ARTICLE SUMMARY
AI Agent governance is the set of controls that defines which model each agent uses, how every execution is tested and monitored, and when a decision requires human approval. It is what allows companies to expand their use of AI with predictable costs, traceability and security.
As AI Agents take on more steps in business processes, the focus shifts from adoption to control: which model runs each task, how much each execution consumes and who is accountable for each decision.
According to Gartner, over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. These factors depend less on the models themselves and more on AI Agent governance.
Pipefy’s most recent updates, released between July and September 2026, were built around that need. In this article, you will find the most impactful ones, organized into three areas: AI governance, process visibility and execution with AI Agents.
[E-Book] 30+ companies that orchestrated their processes with AI in Pipefy
What are the latest Pipefy updates for Q3 2026?
The table below summarizes the Pipefy updates from this period and what each one changes in the daily work of teams:
| Area | Update | What changes in operations |
|---|---|---|
| Governance | Pipefy AI Lite, Standard and Pro models | Model set per behavior, based on the complexity of each task |
| Governance | BYOM with Anthropic models | Claude models connected with the company’s own API key |
| Governance | Company Overview | AI, automation and API consumption across multiple organizations in a single view |
| Governance | AI Agent webhook alerts | Status and reasoning of each execution in the team’s monitoring tools |
| Orchestration | Map and Card Volume | Pipe connections to other pipes and databases, with card traffic for each one |
| Orchestration | Flow View | Phases, forms, automations and AI Agents in a single editable diagram |
| Orchestration | Card Timeline | History showing whether each action came from a person, an automation or an AI Agent |
| Execution | Behavior and automation testing | Scenario validation before publishing, with no impact on production data |
| Execution | Max Effort (Beta) | Deeper reasoning for complex analysis and documents longer than 30 pages |
| Execution | Human Validation | Approval checkpoint within the agent’s behavior before the action moves forward |
| Execution | MCP Client connectors | 12 connectors with one-click authentication and no API keys |
AI Agent governance: models, consumption and alerts under IT control
The first area brings together the features that give IT control over cost, privacy and stability, the foundations of AI governance. AI Agent governance starts with choosing the right model for each task.
Pipefy AI Lite, Standard and Pro models, set per behavior
Pipefy AI Agents now offer three native models, selected per behavior rather than per agent. This way, a single agent can combine different models based on the complexity of each step:
- Lite: simple, structured tasks where response speed is the priority, such as translating a requester’s comment during triage.
- Standard: the default model, suited to tasks such as reading an invoice PDF, extracting the supplier, amount and dates, and routing the card.
- Pro: complex analysis, such as comparing three supplier quotes against RFP requirements and drafting the rationale for the final choice.
BYOM with Claude models from Anthropic
Bring Your Own Model (BYOM) lets companies connect the models they already license to Pipefy, using their own API keys.
Since July, Anthropic has joined the list of providers, alongside options such as OpenAI, Google Vertex and AWS Bedrock, and Claude models can be assigned to each behavior.
Model management also includes a Zero Data Retention option, under which prompts are not stored.
Company Overview: consolidated consumption across multiple organizations
For companies running multiple organizations on Pipefy, Company Overview works as a governance hub, bringing automation health, API consumption and AI Credits for each organization into a single view.
With per-pipe charts and report exports, decisions about where to expand AI use are grounded in data.
AI Agent webhook alerts
Each AI Agent execution can trigger a webhook alert with its status, the card data and, in case of failure, the agent’s reasoning and the point where the execution stopped.
Alerts can be routed to tools such as Slack, Microsoft Teams and Datadog, or to any external endpoint, so teams can spot exceptions before they affect the process.
Agentic orchestration: visibility from the ecosystem down to the card
In agentic orchestration, a single workflow runs across pipes, databases, integrations and AI Agents. The second set of Pipefy updates expands visibility into this structure at three levels: ecosystem, flow and card.
Map and Card Volume
The Map tab places the pipe at the center and displays the pipes and databases connected to it, along with the type of each connection, with no setup required.
Card Volume adds traffic from the last 30 days: line thickness and color indicate the volume of each connection, and dashed lines flag connections with no activity.
Flow View
Flow View brings the pipe’s phases, forms, automations and AI Agents into a single diagram, along with SLAs, labels and admins. The diagram is also editable: clicking any element opens its configuration.
Card Timeline
At the card level, the Timeline records the full history from creation to the final phase and uses distinct icons to show whether each action came from a person, an automation or an AI Agent.
Failed executions and executions with no action now appear as well, with direct access to the logs, and card deletions and restorations are recorded for auditing.
Execution with AI Agents: from testing to human validation
The third area brings together the features that add security to AI workflow automation, from testing to human approval.
Behavior and automation testing before publishing
Behavior testing simulates what an AI Agent would do based on a real card from the pipe, without changing production data, so IT and process owners can refine instructions before publishing.
For automations that move cards, the test checks the pipe’s actual configuration and points out the reason for any failure, such as an empty required field.
Max Effort for long documents and complex analysis
Currently in Beta, Max Effort expands the agent’s reasoning: it plans, searches for specific context, looks up particular pages and reviews its own analysis before acting.
Enabled per behavior and powered by the Pipefy AI Pro model, it is designed for contracts, invoices and reports longer than 30 pages.
Human Validation within agent behaviors
Human Validation adds an approval checkpoint to the agent’s behavior: the execution pauses, a task goes to the person in charge to approve or reject the proposed action, and the next step moves forward only after the review.
In an expense reimbursement, for example, amounts below the limit set by the company proceed automatically, while the rest go to the manager for approval.
MCP Client connectors with one-click authentication
MCP Client lets AI Agents query and act on external systems within their behaviors, such as retrieving lead data from Salesforce or HubSpot.
In August, the feature gained 12 connectors for the most requested operational tools, such as Google Calendar and OneDrive. Authentication takes place on the external platform’s own login screen, with no API keys.
- Learn more: Pipefy Updates: more control as you scale AI
How the Pipefy updates work together in operations
The greatest gains from the Pipefy updates come when the features work together within the same process.
Imagine an expense reimbursement process. Triage could use the Lite model, while a behavior with Max Effort would review the longest reports against internal policy. Amounts above the limit would go through Human Validation.
After being validated through behavior testing, the workflow would go live: failures would reach IT through webhook alerts, the Timeline would show who performed each step and Company Overview would show AI consumption.
This combination turns AI Agent governance into routine within agentic orchestration. After all, companies don’t scale with prompts. They scale with orchestrated processes.
For operations at enterprise scale, these features are complemented by certifications such as SOC 2, ISO 27001, ISO 42001, GDPR and CCPA. Governance for IT, speed for teams and ROI for the business.
See these updates applied to your own processes:
