Where AI Fits in a POP Display Manufacturer's Workflow

Almost everything written about AI and design assumes the reader is a designer sitting alone with a brief. A custom point-of-purchase shop does not work like that. There is a salesperson who took the meeting, a designer who gets handed whatever came out of it, an estimator who has to put a number on it, an engineer who turns an approved picture into something that can be cut, and a plant that has to build the thing at a real unit count.

I have worked in point-of-purchase for close to four years and I build AI POP Displays. This is the map I would hand a shop owner, step by step, because deploying AI at the wrong step loses more time than it saves.

Here is the nine-step flow of a normal custom job, and what actually happens to each step.

Step 1: the sales meeting

Untouched, and it should stay that way.

Somebody sits with a brand or an agency and finds out what the campaign is, what the retailer will allow, what the budget really is, and which of those three the customer has not thought about yet. None of that is a drawing problem. The output of this meeting is not an image, it is a set of constraints, half of which arrive as offhand remarks.

The temptation here is to generate something live in the room. I would be careful with it. A picture in a first meeting sets an expectation about a structure nobody has costed, and walking that back later is more expensive than the meeting was worth.

Step 2: the brief reaches design

Mostly untouched. This is a reading comprehension problem.

What lands on the designer's desk is a forwarded email with a paragraph of requirements, a logo, and maybe a photo of a competitor's unit. Somebody has to work out what was meant. Is that counter or floor. Is that promo corrugated at 400 units or a permanent metal unit at 40. Is the retailer's footprint rule a hard limit or an opening position.

The part a tool can take is the sorting. In AI POP Displays the briefing step reads that email and lays it out against a display taxonomy, so format, material, environment and product load land in named fields and the gaps show up as gaps. The judgment calls stay with the designer, and the thinner the brief the worse everything downstream gets. I wrote the eight sections a brief actually needs in how to write a POP display brief, and the reason it matters here is mechanical. Ambiguity at step two does not stay at step two. It reappears at step five as a change request.

Step 3: the concept

Compressed hard. This is the step AI took.

Before, putting three directions in front of a customer meant three speculative builds, or three careful hand visuals, and a decision up front about which one to bet the hours on. That decision is where a lot of display work goes quietly wrong. You pick a reading of the brief, the customer sees it, and their reply tells you they meant something else.

With a model in the loop you answer the ambiguity by showing it. The same brief, three readings, in the time the first one used to take. What makes that work is the layer that sits on top of the image engine, and it is the work that decides a display. A general model reads a format name as a style and a material as an adjective, composes the product load instead of merchandising it, lets unlabeled references fight each other, and never tells you when it has quietly ignored a constraint. AI POP Displays is built for that job. A display taxonomy where each format name carries its proportions, a materials list that behaves the way the material does, retail environments as options rather than prompt text, product load taken from the packs the customer actually sells, references labeled by role with the sketch always last, and the briefing step from the previous section feeding all of it. That work is done once, in the tool, instead of retyped on every brief.

AI POP Displays runs on Google's Nano Banana Pro, shipped as gemini-3-pro-image, with OpenAI's gpt-image-2 on part of the pipeline.

If the constraint in your shop is that a designer can only afford to show one direction per brief, fifteen free credits is enough to find out whether that constraint is real.

Step 4: the quote

AI should stay out of this one entirely.

An estimator needs a bill of materials, a board grade, a real footprint in millimeters, a print spec and a unit count. A render supplies none of those. It has no measurements in it at all, which means proportions can be completely believable and still not be the retailer's gondola end.

The failure mode is specific and I have watched it happen. The concept gets sent to the estimator as if it were a spec, somebody prices what they think they are looking at, and the number is wrong in whichever direction hurts. The render is an input to the conversation about what to quote. It is not the thing being quoted, and the discipline of keeping those two separate is worth more than any speed gained upstream.

Step 5: the client change rounds

Compressed, and this is where the hours actually were.

This is the step almost nobody writes about, and it is the one that decides whether a custom job is profitable. A job typically runs three to six change requests before approval. Taller header. Different material. Move the branding to the side wing. Show it in a pharmacy instead of a supermarket. Can we see it from the other side.

Historically each of those meant going back into the model, re-rendering, waiting, and sending. The round trip was measured in days and most of it was not thinking, it was rebuilding. That is what changes when the concept lives as something you edit rather than something you rebuild. In our own console the operations that exist today are changing the ratio, changing the retail background, generating other angles, converting to a line drawing, converting to a photographic read, removing product, applying a graphic, and a free-text edit for everything else.

The thing to watch, and it is a real limitation rather than a footnote, is drift. Ask for one change and other details in the frame can move with it. Holding the geometry of an object across six versions is hard, and no vendor has it fully solved. The way to judge it is on a real change round with your own packs, and the free plan covers that on one real brief.

Step 6: approval

Partly compressed, mostly logistics.

Getting a decision out of a customer is a chasing problem, not a rendering problem. What AI changes here is small but real. When the approval version can be shown in the aisle it is going to sit in, the customer is reacting to what their shopper sees rather than to an object on a white sweep, and that shortens the conversation.

What has not changed is that somebody senior still has to say the structure is buildable before the file moves on. Skipping that check because the picture looks resolved is the most expensive mistake available at this stage.

Step 7: production drawings

Untouched, and not close to moving.

This is where the approved image becomes a dieline, a board spec and a cutting layout. That work lives in structural CAD run by people who build displays for a living. Esko's ArtiosCAD, which names POP, POS and FSDU work on its own product page, shipped releases in April and July of 2026 with no AI capability in either. The AI Esko has shipped in that cycle is in other products, object selection in ArtPro+ and scripting and project creation in WebCenter, not in the structural tool.

The wider render market has moved, which is worth noting precisely because it moved somewhere else. KeyShot Studio, now on 2026.2, added a text-prompt feature called AI Shots in its 2026.1 release, running locally. That is generative imaging arriving inside a CAD rendering tool. It is still not a dieline.

So the pattern holds. Nobody in structural CAD is generating concepts, and nobody generating concepts is producing cutting layouts. The two halves of this job are converging from opposite ends and have not met, which is the gap we are building into next, and I come back to it at the end. I went through the full version of the argument in can AI design a POP display.

Step 8: the sample

Untouched. It is a physical object.

A white sample gets cut, usually on a digital cutting table from a vendor like Kongsberg Precision Cutting Systems, and then somebody puts real product on it and leans on it. That is the moment the job finds out whether the shelf carries a full facing and whether the base ships flat. No image has an opinion about any of this, and none ever will, because the question is about a material under load.

Step 9: production

Untouched.

Tooling, print, converting, packing, delivery windows, retailer compliance. Nothing in the generative layer reaches this.

So what does the map actually say?

Two of nine steps compress, and a third gets its sorting done for it. Two of nine are places AI should be kept out of, which are the quote and the structural sign-off. The rest are unchanged.

That is a smaller claim than most coverage in this category makes, and I think it is the one that survives contact with a real shop. What makes it worth doing anyway is which two steps they are. The concept and the change rounds are exactly where the unbilled hours live, because concepts get made for pitches that lose and change rounds are almost never invoiced separately. The steps that did not move are the ones you were already being paid for.

The related read, if what you are weighing is headcount rather than process, is AI versus hiring a designer, which splits the same work by who does it instead of by when it happens.

The gap in step seven is what we are building into now. Bellto is an AI agent for brands and agencies that takes a POP campaign from brief to shop drawings. The brand tells it what it is launching, the agent asks for what is missing, proposes the campaign mix with a budget split per store, and designs every piece at real scale in parametric 3D, with proportions derived from the product dimensions and facings. Every approved piece comes out as a 3D model, a part-by-part cutlist, dimensioned drawings, STEP of the assembly and a cut DXF per part, and the agent suggests manufacturers that fit by material, format and volume. For a shop, that means jobs arriving with dimensioned drawings, cut DXF and STEP instead of a photo and a paragraph, and the quote still comes from you. The waitlist is open to any brand or agency, and POP manufacturers can sign up too. We are contacting the first ones soon to run the first real campaigns end to end. It opens by invitation, in small groups, the campaign described at sign-up sets the place in line, pricing goes first to the people on the list, and there is no card.

If there is a brief on your desk this week, signup is free with 15 one-time credits and no card required, and Pro is $49 a month for 150 concept generations at the founding price, with the $69 list price stated openly. Nothing you upload or render trains AI models, on any plan, because most of this work sits under NDA and that should not be a paid feature.