Mediana Smart Campaign

Mediana's users are not marketers. They are shopkeepers, clinic managers, the PR person at a language school — people who know exactly what they want (“bring back my old customers”, “sell this product”) but not how to translate it into a campaign: which segment, what copy, which channel, what budget. The classic campaign builder assumed that translation was the user's job. Every question it asked — audience? copy? schedule? — was a question our users couldn't confidently answer. Campaigns didn't get made, or got made badly, and Mediana stayed what it had been for 14 years: a place to send raw SMS, not a place to do marketing. The business goal was equally clear — move Mediana from selling messages to being the tool where marketing decisions get made.

Thirty out of thirty. We rejected the trend and named what we built instead: AI-Assisted UX. The structure stays familiar, with visible steps, clear progress and everything inspectable, and AI enters only at the decision points where users actually get stuck: writing the copy, choosing the audience, framing the marketing approach. The user always sees a plan before anything runs, and always decides.
‍

The user states the goal, the system drafts the plan, the human approves the plan. What the research changed was not just the interface pattern but the order of priorities. Clarity of the path scored higher than ease of use in every session, so the design optimizes first for "I know where I am and what happens next", and only then for speed.

Three ways in: a product URL, nothing at all, or a plain-language request. The smart campaign is one of four campaign types on the platform — its siblings cover manual audiences, VIP-curated banks, and retargeting by behavioral tags — but it's the one built for the user who has a goal and nothing else. With a product URL, the system reads the page and infers the campaign's intent from the artifact itself. With nothing online at all, the user enters raw product details and the AI generates a landing page — for many offline businesses, their first usable digital asset. Or the user simply describes what they want.
‍

The system doesn't execute; it proposes. From that intent it drafts complete campaign packages, each bundling a suggested audience, message copy, and the marketing approach behind them. The approach is stated explicitly, because a plan you can't interrogate is a plan you can't trust. Every part of a package is editable, and the parts stay honest with each other: edit the message, and the audience is re-proposed to match it. The campaign's intent lives in the message the user actually sends, not in a form they filled three steps ago.
‍

Nothing runs until the user says so. Before anything is sent, the whole plan is laid out on one screen — the message, the proposed audiences with the reasoning behind them, the channels, the schedule, the cost — under one heading: “Needs your approval.” It is the least glamorous screen in the product and the most important one: it is where “AI proposes, the human decides” stops being a principle and becomes an interface.
‍

“AI proposes, human decides” has to be engineered, not promised. “Won't it just send it to five million people?” On a platform where every send costs real money and cannot be unsent, four mechanics carry the weight, and each one answers a specific fear we heard in the room:
‍

1. Subset control. “It will send to everyone.” After audiences are proposed, the user chooses to send to all of them or only part. The system's output is a starting point, never a commitment.
‍

2. Exclusion. “It will bother people I already messaged.” Users exclude groups such as recipients of previous campaigns, turning mass sending into relationship management.
‍

3. Gradual delivery. “Once it goes, it's gone.” Alongside “now” and “scheduled”, a campaign can run gradually, so the send becomes a monitorable process instead of a single irreversible action.
‍

4. A report that talks back. “I won't understand what happened.” After (and during) a campaign, users get delivery, engagement, cost and demographic breakdowns plus an AI-written summary they can question conversationally. On gradual campaigns that closes the loop: analyze mid-flight, adjust, continue.
‍

This is the one place we kept a conversation, on purpose. Many of Mediana's customers have no website, no product page, nothing to point an AI at. For them the smart campaign generates the missing artifact: the user describes the product in plain terms, a short conversational assistant shapes the details, and the system produces a landing page, then runs the same propose-and-approve flow on top of it. The first version of this step was a form; it never shipped. Instead we chose the lowest-risk step in the flow, one with no money on the line and a visible outcome, and left the chat there as a live experiment. If people are comfortable talking to the system when nothing can go wrong, a fully conversational option becomes something we can offer later, backed by evidence rather than by trend. For an offline shop the result is often the first digital asset they have ever owned.

Live on both panels; results claims on hold until the rollout completes. The smart campaign runs on the direct Mediana Next panel and the reseller-facing IPPanel, which puts it in front of the platform's full base of ~500K business users. The signal we watch most closely is the edit rate: how often people correct the system's proposals is the truest measure of whether those proposals deserve the trust the interface asks for.
‍

1. The prompt box is not the product. Our users' problem was never typing; it was decision confidence. Structure gives confidence; AI fills the gaps inside it. Thirty out of thirty told us so before we wrote a line of production code.
2. Explain the plan, not just the output. Bundling the marketing approach with each proposal did more for trust than any accuracy improvement.
3. Listen for the fear, not just the friction. The most useful sentence in the study wasn't about ease of use; it was "won't it just send it?" Usability metrics would never have surfaced that. Think-aloud did.
4. Keep intent and execution coupled. Re-proposing audiences when the message changes is one small interaction, and it's the difference between a form with AI sprinkled on it and a system that understands what you're trying to do.
‍

Each release, the system earns the right to one more decision. The same philosophy is expanding deliberately, one decision at a time: channel selection is next in line to move from the user to the system. That sequencing, more than any single feature, is what I'd call the design of the product.