Overview
Bring Your Own Model (BYOM) is now available for all companies on Pipefy. It lets you register multiple LLM providers and models in your organization and assign the right model to each Agent Behavior directly in the Agent Builder, as long as the Organization admin configures it previously. Each step in your process runs on the model that fits it best, whether that's a lightweight model for extraction or a powerful one for complex reasoning, all within the same pipe.
What's new?
- Multi-provider registration in the Model Registry: Company Admins and Super Admins can register more than one LLM provider at a time under Admin Panel → AI Settings → Bring Your Own Model, set credentials per provider, and designate one model as the organization default.
- Per-Behavior model selection: Pipe Admins assign a specific model to each Agent Behavior directly in the Agent Builder, using the model dropdown at the bottom of the Behavior panel. Behaviors without an explicit assignment automatically fall back to the org default.
- Three-layer fallback: the system first tries the model assigned to the Behavior; if none is assigned, it uses the org default; if no custom configuration exists or a failure occurs, it falls back to Pipefy AI.
- Available on all plans: BYOM is accessible to all Pipefy companies regardless of plan.
- Company Admin as Org Admin: Company Admins are now also added as Org Admins, giving them full access to BYOM configuration without depending on other admin roles.
- Supported providers: OpenAI, Azure OpenAI, Google Vertex, Oracle OCI Generative AI, AWS Bedrock, and custom providers via a self-hosted inference endpoint (OpenAI-compatible format required).
How does this help?
- Right model for every step: lightweight and fast models handle extraction and classification; more capable models take on complex reasoning and analysis. All in the same pipe, no split workflows.
- Full autonomy for Pipe Admins: model selection happens directly in the Agent Builder, with no IT ticket needed for each new Behavior or adjustment.
- Control over sensitive data: route each type of information to the provider approved by your security policy, with visibility into which model processes each Behavior.
- Continuity guaranteed by automatic fallback: no Behavior stops running due to a missing assignment. The system always resolves to a model.
- Eliminates multi-org workarounds: teams that split workflows across separate Pipefy organizations to simulate per-model behavior can now manage everything in a single org.
Use cases
1. IT ticket triage and root cause analysis with different models in the same pipe
A Pipe Admin configures two Agent Behaviors in an ITSM pipe. The first uses a lightweight model to classify incident type (hardware, access, software) and assign SLA by category automatically. The second, triggered only for tickets requiring deep root cause analysis, uses a more capable model to write the diagnosis and recommend a resolution path. Simple incidents get immediate triage; complex ones get analytical depth, without inflating cost across the entire queue.
2. Data extraction and compliance validation in procurement
In a purchasing process, one Agent Behavior extracts fields from the requisition (supplier, amount, spend category, cost center) using a fast, lightweight model. A second Behavior, responsible for evaluating policy compliance before routing to approval, uses a model with stronger contextual reasoning. Everything runs in the same pipe, with no duplicated workflows and no manual review step needed for the compliance check.
3. Employee onboarding with role-specialized Behaviors
An HR onboarding pipe runs three Behaviors: the first extracts data from the admission form using a lightweight model; the second generates a personalized checklist by role and department using an intermediate model; the third answers open-ended questions about internal policies using an advanced model. Each step runs on the most efficient model for its function, without separate pipes and without unnecessary cost on simpler tasks.
4. Customer support routing by request complexity
In a Customer Success pipe, the first Behavior classifies the request by type and urgency using a lightweight model. Simple requests move to automated resolution; requests that require a contextualized response are routed to a Behavior powered by a more capable model. Routing and fallback happen automatically within the same pipe, with no manual triage step.
5. Data governance by approved provider
A company with distinct regulatory requirements by data type configures Behaviors with different providers: financial data processed exclusively by a Legal-approved provider; general operational data using the org default model. Configuration is managed centrally by the Company Admin, with no need for multiple Pipefy organizations to isolate providers by data sensitivity.
Most relevant for
- Pipe Admins building workflows with Agent Behaviors
- Company Admins and Super Admins responsible for AI governance in the organization
- IT teams that need to control which LLM providers process internal data
- Procurement, HR, and Customer Success operations with pipes that combine simple and complex steps
- Companies with data sovereignty requirements or security policies by information type
Documentation
For the full step-by-step, check the Help Center article.
