- Relay.app shut down for free users on August 15, 2026, and will shut down for paying customers on September 14, 2026, eliminating one of the few automation platforms built around human-in-the-loop AI approval steps.
- Zapier’s task-based pricing counts advanced AI steps as three to five tasks each, meaning complex multi-step AI agents can exhaust plan limits fast and push teams into significantly higher cost tiers.
- n8n’s per-execution billing, where a 20-node AI workflow counts the same as a 2-node one, makes it materially cheaper than task-metered platforms for teams running high-complexity agent workflows at volume.
Relay.app built something genuinely useful: a structured approval layer between AI outputs and business-critical actions. Free users lost access on August 15, 2026, and paying customers have until September 14, 2026, before the teams that relied on it face a forced migration to Zapier or n8n. Neither is a direct replacement, and the cost gap between them grows fast once workflows get complex.
What Relay.app Actually Did
The platform let you wire human decision-making directly into an automated workflow. An AI step would draft a response, a human would approve or edit it, and execution continued only after that. It also offered Custom Prompt AI and dedicated AI Agent steps, with integrations across multiple AI models. Users consistently rated the interface as more intuitive than Zapier, Make or n8n.
That design addressed something most platforms ignore: non-deterministic AI outputs in business-critical processes are risky, and treating human review as an afterthought is how errors compound. No current replacement handles this at the same level of workflow integration.
Zapier’s AI Pricing Problem
Zapier has pushed hard into AI tooling over the past 18 months. Its Copilot workflow builder launched in January 2024; Zapier Agents, rebranded from Zapier Central, followed in January 2025. The platform now covers text analysis, translation, summarisation and classification across more than 8,000 app integrations.
The pricing is where things get complicated. Every successful action counts as one task. Standard AI model calls cost one task each; more capable AI steps cost three or five. On the free tier, that’s 100 tasks per month, enough for a simple two-step workflow, not enough for anything resembling a real agent. Task consumption compounds fast, and teams building multi-step AI workflows will hit paid tier thresholds quickly. The visual builder and integration depth make Zapier the obvious entry point for non-technical teams. The cost unpredictability at scale is the genuine problem. For a closer look at where task-metered billing creates friction in production, see our breakdown of hidden costs in enterprise AI automation workflows.
n8n: Execution-Based Billing
n8n charges per execution: one workflow run is one unit, regardless of node count. A 20-node agent with five AI steps costs the same as a two-step workflow. For complex automations running at volume, that billing model changes the economics substantially, and it’s where n8n tends to pull ahead of Zapier in high-volume stress tests.
Teams self-hosting on a small VPS report infrastructure costs of roughly $5 to $10 per month. The tradeoff is operational maturity: n8n rewards builders who are comfortable managing their own infrastructure and debugging at the node level.
The Real Deployment Challenges
Switching platforms solves the pricing question. It does not solve the harder problems in production agent deployments.
Multi-step agent processes are prone to cascading failures, particularly with advanced models on complex tasks. Data quality and integration complexity remain the dominant friction points, and no platform makes those go away. Relay.app’s shutdown removes the one tool that took human oversight seriously by design, baking approval gates into the workflow rather than bolting them on after.
Zapier offers reach and accessibility, but costs mount as AI usage scales. n8n offers cost control and developer flexibility, but requires more operational maturity to run well. Neither eliminates the governance work that serious agent deployments require.
