The single most important design decision in AdNexus is also the most boring-sounding one: nothing the AI proposes goes live until a human approves it. We call it draft-first, and it shapes everything else.
It is tempting to let an AI agent act autonomously. The demos look magical. But ad budgets are real money, and a confident model that misreads a trend can burn through a daily budget before anyone notices. Draft-first turns that risk into a review step.
What “draft-first” means in practice
- Every write is a draft. Budget shifts, pauses, audience tweaks, creative swaps — all of them land in an approval queue, not on the live campaign.
- Every draft explains itself. The metric that triggered it, the change proposed, and the expected impact are attached, so you can judge quickly.
- You can edit before you approve. Change a number, narrow an audience, then ship — or reject outright.
- Everything is logged. Who approved what, when, and why is recorded for a full audit trail.
Why we made approval the default, not an option
A setting that defaults to “auto-apply” eventually gets switched on under time pressure, and then a bad change ships unseen. By making approval a structural part of the architecture rather than a toggle, the unsafe path simply does not exist. That constraint is a feature.
The trust dividend
Counter-intuitively, asking for approval makes the AI more useful, not less. Because you trust that nothing ships behind your back, you let the agent monitor far more aggressively — every account, around the clock — and you act on more of its suggestions. Trust is what unlocks scale.
An assistant you have to babysit is slower than doing it yourself. An assistant you can trust is the only kind worth having.