gets staffed with people — and the hiring plan quietly turns into an integration plan nobody has written.

It stays that way because fixing it is not a configuration job. It is software: API clients, pagination and rate limits, idempotency so a retried job does not send a second purchase order, schema drift when a 3PL renames a column, credential rotation, and somebody noticing when the Shopify Admin API deprecates the version you built against. A marketing-led agency is not staffed for that work and does not want it. This is the one page on this site where the differentiator is simply that we write code for a living instead of configuring other people’s.

It is worth saying the unflattering half too. A good deal of what gets sold as an “AI agent” in this category is a rule with a language model bolted on for decoration. Some of what we build is exactly that — a rule — because a rule is cheaper per run, testable, and incapable of inventing an answer. The model earns its place only where the input is genuinely unstructured: a customer’s actual sentence, a supplier’s PDF, a review, a product description that has to be written rather than looked up.

And there is a second reason these projects never get built: the tools most agencies reach for bill in a way that punishes exactly this shape of work — many steps, run many times — so the automations that would pay for themselves get quoted out of existence before anyone writes a line. That arithmetic is worth doing in public.

Anything where a wrong answer costs more than a human minute

The test is not “can a model do this” — it usually can. The test is what a mistake costs. If being wrong means a compliance problem, an incorrect ingredient or feeding statement, a claim you cannot make, or a subscriber who leaves, then the expected value of automating it is negative even at a two-percent error rate. Agents belong on work where the worst outcome is that somebody redoes it.

Anything touching money, without a person on the end

Refunds, discounts, payment-method changes, credits, cancellations that trigger a refund, subscription price changes. An agent may assemble all of them — gathering the order, the policy and the history, and drafting the action — but a human presses the button. That is a deliberate boundary rather than a technical limit, and it is the same one we hold on payment recovery: billing is the one place where an automated mistake compounds instead of ending.

Anything sitting on data that is not already reliable

An agent reading a catalogue where a third of the ingredient fields are stale will produce confident, fluent, wrong answers considerably faster than a person could. The same is true of a 3PL feed nobody reconciles and a subscription record that disagrees with Shopify. Fix the source of truth first — usually catalogue and feed work or a reconciliation job — then automate on top of it. We will say so in the audit rather than sell the agent anyway.

There is a fourth, quieter one: anything a rule already handles. Deterministic workflows are cheaper per run, easier to test and impossible to talk out of their logic. Most of what a brand actually needs from ops automation is rules, and the agent layer is a smaller, sharper thing sitting on top of them.