Inside DigiSign's AI Content Studio: shipping generative AI to production

DigiSign's AI Content Studio takes a one-line brief and returns finished screen designs, composed on a 336-template library, ready for a human to approve and schedule to real screens. We built it because of a bottleneck we kept running into inside DigiSign, our cloud signage platform: screens go stale, and they go stale because every fresh design needs a designer that the shops and clinics running those screens do not have. The manager who owns the screen can write a sentence. Opening a design tool at nine in the evening and producing something the brand would sign off is another matter entirely. So last month's offer keeps playing until someone finds time, and nobody finds time.

Why staleness kills a screen is a subject of its own; our signage content guide covers what stale screens cost and what to put on them instead. This is the story of the fix we shipped, and of what taking generative AI to production genuinely involved: the constraint that made it work, the pipeline, the running costs, and what we would tell anyone attempting the same.

What does the Content Studio actually do?

The flow is short on purpose. A user types a brief the way they would message a colleague on WhatsApp: “Diwali offer, 15% off gift hampers, warm festive look, Hindi and English.” Gemini reads that sentence and writes the on-screen copy. A composition engine lays the copy onto templates from our library and renders finished designs at the exact resolution the display needs. The user picks one, edits a word if a word needs editing, presses approve, and schedules it. From typed sentence to a design playing on a physical screen: minutes, with no designer anywhere in the chain. The person typing the brief never opens a design tool and never sees a toolbar.

It has been running in production inside DigiSign ever since, which is the part that separates this story from most writing about generative AI. Demos are easy. Tuesdays are hard.

Why constrain the model to 336 templates instead of letting it draw?

If your instinct says AI-generated design looks generic and slightly broken, your instinct is right, and we will not argue with it. Our early prototypes let the model compose freely. Text spilled out of its box. The logo wandered. One draft set body copy in a colour the brand manual explicitly bans. Every attempt arrived quickly, which somehow made it worse: freeform generation produces five bad options fast, while constrained composition produces one usable design slowly enough.

So we took the pen away. The Content Studio composes on a library of 336 templates, each designed by a person who understood type hierarchy and viewing distance long before any model got involved. The model chooses and fills; it does not invent geometry. That one constraint buys brand safety, layout sanity and print-grade output every time, and of everything in the build it is the decision we would defend hardest.

What does the pipeline look like in plain words?

Four stages. The brief goes to Gemini, which works out what is being asked for and writes the actual words: headline, supporting line, offer terms, small print. A composition engine lays those words onto candidate templates and checks the fit (does the Hindi headline overflow, does the contrast survive shop lighting). Finished designs render at the display's native resolution. A person reviews the results, approves one, and schedules it to screens. That is the entire machine; there is no fifth step.

Notice what the model never does. It does not decide that a screen needs fresh content, and it carries no accountability when a price is wrong. Generative AI did not remove the human from this system; it removed the blank canvas. Gemini will write a confident Diwali headline, but it cannot know that the discount was actually approved at 10% rather than 15%, or that the hamper in the photograph was discontinued in March. Judgement stayed with people. What moved to the machine was production: the hours between deciding what a screen should say and having something fit to put on it.

Nothing reaches a screen without a person pressing approve.

What does it honestly cost to run?

Less than people expect, and not where they expect it. At Gemini's advertised API prices, the model spend behind one design's intent and copy is small change, the sort of line item that disappears into any cloud bill. Even across a busy deployment, run costs for a feature like this sit comfortably inside what typical AI features cost to run in a month: thousands of rupees, not lakhs. There is no per-design fee to budget for, and no scaling cliff as the screen count grows. The model was never where the money went.

The number that matters is a different one: the marginal cost of a fresh design, which collapses toward zero. When a new design costs a designer's day, updates get rationed, and screens receive their quarterly refresh whether the business changed or not. When a new design costs one sentence and a minute of review, the manager updates the screen because a thought occurred to them on the drive in. Screens stay fresh because fresh stopped being expensive, and fresh screens are the entire point of the medium.

What would we tell a client who wants to build something similar?

This request now reaches us regularly under different names: product posters for one business, menu boards for another. What follows is the advice we give when we scope generation systems in our AI and automation practice — advice we followed ourselves, apart from the parts we learned by not following them.

• Constrain the output to a format you can guarantee. Templates and fixed schemas turn generation from a party trick into something you can promise a customer; freeform output turns every result into a review job.
• A person stays on the approve button, permanently. Not as a safety measure you remove once trust builds: as the design itself. Approval is where accountability lives, and it takes seconds when the options in front of you are already good.
• Measure whether people use it, not whether the demo impresses. A feature that delights a boardroom and gets opened twice a month has failed, however the demo felt.
• Expect the boring parts to be most of the build. The Gemini integration was a small fraction of our engineering time; queues, retries, timeouts, malformed-response handling and template versioning were the bulk of it, and they are the reason the Studio gets through its Tuesdays.

And before any of that, the question we ask first: do you need generation at all? If your screens change once a month and someone on the team makes the designs happily, you do not need this yet. A template library and that person are cheaper, and at that volume no slower. Build the pipeline when the person becomes the queue, when designs are wanted faster than anyone can make them; that is exactly the moment we built ours.

That is the honest shape of shipping generative AI to production: a capable model, deliberately caged, doing the one thing that was genuinely scarce. The blank canvas is gone from DigiSign. The judgement never left.