Building an AI agent to cut down Carta's onboarding time by 63%

TL;DR

Carta is at the forefront of AI-driven development, and we made a strategic bet on incorporating agentic workflows into the onboarding of multi-entity companies. This was previously an entirely manual, undocumented process averaging 18 days per case, and it was one of the first to prove that bet in production. I led design strategy, workflow architecture, and implementation end to end, from designing how humans and Claude agents would collaborate to implementing the skill orchestration system that shipped.

Carta has kept building publicly on that bet since: Carta Plugins on Claude announced by CEO Henry Ward.

Problem

Onboarding is a crucial point for any business and is usually a timeboxed process. Implementation Managers were generally required to onboard multiple multi-entity companies simultaneously, with the expectation of having all customer data ready and in order at go-live. This became less efficient for large private equity firms with multi-entity companies, who usually have voluminous, ungrouped documents and very complex setups, making the process time-consuming and inefficient overall.

Approach

Understanding the current Implementation onboarding workflow. Started with direct observation, mapping how IMs worked through both single-entity and multi-entity onboardings, and identifying the most time-consuming tasks and where an agent could actually help.

Expand

Designing a system where humans and the agent collaborate. This required multiple touchpoints across the IM, Carta, Claude, and Slack, and defining exactly what handed off between them at each step.

Expand

Solution

• Surface ambiguity, don't resolve it. Claude handles the deterministic work. Uncertain extractions, ambiguous matches, unresolved IDs are flagged and routed to the IM, never auto-resolved.
• Checkpoint by default. Every step writes output to disk. IMs resume from any point without rerunning earlier work.
• Source traceability at every step. Every inference traces to the source document. IMs verify claims on live client calls without leaving the plugin.

A Claude Code plugin built as six orchestrated skills, each owning a single domain, outputting structured data for the next step.

onboarding-orchestrator sequences the full flow and handles resume from any step. setup-entitygroup creates the entity group in Carta. link-entity-stakeholder infers and links entity relationships across cap tables. setup-corresponding-interest parses legal agreements to extract share class mappings and CI configuration. link-securities matches grants across OpCo and ManCo by quantity and issue date. multi-entity-provisioning handles downstream configurations post-linking.

Every skill reads and writes through CLI commands built for this workflow, fetching live Carta data, creating stakeholder links, and confirming security matches, so every action is grounded in the same system of record IMs use directly.

Three review artifacts sit at the center of the IM experience: interactive interfaces rendered inside Claude showing the current state of the data, source document overlays for every inference, and inline actions to accept, override, or flag. Designed with deliberate visual familiarity to Carta's core UI so IMs orient immediately.

Expand

Expand

Expand

Expand

Slack delivers the handoff. When an artifact is ready for review or a pre-flight confirmation is needed before a Carta write, a notification surfaces that moment to the IM. The plugin does the work. IMs act when the decision is theirs to make.

Impact

Converted an undocumented, IM-memory-dependent process into a repeatable, auditable system. Onboarding time dropped from 18 days to 6.5 days. IMs handle 2.4× more cases simultaneously. Silent failures on entity linking dropped to zero.

The plugin established the pattern for how agentic workflows are designed and shipped across Carta's internal staff tooling.

Key learnings

• Trust is earned through transparency. The moments that built the most IM confidence were when Claude showed its working: source overlays, flag gates, explicit confirmation before writes. Automation without proof erodes trust faster than it builds it.
• Artifacts take over the interaction. When a visual artifact is present, the chat thread disappears from the IM's attention entirely. What lives in the artifact versus the conversation is a consequential design decision; mixing action surfaces breaks the mental model.
• Tracking agentic workflows is an unsolved problem. Session analysis depended on user-submitted Claude usage reports, not instrumented data. As agentic AI expands across Carta, native observability into where agents act and where they hand off to humans is a critical open design problem.
• The gates matter more than the skills. The hardest design work was defining exactly where Claude stops and waits. Too many gates and the plugin loses value. Too few and IMs lose trust. The right gate is wherever a wrong decision in Carta is irreversible.