AI agents in Zapier: launch 35+ models without API keys
Zapier brings together autonomous agents and 35+ AI models in a single editor. We look at how businesses in Uzbekistan can automate processes without separate API contracts.
On October 1, 2026, Zapier announced that it was merging its standalone product, Zapier Agents, with its core visual script editor under a single “AI by Zapier” experience. At the same time, the service provided built-in access to more than 35 language models and implemented support for the open Model Context Protocol (MCP). For executives and COOs of companies in Uzbekistan and Central Asia, this means the ability to implement multi-stage autonomous agents in CRM, warehouse databases and work chats directly through a standard subscription - without purchasing separate API keys, currency accounts for developers and complex administration of token balances.
What has changed in the Zapier platform
Previously, digital automation in Zapier was built on strictly deterministic chains: the emergence of a trigger (for example, a new application on the website) caused a strict sequence of predetermined steps. A separate experimental module, Zapier Agents, existed separately, which complicated the process architecture. Now this separation has been completely eliminated.
Autonomous decision-making steps (agentic reasoning) and launching external tools (tool-calling) are now built directly into the standard Zap step-by-step editor. This allows you to combine classic linear actions with stages where the model evaluates the context, decides on the next action or queries missing information from connected databases.
Key technical changes of the platform:
- Unified model library:The system integrates about 35 models from providers OpenAI, Anthropic, Google, Moonshot AI and Z.ai without the need to connect your own API keys to each to the provider.
- Write-off through platform tasks:the costs of running neural networks are included in the overall monthly Zapier task limit through a system of coefficients (multipliers).
- One-click migration mechanism: Existing classic multi-step Zaps can be converted to agent-based logic with one click, preserving existing triggers and integrations.
- Support for Model Context Protocol (MCP):External AI assistants can now securely query data and trigger actions through the Zapier catalog, covering more than 9,000 enterprise applications and over 40,000 applied actions.
Where applicable: business scenarios in Uzbekistan and Central Asia
The main barrier for many regional companies in creating complex AI processes previously lay in the infrastructure: direct use of advanced neural networks required foreign registration corporate developer cards, set up payments for tokens and attract programmers to process responses. The introduction of models directly into the Zapier ecosystem removes this barrier, since the service is available in Uzbekistan on the basis of standard corporate accounts.
In the practice of automating business processes, there are three applied areas where such agents provide an immediate effect:
- Qualification and routing of incoming requests: When a request arrives from corporate mail or instant messenger, the agent can analyze the essence of the message, check the client’s history in CRM, and compare the request with the current one loading of managers and independently choose whether to make an appointment, create an invoice or transfer the task to a specialist.
- Reconciliation and processing of primary data:The agent is able to take unstructured text or an incoming report, extract items, check them with the warehouse table and create an adjustment entry in the accounting system without writing regular expressions and strict formulas.
- Preparation of analytical summaries for management:The script can regularly poll multiple connected tools, collect project statuses, identify delays and publish a structured summary to the team's work channel.
Step-by-step plan: how to implement AI steps in workflows
To move from rigid scenarios to flexible agent chains, companies are recommended to follow a consistent plan:
- Audit current scenarios:Highlight processes where regular failures occur due to non-standard format of incoming data or where employees have to manually transfer information between systems.
- Select the appropriate model class:Assess the complexity of the task. For simple sorting and classification, standard models are sufficient; for in-depth analysis of context, legal text or multi-level terms, choose advanced or premium options.
- Set up access to tools:clearly indicate which applications from the Zapier catalog the agent can access (for example, only reading cards in CRM and creating draft messages).
- Set control points (Human-in-the-loop):for all critical steps - deleting data, invoicing, sending public letters - include mandatory approval by the responsible employee before execution.
- Conduct testing on the execution log:run the process on a sample of real historical incidents and analyze detailed step logs (execution history) to adjust the agent’s instructions before transferring the process into permanent operation.
Tariffing, restrictions and operational risks
The transition to agent-based scenarios requires an accurate understanding of the economics of the service. Zapier abandoned token accounting in favor of the usual scale of tasks (tasks), introducing call multipliers:
- Standard models: consume 1 tariff plan task per call. Suitable for basic formatting, routing and simple responses.
- Advanced models:Charges 3 tasks per hit. Optimal for multi-stage document analysis and working with complex logic.
- Premium models:they write off 5 tasks per request. Used for resource-intensive tasks that require maximum precision in reasoning.
Despite their convenience, autonomous steps have operational limitations. An agent with open access to tools can potentially perform an erroneous action when given an ambiguous instruction. Therefore, the security architecture must rely on detailed audit trails and manual reconciliation checkpoints. It is also important to monitor quota consumption: cyclical requests from an agent with an incorrectly assigned task can quickly exhaust the available package of tasks for the month.
The place of autonomous agents in complex automation
The appearance of AI steps in no-code platforms does not eliminate the need for strict business logic. The experience of automation projects shows that autonomous agents are most effective not as a replacement for fundamental databases or accounting systems, but as a flexible “bridge” between them. They take care of routine tasks that require human judgment on a small scale - parsing emails, categorizing, filling in missing fields, and routing tasks.
Frequently Asked Questions
Do I need to buy separate OpenAI or Anthropic API keys?
No. Access to all 35+ models included directly in the Zapier platform. The user does not need to register developer accounts with model suppliers or top up the balance in foreign currency.
How to control that the agent does not make an error in the working database?
The platform allows you to implement a mandatory human confirmation stage (human-in-the-loop). The agent prepares the decision and parameters of the operation, but the action itself on the target system is executed only after the button is pressed by the responsible employee.
What does Model Context Protocol (MCP) support provide?
The MCP protocol allows external AI tools and local assistants to securely use the Zapier ecosystem as a single directory of connectors. This opens up the ability to access more than 9,000 services and 40,000 actions through a single, standardized interface.
The NDC (Digital Business Solutions) team helps companies design a robust digital architecture and implement process automation - contact us to understand your company's challenges.
