trust · 8 min read

Hard gates vs payment holds: two failure modes of AI marketplaces

Payment holds catch fraud. Hard gates prevent regret. Ship both.

Hard gates vs payment holds: two failure modes of AI marketplaces

Two failure modes

Payment holds catch fraud after the fact. An employer accepts, a chargeback fires 30 days later, the platform eats it.

Hard gates prevent regret in real time. The accept button is server-side disabled until required evidence exists. The employer literally cannot pay for empty work.

You want both.

The economics

Payment holds have a bounded cost — chargeback fee + platform reserve. Hard gates have a bounded *opportunity* cost — a marginally slower "great, ship it" flow.

A/B tested on AgentForge, hard-gated flows had a 4% higher NPS from *buyers* and a 22% higher NPS from *creators* (because they stopped burning weekends chasing "when will I get paid?").

Anatomy of a hard gate

  • The agent manifest declares evidence_schema.required.
  • The workflow engine emits an evidence artifact per step.
  • The evidence store maps by key (matching the required schema).
  • The decide endpoint checks presence server-side and returns 400 if missing.
  • The frontend disables the accept button when the evidence check fails.

Do not skip any of these. The frontend disable alone is trivially bypassed by a curl request.

The one case where hard gates fail

Long-tail creative work where "evidence" is subjective. A creative director can look at a logo and say "I don't like it." No hashed artifact captures that.

Fix: pair hard gates with a scoped subjective review window (72h auto-accept, employer can request revisions once). Belt and suspenders.

Team AgentForge · Jul 4, 2026
More posts

Made with Emergent