Rippling builds AI agents that answer and act on admin questions across its HR, payroll, IT and finance products, according to a recorded fireside chat between Ankur Bhatt, Head of AI at Rippling, and Harrison Chase, CEO of LangChain, published on November 26, 2025.
TL;DR: Rippling's AI work moved from summarization features to standalone AI products and then to agents that act like a systems analyst for admins. Ankur Bhatt describes a shift from per-domain subagents behind a fixed router to a deep agent that reasons over tools, because people phrase the same question in many ways. Rippling ships agents on a shared platform, tests them on production data snapshots, and turns them on internally first. Bhatt reports 150 AI projects in the last hack week, about 50 of them finished (04:56). The chat gives no accuracy or adoption figures for the agents.
| Fact | Detail |
|---|---|
| Source | View source |
| Source type | YouTube fireside chat (LangChain), 44 min |
| Source published | 2025-11-26 |
| Speakers | Ankur Bhatt, Head of AI, Rippling. Harrison Chase, CEO, LangChain |
| Industry | HR, payroll, IT and finance software |
| Function | HR, payroll, IT and finance administration |
| Techniques | Deep agent, subgraphs, workflows wrapped as tools, permission inheritance, tracing, internal rollout |
| Tools named | LangChain, LangGraph, LangSmith, Databricks, Salesforce |
Independent summary of public material. Rippling is not affiliated with Taskade.
The system Rippling built
Rippling sells one suite for HR, payroll and benefits, for IT identity and device management, and for corporate cards, travel and expenses. Ankur Bhatt describes three waves of AI work in that suite (02:43):
- Features that summarize unstructured content.
- Standalone AI products. One example is Talent Signal, which calibrates individual work output across an organization. Another is shift forecasting in the time product.
- Agents that work like a systems analyst for admins and answer complex questions about the account.
One example question is why an employee did not get paid. The agent must trace the payroll data, for example a change of country or address that changed the deductions. Bhatt says these agents take far more design work than a summary feature, because the answers must be accurate.
Architecture of the system
Rippling first built one subagent for each domain, such as IT identity, devices or payroll, with a fixed router in front. Bhatt reports that the router broke on real questions, because people phrase the same request in different ways, for example "onboarded last week" against "hired last week". About a month before the talk, the team started to test a deep agent design (19:32). The model gets a set of tools and example paths, decides which domain a question belongs to, and picks the tools itself. Bhatt says the results surprised him.
Bhatt gives a sales briefing agent as an example of the limit of fixed workflows. The agent summarizes an account executive's intro call with a prospect, writes the summary to Salesforce and creates a document. When the prospect asks about security or legal terms, the fixed flow does not cover it, and the account executive still has to check the output.
The team keeps deterministic steps where an agent takes an action. Bhatt describes a workflow that must capture data in a fixed order as a good tool for the agent to call. The model chooses when to call the tool, and the tool does the work the same way each time. To keep the tool list manageable, the team groups the tools for one domain, such as payroll, into a subgraph with its own paths.
Other parts of the platform:
- A shared base that every product team can use: a data layer on Databricks, an agent layer built with LangChain, and an eval system. Bhatt calls it a "paved path" from prototype to production.
- Snapshots of production data for tests, because a demo instance does not show real behavior.
- Permission inheritance. The agent takes the role of the user who asks, so a user without access to a salary does not get it through the agent.
- Tracing of every production run, which the team uses to find why a call or a route behaved as it did.
Results the source reports
- Ankur Bhatt reports 150 AI projects in the last company hack week, and about 50 of them reached completion (04:56). Rippling runs a hack week every six months (04:48).
- Bhatt reports that the first alpha of each agent goes live inside Rippling, so employees, including the CEO as a super admin, give feedback right away.
- Bhatt reports that outside AI vendors must meet a zero data retention and no-training rule, and that Rippling shut down some pilots that did not fit its responsible AI checklist (27:49).
- For engineers who use AI coding tools, the internal rule is "AI is your superpower but you are still accountable for the code you're pushing to production." (36:25)
Critical assessment
The chat is an interview, not a technical report, and it gives no accuracy, usage or time-saved figures for the agents. The deep agent work was about one month old at the time of the chat (19:32), so the results describe early tests. The host is the CEO of the framework vendor that Rippling uses, and the event promotes that vendor's tools. The hack week numbers describe internal experiments, not shipped features. No outside party checked any claim. The source is a useful record of why one large HR and payroll vendor left fixed routing.
Build this in Taskade
- Build an AI agent with persistent memory and file analysis that answers questions about your HR policies. Start at AI agents.
- Start from a template in human resources agents.
- Run onboarding and request steps with HR automations.
- Send a request to an automation from another system with a webhook trigger, on Pro or higher. See the webhook trigger.