What AI does not do in a print workflow

What does "what AI does not do" mean here?

Most print industry writing about AI lists what a model can attempt. This article lists where it stops, in five specific jobs a print workflow is often sold an AI feature for: artwork quality checks, background removal, upscaling, auto-layout, and support agents. None of these claims are hedged for marketing effect; they describe where a model's output is a suggestion, where a deterministic rule has to be the one that decides, and where a named person has to approve the result before it reaches production or a customer.

This matters because a wrong guess in print is expensive and hard to reverse. A bad background cut or an invented price does not get a second edit once a job has run. The qualifying question this article is written to answer is not "can AI do this job," but "which part of this job can AI not be trusted with, and who or what decides instead."

Where does AI fall short on artwork quality checks?

A model can be useful for spotting a likely problem in an uploaded file: a pattern that resembles a low-resolution image stretched too far, or a colour mode that looks wrong for the product. What it should not be trusted to do is decide, on its own, whether a measurable property of the file meets a printer's specification. Resolution at final size, bleed geometry, embedded fonts and total ink coverage are all properties the file itself carries, and they are exactly checkable by a deterministic rule rather than by a model's judgement call.

Files that pass some checks and leave others unchecked still wait in a queue for a person to review.

The reliable split is: a model flags a pattern it was trained to recognise, and a deterministic preflight rule enforces the exact threshold a printer's specification sets. Automated preflight: designing pass and fail rules for print files covers how those thresholds are written and tested; an AI flag can point a reviewer at a file faster, but it does not replace the rule that decides pass, warn or fail.

Where does AI fall short on background removal?

Background removal tools fail in predictable places: fine hair, glass, reflective surfaces, fur, and fine text or line art sitting against a busy background. A model trained on general photography can produce a convincing cut for a simple product shot and a visibly wrong one for exactly the images a print business is most likely to need removed, because packaging mockups, apparel photography and jewellery are disproportionately the hard cases. A wrong cut that reaches a press is not a quick fix; it is a reprint.

The limit here is not that the tool is bad; it is that the tool cannot see the edge case it has just mishandled. A person reviewing the result before it reaches a template or a proof can catch what the model cannot flag about its own output.

Where does AI fall short on upscaling?

Upscaling a low-resolution image does not recover detail that was never captured; it invents plausible-looking detail to fill the gap. At normal viewing distance on a small print, an upscaled image can look acceptable. At the resolution a large-format or close-viewed product demands, the invented detail shows as a texture that does not match the rest of the image, and it shows worst exactly where the customer looks closest: a face, a logo, fine text.

A deterministic resolution check, comparing the image's native pixel dimensions against what the product's size and viewing distance require, still has to run regardless of whether an upscaling step was applied. Upscaling can make a borderline image usable for a tolerant product; it is not a substitute for the resolution threshold the preflight rule enforces, and it does not turn a genuinely inadequate source file into a correct one.

Where does AI fall short on auto-layout?

An auto-layout feature can propose a plausible arrangement of text and images inside a template, and that is a real time saving for a first draft. What it does not know on its own is a printer's production geometry: the exact trim box and bleed a product needs, which elements must sit clear of a fold or a die line, and which layout changes would make a finishing operation impossible on a given substrate. Those are deterministic constraints, not stylistic preferences, and a model optimising for a pleasing arrangement has no reason to respect them unless a rule checks the output afterwards.

The safe pattern is the same one that governs artwork checks: a model proposes, a deterministic geometry rule validates the trim, bleed and fold constraints the layout must respect, and a person approves the result before it is accepted as a template.

Where does AI fall short as a support or sales agent?

A support or sales agent built on a language model is useful for extracting a structured specification from a free-text enquiry and for answering a status question by reading a system of record. It should never be trusted to set or adjust a price, confirm a delivery date, approve a refund, or make any other commitment on the business's behalf, because none of those are facts the model can know; they are decisions that belong to a pricing engine, a production schedule and a named person. AI agents for print sales and support sets out a worked example of exactly this boundary, including the full table of what such an agent never decides.

Responsibility table: who decides?