Your AI wrapper is really a workflow

The laziest insult of the last few years has been wrapper: a thin interface over a rented model, one API key away from being rebuilt by any competitor. It contains a grain of truth, roughly one part in ten. Most products wearing the AI label are not thin, and the part supplying the thickness is not the model. Delete the inference call and most of these companies still own something substantial: a schema, a state machine, human checkpoints, an audit trail, a queue, a retry policy. Models get replaced; the scaffolding persists. That is not a wrapper. It is workflow software with a probabilistic component in the middle.

The demo is the model, the product is everything around it

Take document processing, the canonical case. An invoice arrives by email. The demo moment is the model reading it and emitting structured JSON in four seconds. The product moment is everything before and after that: pulling attachments out of a PDF that is really a scan, validating extracted numbers against the customer's chart of accounts, routing a low-confidence line to a human instead of quietly inventing one, posting the approved entry to the accounting system, recording who confirmed what, and staying usable when an inbound webhook is delivered twice or the model vendor announces deprecation with ninety days of notice.

Not one of those steps is inference. All of them are ordinary software engineering with exacting requirements. The same shape recurs across the category: support triage classifies and routes and escalates, code review tools propose patches for a human to accept, enrichment tools fill in fields against a source of record. In each case there is a deterministic skeleton with a single stretchy step in the middle that calls a model. The skeleton is where the work went, and it is where the value stays.

The state machine is the durable artifact

The skeleton deserves respect, because it is what survives if every model you rely on vanished tomorrow. It has states: received, extracted, needs review, approved, posted, failed. It has guarded transitions: nothing reaches approved without either a confident model output or an explicit human click. It is resumable, so a crash between extraction and approval loses nothing, and idempotent, so a replayed webhook produces the same result as the first delivery.

Portability is the part worth dwelling on, because models get replaced whether you like it or not. Vendors deprecate on months of notice, prices move, quality shifts release to release, and a competitor shows up with a better model at a better price. When the probabilistic step lives behind one node in the graph, the swap is contained: point that node at the candidate model, replay last month's real documents through a shadow path, diff the outputs against what the current model produced, flip the configuration when the error rate is acceptable. The eval set built from human-corrected answers becomes migration fuel. Without the state machine and the eval set, the same deprecation notice is a rewrite.

The human checkpoint is a feature, not an apology

The cheap version of an AI product hides its uncertainty and ships confident wrongness straight to the customer, which is how the support queue learns about your accuracy rate before you do. The good version schedules a person.

Confidence thresholds route low-certainty extractions into a review queue. Amount gates send anything over a certain size to human eyes even when the model claims to be sure. The correction interface matters as much as the automation: when a person fixes a value, that fix writes back into a golden set of examples, and next quarter's candidate models get evaluated against it before they ever touch production. Over time the checkpoints shrink as the eval set grows, but the design assumption stays put: the machine is allowed to be wrong, the system is not allowed to be silently wrong.

Provenance lives in the same category. Which model version read this field, under which prompt revision, confirmed by whom. That audit trail is boring, and it is what accountants, regulated customers, and support engineers actually need. No base model ships with one.

What survives a model swap

Rent the commodity, own the context. The model is priced like a commodity because it is becoming one, and the durable inventory accumulates on the other side of the line: a domain schema that encodes your customers' chart of accounts rather than a generic invoice shape, validation rules discovered one incident at a time, an eval set made of real corrections, integrations with the systems of record people actually pay for, permissioning, retention rules, cost and latency budgets per step. A competitor licensing the same model tomorrow starts with none of the scar tissue that taught you where validation belongs.

That is why wrapper was always the wrong word for most of this category. The honest name is workflow software whose middle step happens to be a probability distribution. The workflow is the product, it persists, and the middle keeps improving and keeps getting swapped, which is exactly what you want from the commodity part.

From the desk that wrote this

This blog is run by the team behind Layerbase, so every post comes with a conflict of interest attached rather than concealed behind it.