Skip to main content

​
What Are Data Pipelines?

Data Pipelines are pre-computed data infrastructure that give agents and Human-in-the-Loop (HITL) experiences the context they need to operate effectively. Rather than fetching data on-demand during an agent run, Pipelines extract and transform data ahead of time — so it’s ready when your agents and dashboards need it. Screenshot2026 03 30at9 48 22PM A Pipeline connects to an external system via a Connector, pulls and transforms the data according to logic you define, and stores the result as structured Pipeline Outcomes in the Internal Data Store. These outcomes can then be referenced by Agents during execution or displayed in HITL Experiences.

​
Key Characteristics

Extract & Transform. Pipelines pull data from external systems (via Connectors) and transform it into clean, structured formats your agents can reason over. Stateful Infrastructure. Unlike Tools, Pipelines run on a schedule and their outcomes persist between runs. The data is always available — agents don’t need to wait for a live API call. Platform-Level Logic, Project-Level Credentials. The extraction and transformation logic is defined once at the platform level. Each project brings its own connector credentials, so the same Pipeline can serve multiple customer environments. Named Outputs (Pipeline Outcomes). Each Pipeline produces one or more named output tables — for example, deals_over_10k or aggregated_expenses. These are the outputs that Agents and Experiences reference.

​
Pipelines vs. Tools

Pipelines and Tools may seem similar, but they serve different purposes in the Adopt AI architecture.
AspectToolPipeline
NatureCapability (what an agent can do)Data (what an agent knows)
ExecutionCalled at runtime, on-demandRuns independently on a schedule; outcomes pre-computed
StateStatelessStateful — outcomes persist between runs
Primary ConsumerActionsAgents (pipeline node), Experiences (data binding)
CredentialsPlatform-levelProject-level
Think of Tools as verbs (send email, create ticket, look up a record) and Pipelines as nouns (a pre-built dataset that describes your business reality). Agents use both — Tools to act, Pipelines to understand context.

​
Pipeline Outcomes

A Pipeline Outcome is a named output produced by a Pipeline run — a structured table of data stored in the Internal Data Store. Examples of outcomes:
  • deals_over_10k — all open CRM deals above a threshold, refreshed hourly
  • aggregated_expenses — expenses grouped by category, updated daily
  • customer_support_tickets — recent support tickets with status and priority
Outcomes are referenced by name in Agent workflows and Experience data bindings. They are always available without making a live external API call.

​
Where Pipelines Fit in the Platform

Pipelines sit in the Foundation layer of the Adopt AI platform, alongside Connectors and Tools. 
Foundation Layer
├── Connectors  →  connect to external systems (HubSpot, QuickBooks, AWS S3...)
├── Tools       →  callable actions agents can perform on those systems
└── Pipelines   →  scheduled data extraction and transformation

Orchestration Layer
├── Actions     →  step-by-step instructions agents follow
├── Agents      →  orchestrators that use Tools and Pipeline data
└── Experiences →  HITL dashboards that display Pipeline Outcomes

​
In Agent Workflows

When an Agent runs, it can have a Pipeline node in its canvas. The Pipeline node injects pre-processed data from a Pipeline Outcome directly into the agent’s execution context — no live API call required.

​
In HITL Experiences

Experiences are dashboards where human reviewers inspect and approve agent work. Pipeline Outcomes can be bound to Experience components — tables, charts, and reports — so reviewers see up-to-date data without the agent needing to fetch it in real time.

​
Pipeline Lifecycle

  1. Draft — Pipeline is created and workflow is AI-generated but not yet running.
  2. Active — Pipeline runs on its defined schedule; outcomes are continuously updated.
  3. Paused — Pipeline is temporarily stopped; the last outcome is preserved but not refreshed.
Pipelines run independently from Agents. They pre-compute data on their own schedule so that when an Agent executes, it can access fresh, structured data instantly — without incurring API latency during the run.

​
Next Steps