Customer support · Production

Ramp builds AI Tour Guide agent to help users navigate its financial operations platform

The problem

Ramp's platform has many layers of functionality and users face an onboarding curve, requiring time to become platform experts. Ramp wanted users to self-serve faster without calling customer support.

First attempt

An initial multi-step agent making two separate LLM calls — one for planning and one for grounding — was accurate but too slow for an acceptable user experience. Context stuffing with user screenshots was also found to be less effective than focused, well-enriched interactions.

Workflow diagram · grounded in source

Classifier routes query

routing

“the Ramp team developed a classifier that intelligently identifies relevant queries and automatically routes them to the Tour Guide feature when appropriate”

App state ingestion

ai_action

“The Ramp team designed the agent to take as input the current state of the web app session and suggest the next best action”

Single action generation

ai_action

“the agent generates exactly one action – scrolling, clicking, or text fill – at a time. The resulting altered session would then be fed to generate the next action on the tour”

Cursor control with explanations

output

“Tour Guide takes control of the user's cursor to perform actions a human would do in Ramp (e.g. clicking a button, navigating a dropdown, or filling out a form). As the AI navigates through the interface, it provides step-by-step explana…”

User interrupt or takeover

human_review

“Users can see all the agent actions and interrupt or take control of the agent at any point, rather than just running it in the background”

Reported outcome

Tour Guide increases user productivity and platform accessibility while building user trust through step-by-step transparency.
Letter-labeling of UI elements in the prompt led to a significant improvement in output accuracy.

Reported metrics

User productivityincreases user productivity

Output accuracysignificant improvement in output accuracy

Reported stack

LangChainLangSmithLangGraph

Source

https://www.langchain.com/breakoutagents/ramp

Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Tour Guide increases user productivity and platform accessibility while building user trust through step-by-step transparency.

What tools did this team use?

LangChain, LangSmith, LangGraph.

What results were reported?

User productivity: increases user productivity; Output accuracy: significant improvement in output accuracy (source-reported, not independently verified).

What failed first in this deployment?

An initial multi-step agent making two separate LLM calls — one for planning and one for grounding — was accurate but too slow for an acceptable user experience.

How is this customer support AI workflow structured?

Classifier routes query → App state ingestion → Single action generation → Cursor control with explanations → User interrupt or takeover.

Tour Guide increases user productivity and platform accessibility while building user trust through step-by-step transparency.Letter-labeling of UI elements in the prompt led to a significant improvement in output accuracy.