A campaign agent in 30 minutes, built from a public offer

The situation, and why it is hard by hand

A Baltic bank publicly announced a new card with cashback at a large grocery chain. Good offer, ordinary problem: the people worth telling are scattered across an existing customer base, and telling them means either advertising that reaches everyone indifferently, or a calling campaign somebody has to staff.

Offer promotion is one of the campaign shapes we build for lenders every week, the same machinery as a re-activation or a win-back campaign, pointed at a product instead of a balance. So my question was not whether the platform could run it. It was how long it takes to get from reading about an offer to having the campaign that promotes it.

What was built

I authored a deterministic flow the way any campaign is authored. An opening that says who is calling and why. A step that establishes whether the person is interested at all. A small set of questions the agent may answer: what the cashback is, where it applies, how you would get the card. And a clean exit for anyone who says no.

The design work was in what the agent may not do, which is where your compliance function will start reading. It gives no financial advice. It confirms no eligibility and implies no approval. Anything outside its answer set becomes a captured question with a call-back, or a transfer to a person. Those boundaries are steps in the flow rather than instructions in a prompt, which is why they hold on the thousandth call as well as the first.

What comes back

A call log tells you a number was dialled. What lands in your CRM here is a disposition per contact, the objection as a code rather than a note someone on your team has to read, the recording, the transcript, the flow version that ran and every tool call the agent made.

That is what makes your second campaign better than your first. "Not interested" is a dead end. "Not interested, already holds a competing card" is a targeting rule for the next run.

What it took, and what it did not prove

Under thirty minutes, on my own, with no engineering ticket. I authored it through Convershake's agent tooling and reused the campaign scaffolding that already exists: the opening, the consent and opt-out handling, the disposition set.

Two honest caveats. This was a build, not a deployment: I loaded no list and called nobody's customers. And thirty minutes buys a working flow rather than a launched campaign. The calendar in a real launch is set by connecting telephony, agreeing the contact rules for the market, and the customer's own review of the script.

What people said about it

The offer still has to be worth the call. Voice AI makes a good offer reach further; it does not make a bad one better.

I asked whether people would rather hear about a card offer from their bank by phone than through advertising, and the answer split. Some said yes, gladly: a straight answer in one minute beats a week of banner ads. Others said they would rather find products themselves and did not want the call at all.

Both reactions are constraints you can design for. Your consent and suppression rules decide who is called. Relevance decides whether the call is welcome. A one-sentence opt-out that actually works decides whether a second call ever happens. Ignore any of the three and your campaign is just louder than the advertising it replaced.

Where this applies, and what I would watch

If you lend, this is the revenue side of the same Convershake platform that runs your reminders and early collections: re-activation, win-back, declined re-application, lead conversion. Same authored flows, same guardrails, same evidence per call, pointed at growth instead of recovery.

What I would watch before calling it a win: whether the list is genuinely segmented, whether the opt-out is honoured the first time it is said, and whether the second campaign beats the first because of the reasons the first one captured. A campaign that cannot show me those three is just activity with a phone bill.