Most marketing AI stories are demos. This one is a build log. Over three two-week sprints, Aicadium built and tested SPECTRUM (Strategic Personalized Engagement Campaigns Through Relevant Unified Marketing). It is an AI-assisted engine that takes a campaign from a structured brief to a complete, persona-differentiated copy package and creative brief. This is what we built, what broke, and what we would tell any marketing leader attempting the same.
Our manual baseline for one campaign is about two weeks. It covers campaign asset planning, persona definition, research, copywriting, a first round of review and edits, and the creation of the creative brief. SPECTRUM does the same span of work. It generates personas, generates the asset copy, runs that copy through the copy evaluator, edits it, and writes the creative brief. It does all of that in about 25 minutes of compute time for roughly US$0.10 as measured when creating four persona-specific campaigns that included blog posts, landing page copy, email drip series, social posts, and a video script.
The bet we set out to test
The hypothesis was simple. AI-assisted marketing can multiply a team’s impact roughly tenfold without growing headcount. The bottlenecks included the marketing team’s bandwidth, the wait for other teams’ inputs, and the approvals a campaign needs before it can move. SPECTRUM changes both ends of that wait. It uses AI to fill in where we do not already have good information, while still accepting human-generated material wherever that exists and is preferable. Its built-in copy evaluator then shortens the approval cycle by catching issues before a human reviewer ever sees the draft. Systematise those handoffs, and the marketing team gets a streamlined, AI-driven multiplier.
We scoped the first phase to the creation half of the lifecycle. Brief, personas, copy, quality gate, and creative brief. Distribution and optimisation came later. The target was to compress a two-week manual process into a single short run, with quality high enough to publish after light editing.
What the working system does
SPECTRUM runs end-to-end as a real application. A marketer answers structured intake questions through a step-by-step wizard. It asks for the product, its value proposition, the target audience and geography, the campaign goal, the channels, and any brand guidelines or source material. The system drafts targeting across industries, roles, and pain points, and the marketer can edit, regenerate, or delete any of it. It then generates three to five genuinely differentiated personas and lets the marketer select among them. Copy follows, per persona and per channel:
-
Blog posts
-
Email campaigns
-
LinkedIn posts and ads
-
Landing-page copy
-
A video script
That is the same multi-day drafting work across blogs, emails, social posts, ads, landing pages and scripts that used to fill much of the two-week baseline. Now it arrives as a review-ready first draft in minutes.
Two design choices did the most work. Copy is completed and approved before any creative is produced, because finished copy is the best possible art direction. Within the same run, SPECTRUM writes a detailed creative brief that can direct an AI assistant, a specialised generation tool, or a human designer. And every AI step has a human checkpoint before the next stage begins: approve, edit, or reject. See it in action in our demo video.
What did it cost?
In our runs, a full campaign covering two industries and two roles cost roughly US$0.10 and about 25 minutes of compute, across 65 model calls. It produced the copy package plus a detailed creative brief. As we optimised, our working benchmark tightened to under US$0.025 per campaign run. Producing the finished visuals is a separate downstream step: generating a set of ten brand-consistent campaign stills from the brief cost about US$1.
Those numbers are our own measurements, from our own campaigns, and we publish them as observations rather than industry benchmarks. What matters is the order of magnitude. At these costs, budget is no longer the constraint on personalised campaign production.
Why we built an application instead of relying on AI chats
The most scrutinised decision of the build was architectural. Why not do all of this in a single AI-assistant chat? Because campaign creation is interactive, multi-step work. Uploading documents, editing suggestions, selecting personas, and triggering downstream actions are interface problems rather than prompt problems. Structured inputs and outputs let us validate what the model returns and correct it when the output drifts from the instructions.
Building an application also kept us model-agnostic. SPECTRUM calls whichever AI assistant the user prefers through the provider’s API, rather than being locked to one assistant’s proprietary skill format. That flexibility is itself an argument for an app over a single-assistant skill. The engine, the interface and the quality gates stay put, while the underlying model is a choice you can change.
We still use narrow AI skills where they fit. Our copy evaluation gate is exactly such a skill, plugged into the application. The principle we landed on has two halves. Use a skill where the task is narrow and predictable, and build software where the problem spans many inputs, model calls, and human decisions.
The takeaway for marketing leaders
The technology worked, but the process design mattered more. Three lessons matter most for a marketing leader considering an engine like this:
-
Systematise your inputs and approvals first. The teams that benefit are the ones that can specify their intake questions, their quality bar, and their approval points precisely enough to hand off.
-
Treat it as a repeatable capability, not a one-off. A structured application turns campaign production into a process teams can adopt, scale, and govern. That is the business value that outlasts any single campaign.
-
Decide ownership and access early. Who owns the tool, where it runs, and who can use it are the questions that stall adoption.
Two lessons from the engineering are worth keeping for the teams who will build or buy. First, application infrastructure (persistence, authentication, deployment) was the largest source of unplanned work and is invisible in a demo. Second, keeping prototypes disposable meant two stack pivots cost days rather than weeks.
Want to learn more about our AI marketing transformation? Read the audit that shaped our stack decisions.
