The most cited statistic in GTM engineering right now comes from Brendan J. Short’s The Signal, which examined the stacks of 62 fast-growing B2B companies at $10M+ ARR: 65% use Clay. Zero — none — use an off-the-shelf AI SDR. One dataset, undisclosed selection methodology, so hold it loosely. But it rhymes with everything else we can verify: UserGems-reported annual churn of 50–70% on AI SDR tools (vendor-sourced, echoed across every 2026 buyer’s guide), Clay crossing a company-announced $100M ARR, and GTM engineering headcount doubling year-over-year per Sumble’s job-post data. The market has voted, and it voted for neither “build” nor “buy” as the debate was originally framed. It voted for a third thing.

What actually failed

The product that churned was the digital worker: an autonomous rep-replacement that owns the whole outbound job. The failure pattern is consistent with what we documented in Why Most AI SDRs Fail — the job was never one job. Prospecting, qualification, personalization, sending, and follow-up have different error tolerances, and a system marketed on replacing all five gets fired the first quarter it embarrasses someone in front of a real account. Even the vendors concede the framing shift: the practitioner consensus has moved from “replace the SDR” to automating specific jobs inside the workflow — and Actively AI raised its Series B on an explicit “AI SDRs failed” thesis.

What actually won

Orchestration plus judgment. The Clay-centric pattern that dominates The Signal’s sample is not “build everything”: it’s a composable middle — enrichment waterfalls, signal routing, personalization assembly — operated by a human who decides what good looks like, with sending delegated to boring, reliable sequencers. That’s why the same Sumble data shows GTM engineering as the breakout role (400+ at US digital natives, doubling yearly) while overall GTM job posts fell 15%: companies are hiring the person who composes the system rather than licensing a system that claims to need no one. Our Clay waterfall playbook and the GTM engineer’s toolkit map that architecture.

Where buying still wins

Three places. Point capabilities with hard floors: deliverability infrastructure, visitor identification (RB2B), calling agents (Bland) — building these yourself is malpractice. Tier-2 coverage at scale: a packaged agent (11x, Artisan) on accounts no human would ever touch is a different risk calculus than replacing your named-account motion — our Series B stack prices exactly that split. The reasoning layer, maybe: Actively’s per-account intelligence bet is “buy,” but it’s buying prioritization for human reps, not their replacement — which is precisely why it’s the one agent category enterprises aren’t churning out of yet.

The decision rule

Buy anything whose failure is invisible to prospects. Build (compose) anything that decides who you contact and what you say. Never buy anything marketed as needing no operator — the churn data, vendor-sourced as it is, all points the same direction. And staff the composer: on current hiring evidence, one GTM engineer with Clay, a sequencer, and taste outperforms every digital worker on the market. The companies in the data already know it.