AI Workflow Automation for Unstructured Work
Zapier automates the predictable. But most real work is unstructured — a message, a decision, a blocker. Here's how Crestline turns plain English into a running workflow.
Classic automation tools are great at the predictable: when a row is added, send an email. But most of the work that actually slows a company down is unstructured — a decision made in a thread, a task that just got blocked, a new hire who joined yesterday. You can't pre-wire every branch of that, and writing it as a flowchart by hand is its own full-time job.
Crestline takes a different path: you describe the automation in plain English, and the system builds it.
Five agents turn a sentence into a workflow
When you type something like "every Friday at 4pm, email each team lead a summary of their team's open blockers," Crestline runs a five-stage pipeline:
- Intent — restate the request as one precise sentence: trigger, action, destination.
- Planner — propose the concrete steps in plain English, grounded in your org context.
- Drafter — emit a typed workflow specification as JSON.
- Validator — check it with a schema and cross-step rules, soft-repairing obvious gaps before falling back to an LLM repair pass.
- Narrator — write a two-to-three sentence explanation and propose a name and tags.
You watch each agent light up as it finishes, see the trace, and get a draft you can activate, save, or tweak. The plan is never a black box.
Triggers and steps that understand your org
What makes this more than a prettier Zapier is that the building blocks are native to your organisation, not generic webhooks.
Triggers fire on real events: a task assigned, a status flipped to blocked, a mention received, a member created, a schedule, or an inbound webhook. Steps include flow control (branch, wait, parallel), AI operations (summarise, extract), integrations (email via Resend, generic HTTP), and Crestline-native actions — create a task, write to a memory layer, reply in a channel, notify a person, or log a verified decision.
Org-graph tokens instead of hard-coded emails
Here's the detail that matters at scale. A workflow never stores a person's email address or a raw ID. Instead it uses org-graph tokens that resolve at run time:
assignee: @manager_of(@self)
notify: @dept-finance
escalate: @cxo
When someone changes managers or switches departments, the workflow keeps working — because it resolves the relationship, not a frozen identifier. Your automations survive your org chart changing, which is exactly the thing that breaks brittle automations everywhere else.
Automating the work that was never a form
The unlock isn't faster forms-to-email. It's that unstructured events — a blocker, a decision, a mention — can now trigger structured, auditable action, with a full run trail for every step. The work that used to require a human to notice, interpret, and chase finally moves on its own.
See it in action on the Workflows page, or get started.