Scaling Financial Software with GenAI and Production ML

ramp · Finance · 2023 · source Agentic: yes · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: financial operations software and production ML

Ramp needed to use AI across expense management, underwriting, fraud, accounting, sales, and early agentic product workflows while respecting reliability requirements in finance.

Constraints:

• Financial workflows involve high-stakes decisions such as underwriting, fraud, compliance, and accounting.
• Some workflows are productionized, while agentic UI-control features remain experimental.
• AI must simplify operational processes without reducing finance control or reliability.

Architecture

Ramp uses an applied AI team with a broad remit, production ML for underwriting/fraud/receipt/accounting workflows, Ramp Intelligence for price, expense, and accounting assistance, and alpha agentic capabilities that can observe the UI, guide users, click through flows, and take bounded actions such as issuing cards or booking travel.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["Applied AI team"] --> B["Manual finance workflows"]
  B --> C["Data signals"]
  C --> D["Ramp Intelligence suite"]
  A --> D
  A --> E["Alpha agentic assistant"]
  E --> F["User interface"]

Components

Data flow

Applied AI teams identify manual tasks, use data signals and automation to reduce decision time, and productize successful capabilities into Ramp Intelligence or internal workflows. For agentic UI assistance, a multimodal model sees the user's screen, guides navigation, and can take actions in controlled alpha scenarios.

Why this architecture

Ramp applies different AI forms to different finance problems: production ML and GenAI simplify routine workflows, while more autonomous agentic UI actions stay in alpha because reliability varies by task.

Models used: GPT-powered Ramp Intelligence · multimodal models · traditional ML models

Scale

• Ramp was valued at $7.6 billion in the referenced funding round.
  “raised a huge round at $7.6 billion valuation.”transcript, stated
• A prior agentic workflow at Paribus processed over 100 million emails a day.
  “assess vast amounts of data, over 100 million emails a day, and send price adjustment requests.”transcript, stated
• Ramp says average SDRs book three to four times as many meetings as the next closest competitor.
  “books three to four times as many meetings as their next closest competitor.”transcript, stated
• A manual underwriting process fell from two days to roughly a quarter to half day in a two-week sprint.
  “It could take an underwriter two days to do that. They sat side by side with the underwriter, and I think within a couple, I think like a two week sprint”transcript, stated

Alternatives considered (roads not taken)

Buy or build a monolithic AI SDR

Rejected because: Ramp instead decomposed the SDR workflow into data and automation pieces that make human SDRs more productive.

Expose broad autonomous agents immediately

Rejected because: Ramp keeps agentic UI-control use cases in alpha because task success varies.

Force users to learn every SaaS workflow manually

Rejected because: Ramp is testing agents that can see the UI and guide or execute actions.

Patterns

tool-use-orchestration (agent-architecture)

The alpha agentic assistant can observe the product UI, click tabs, enter inputs, issue cards, and book travel/hotels.

workflow-decomposition (proposed)

Ramp decomposes sales and finance workflows into data, decision, and automation pieces instead of automating whole jobs at once.

human-in-the-loop-gate (governance)

Ramp frames AI as helping compliance managers and finance users focus on the right decisions rather than removing their role.

Tradeoffs

Agent rollout

Chose alpha group for UI-control agents over general production launch.

Rationale: The agent can take real actions but only works reliably for a portion of tasks.

Sales automation strategy

Chose workflow streamlining over buying an AI SDR.

Rationale: Breaking the workflow into data, list-building, email lookup, and copywriting can make SDRs more productive without replacing the role.

Gotchas (what bit them)

Agentic UI control is still reliability-limited.

Root cause: Task success varies substantially by workflow.

Resolution: Keep the feature in an alpha group and test bounded actions before broad launch.

Finance automation should improve human leverage rather than hide risk.

Root cause: Underwriting, fraud, receipts, and compliance require business accountability.

Resolution: Use data signals and AI to shorten manual workflows and focus managers on exceptions.

Evaluation

Ramp evaluates impact through underwriting time reduction, SDR productivity, product readiness, and alpha-task success rates.

Metrics: underwriting time · SDR meetings booked · agent task success · workflow automation impact

Outcomes

• Ramp Intelligence production features support price, expense, and accounting workflows.
• An underwriting workflow was reduced from about two days to roughly a quarter or half day.
• SDRs reportedly book three to four times as many meetings as competitors.
• The alpha agent can guide users and take bounded product actions, but success varies by task.

Agentic capabilities are explicitly alpha; production claims are stronger for Ramp Intelligence and applied-AI workflow improvements.