Fixing the content supply chain: how I would architect AI into a retailer’s marketing engine

AI is changing how content gets made. It has not changed what makes a content operation work.

Almost every marketing leader I speak to describes the same squeeze. There are more channels than ever, customers expect content that feels made for them, and the team doing the work has not grown. Adobe’s June 2025 survey of more than 1,600 marketers shows how big the gap has become.

Source: Adobe research, 1,600+ marketers, published 16 June 2025.

The last two numbers are the ones that matter. Making an image or writing a headline is rarely what slows a team down. Everything around it does: requests that arrive half-formed, files nobody can find, approvals that queue one behind another, and no clear signal about what actually worked. Put generative AI into that process and you get more content, waiting in the same queues.

So the architect’s job is not to “add AI”. It is to redesign how work flows from request to result, and then put AI where it genuinely removes time and risk.

Take a UK omnichannel retailer in home and lifestyle, the kind of business this problem hits hardest. It sells online and in stores, changes its promotions every week, and its marketing team is stretched across more channels every year. This is the picture I would expect to find, and the solution that follows is designed for it.

Before designing anything I would spend two weeks tracing ten recent assets from request to live, timing every handoff. In a business like this one, the pattern is usually the same.

Requests arrive by email, chat and in meetings. Nobody can see total demand, so priorities are set by whoever shouts loudest and capacity planning is guesswork.

A free-text brief misses the facts the creative team needs: product codes, price, legal lines, channels and sizes. Each gap triggers another round of questions.

Brand, product, legal and regional teams review one after another, often on email attachments. Comments conflict, versions fork, and a two-day job takes two weeks.

Finals live in shared drives with no consistent naming, no usage rights and no expiry dates. Teams recreate assets that already exist, or use ones they should not.

Each channel team resizes the master asset by hand. Once it is live, performance data stays in the ad platforms and never reaches the next brief.

The two-week trace produces a value-stream map: for each stage, how long someone actually works on the asset (touch time) and how long it sits waiting (elapsed time). The ratio, flow efficiency, is the single most useful number in the programme, because it shows how much of the delay technology alone cannot fix.

I design the solution in five layers that follow the life of an asset, with orchestration, data and governance running across all of them. I have mapped it to Adobe’s stack because that is where most large retailers already have licences, but the layers matter more than the logos. Each layer has one job and one owner.

Every request enters through one form in Workfront with required fields: objective, product codes, offer, legal lines, channels and deadline. A brief assistant drafts the brief from those fields and flags what is missing before it reaches a designer. Marketing leadership gets one calendar showing demand against capacity.

People still make the hero work: the campaign idea, the photography, the master layouts. Variants are where volume lives. Templates in Express and a Firefly model trained on the retailer’s own approved imagery produce on-brand variations, while Firefly Services APIs handle the repetitive work: cutting out products, swapping backgrounds, recolouring and resizing to every channel format in one pass.

AEM Assets becomes the single library for anything approved. AI tags each asset on upload, and the request’s metadata (product codes, campaign, rights, expiry) travels with it. Approvals move from a line to a fan: brand, product and legal review at the same time in Workfront, and each asset is routed by risk. A resized banner with unchanged copy does not need legal; a new price claim always does.

Approved content flows straight out to channels without download and re-upload. GenStudio for Performance Marketing builds and publishes paid social and display variants, AEM Sites and Dynamic Media serve the web, app and in-store screens, and Journey Optimizer uses the same assets in email. Customer segments from the data layer decide which version each audience sees.

Content Analytics links results back to the individual asset and its attributes, such as the colour, product and message. That closes the loop: next month’s brief starts with evidence of what worked rather than opinion.

The layer view says what each part does. The integration view says how they talk to each other. The design principle is simple: Workfront owns the state of work, AEM Assets owns the state of content, and nothing moves between them by hand. Every handoff is either a webhook, an API call or an event.

The single highest-value integration is the one that turns an approved master layout into every channel version. This is the sequence I would build first.

Automation fails quietly when metadata is optional. I would define a small contract that every asset must carry before it can be approved, and enforce it in the AEM metadata schema rather than in a guideline document.

Volume grows by multiplication, not addition. Take one weekend promotion:

No team resizes 1,296 files by hand, so in practice it publishes a fraction and tests nothing. With the pipeline above, the marginal cost of a rendition falls to an API call plus a sampled human check, and the constraint moves from production to deciding which variants are worth testing.

The automated check in step 7 only works if brand guidelines exist in a form a machine can evaluate. Most brand books are PDFs. Part of the work is translating the rules that matter into a versioned configuration owned by the brand team, for example:

The values above are examples. The point is the shape: thresholds a service can test, references to the systems that hold the truth (the PIM for price, legal for terms), and routing rules that decide who must look. When the brand team changes a rule, it changes once and applies to every rendition from that moment.

Every rendition keeps a link to its master and to the request that created it. When a price changes or a licence expires, the system finds every derivative in seconds and either regenerates or withdraws it. Without lineage, an expired photograph can stay live in hundreds of placements nobody remembers making.

This section is for the architects and engineers who will build it. It covers the data model, the lifecycle every asset follows, how the pipeline stays reliable at peak, how environments and releases are managed, and how the system is observed and secured. Where a value is an assumption rather than a product limit, it is marked as such.

Most content platforms fail on data before they fail on technology. Nine entities carry the whole flow. The key design choice is that a rendition is a first-class record with its own identity, product, channel, audience and provenance, rather than just a file under a folder.

“Approved” means different things to different teams until it is defined as a state with entry rules. I model every rendition set as a state machine. Only the transitions shown are legal, each is triggered by a named event, and each guard is enforced by the platform rather than by goodwill.

Black Friday and the Christmas launch are where this design proves itself. The variant pipeline follows a fan-out, fan-in pattern with explicit back-pressure, because generative and imaging APIs are asynchronous, rate-limited and occasionally fail.

Sizing the pool. Throughput is set by three numbers you measure in the pilot rather than guess: the concurrency the API entitlement allows (K), the average job time including polling (t seconds), and the share of jobs that need a retry (r). Sustained throughput is roughly K × 3600 ÷ t × (1 − r) renditions per hour. With an assumed K of 20, t of 40 seconds and r of 5%, that is about 1,700 an hour, so the 1,296-rendition promotion clears in under an hour. If the measured numbers are lower, the plan changes the schedule (generate overnight ahead of a launch), not the architecture.

Every step consumes and emits small, explicit messages. A simplified example of the event the brand-check action receives, and the shape of its handler:

A content pipeline that can publish to paid media is production software and is released like it.

Targets are set with the business in the foundation phase and reviewed monthly. When a target is missed, the error budget decides what happens next: new automation pauses and the team fixes reliability first.

A design that works for one market and one format breaks in predictable places when it meets twenty markets, six languages, video and ten years of legacy files. These are the extensions a global rollout needs.

Brand rules are global; advertising law, language and culture are local. Rather than fork the rules per country, I layer them. Each layer can only add or tighten constraints, never loosen the one above, and the most specific layer wins where they conflict.

Language versions are just another dimension of the variant matrix. Copy is translated once per locale through translation memory and machine translation, then human post-edited only where the risk class requires it (promotional terms, regulated claims, anything tone-critical). Locale fallback is explicit, so a missing Irish-language line falls back to the approved English line rather than to a blank layer, and the fallback is flagged in review.

Not every variant needs to exist as a file. Simple crops and resizes for owned channels can be rendered on request at delivery time; anything containing generated pixels, or bound for an ad platform that needs a file, should be stored as the exact approved file.

The cheapest asset is the one that already exists. The intake form searches the library as the request is typed and shows near matches (same product, similar scene) before a new request can be submitted. On ingest, near-duplicate detection flags images that are visually the same as something already approved, so the library does not fill with copies under different names. Reuse rate becomes a measure the content owner is held to.

Video follows the same states and rules but runs in its own queue with its own service levels, because jobs are larger and slower. Masters are reviewed frame-accurately in Frame.io; derivatives cover aspect-ratio cutdowns per platform, burnt-in or sidecar captions per locale, loudness normalisation and end-card swaps for offers. Captions are mandatory in the rule pack, which also covers accessibility for sound-off viewing.

Alt text is drafted automatically on ingest and reviewed for hero assets; contrast and minimum text size are part of the automated check; email and web templates are tested against WCAG 2.2 AA before they enter the template library. Accessibility built into templates scales; accessibility fixed asset by asset does not.

Legacy libraries are usually several times larger than what the business actually uses. I migrate by value, not by volume.

Content automation needs the same test discipline as any production system, adapted to images and copy.

The most useful decision an architect makes is not which model to use. It is which tasks AI should do on its own, which it should assist with, and which stay with people. I sort every task in the workflow on two questions: how repetitive is it, and how costly is a mistake?

Two rules keep this honest. First, every AI-generated or AI-edited asset carries Content Credentials, so anyone can see how it was made. Second, the retailer only uses models that are licensed for commercial use and, for the custom model, trained on images it owns. Both are easier to set as defaults on day one than to retrofit after a complaint.

A custom model is a product with a lifecycle, not a one-off setup. I would run it the same way the business runs any system that can publish to customers.

Producing 1,296 renditions does not mean testing 1,296 ideas. Most variants exist to fit formats and audiences, not to learn something. I separate the two: format and audience versions simply ship, while a small number of deliberate hypotheses (for example lifestyle image versus product-only, or price-led versus benefit-led copy) are tested with enough traffic per cell to reach a decision. Results are recorded against the creative attributes in the metadata contract, so the learning compounds across campaigns instead of staying inside one ad account.

Architecture diagrams can hide how work actually moves. Here is a single weekend promotion travelling through the new design.

In the old flow, the designer made one asset and four channel teams each made their own versions after approval, so legal effectively reviewed the same promotion several times. In the new flow, the full set of variants is generated before review, checked automatically against brand rules, and approved once. The person signing off sees every version side by side.

Technology changes who does what. Programmes stall when the new work has no owner, so the organisation design is part of the architecture, not an afterthought.

Two commercial changes usually follow. Agency scopes move from paying per resize to paying for masters and ideas, because variants become an internal, automated output. And channel teams stop producing their own versions; they consume approved ones from Content Hub and request new ones through the front door.

A CFO will not fund a diagram. The model below shows how I would build the production-cost case, using the baseline from the trace. The starting values are working assumptions; replace them with your own figures and the result updates. It deliberately leaves out revenue upside from better-performing content, which should be proven in pilots rather than assumed.

In most cases the hours are worth more redeployed than removed. The same team that spent its week resizing can produce more original work, run more tests and support more markets, which was the goal in the first place.

A retailer in this position usually starts at level 1 or 2. The 12-month plan targets level 3 across the business and level 4 for paid social, where the feedback loop is fastest.

I would not switch everything on at once. Each phase has to pay for the next, and each one ends with a decision point for the leadership team.

Stand up the single request form and brief assistant, migrate the assets that are actually in use (not the whole archive) into AEM Assets with rights and expiry, and run automated resizing on one live campaign. The aim is to prove a shorter cycle on real work with real people.

Move every campaign onto the new intake, switch approvals to parallel and risk-based, and extend automated variants to all channels. This is where most of the time savings appear.

Train the custom model on the cleaned library, add AI copy variants with human sign-off, and connect asset-level analytics to briefing and audience targeting.

I agree the measures with the CMO and CFO before the first workflow goes live, and record a baseline from the two-week trace. Without a baseline, nobody can prove the investment paid off.

Production metrics prove the machine works. They do not prove it makes money. For that I separate three questions and use a different method for each: did content get cheaper and faster (operational data, before and after); does more relevant content perform better (controlled tests with holdout groups at campaign level); and does it change revenue at business level (geographic or time-based holdouts, read alongside the marketing mix model the finance team already trusts). Claims are reported only at the level of evidence behind them.

A design is only as good as the decisions behind it. These are the ones I would record and take to the steering group, with the option I would choose and why.

Adobe has introduced Agent Orchestrator in Adobe Experience Platform, with purpose-built agents for tasks such as resizing in content production. In this design, agents belong in the optimise phase, once the metadata contract, approval rules and audit trail exist. An agent working inside clear rules is useful; an agent working in a messy library is a faster way to make the same mistakes.

The technology in this paper is ready. Most of it is already licensed in large marketing organisations, sitting underused. What holds companies back is the work around it: agreeing who owns the flow, writing the rules down in a form a machine can check, and having the patience to clean the library before training a model on it.

None of that is glamorous, and all of it compounds. A team that fixes its foundations this year will be able to hand real work to AI agents next year with confidence. A team that skips them will spend next year undoing what the agents did.

If I had to leave a chief marketing officer with one question, it would be this: if demand for content grew five times tomorrow, which part of your process would break first? Start there.