Startups are putting agents to work on core financial operations but are still selective on what they trust agents to do.
Stripe interviewed founders and builders at eight venture-backed startups, ranging from seed stage to Series B, to understand which tasks startups trust agents to handle, which ones still require human approval, and what would make them comfortable delegating more. Participants had deep, hands-on experience with AI, either by using agents weekly to automate finance work or shipping agentic finance products for customers.
Across the group, participants described using agents—many home-built using tools like Claude or Codex or off-the-shelf agents customized for their needs—to automate spend analysis, cash flow modeling, transaction reconciliation, and anomaly detection. However, none had authorized an agent to move company money without human approval. An incorrect wire can tie up operating cash or disrupt a service customers depend on. A payroll error can leave employees unpaid.
Before enabling agentic money movement, those in our study said agents must earn authority over time. Agents first need to demonstrate that they can correctly process business context and prepare accurate work for human approval. As confidence grows, teams are more likely to authorize clearly defined actions, such as transferring funds between named accounts, paying approved invoices, and converting currency within set limits.
This guide details the work startups are delegating today and outlines a path toward more autonomous finance flows for managing and moving money.
Agents give startups more capacity to grow
Stripe Economics data finds that AI provides flexibility in deciding when to add headcount. When agents take on recurring finance work such as document processing and expense review, founders can manage increasingly sophisticated financial operations before building a finance team.
The three-stage progression toward autonomous finance
Founders in our study said they would be willing to give agents more responsibility as they build confidence in the agents’ output. That delegation progresses through three stages where agents:
- Read and analyze data
- Prepare work for human confirmation
- Execute transactions autonomously
Stage 1: Agents read financial data and surface insights
The founders in our study use agents most often for read-only financial work. Agents surface information the finance team can act on: unusual spending, missing receipts, cash flow changes, duplicate subscriptions, and current financial models. At this stage, an error does not move funds or change a financial record. Each useful result adds to the agent’s record of accuracy and supports greater authority in the next stage.
Common use cases
- Track expenses and burn against targets
- Analyze cash flow and build scenario models
- Flag spending anomalies and missing receipts
- Parse invoices and match transactions across systems
Action for founders: Test early outputs against known business conditions
Start with a defined, read-only task that draws on relevant financial data and business context. Give the agent only the financial data and business context it needs for the task. Then compare the agent’s early output with conditions the team already understands. This helps the agent distinguish a real problem from an expected change in the business.
Use case: Board-ready models in minutes
Founders in our study use agents to keep NRR tracking, growth models, and cash flow scenarios current for board meetings and fundraising. Work that once consumed a weekend of spreadsheet updates can run against live financial data in minutes. The founder of a Series A security startup said: “I can generate cash flow reports, custom queries, anything I want very quickly, and I’m not a technical CEO. This has been a massive step change for me.”
Stage 2: Agents prepare financial work for human confirmation
Here, agents move beyond analysis, preparing financial records or queuing actions for a human to review and approve. Most startups in our study are comfortable allowing agents to handle low-risk work like preparing expenses or payments, as long as a person signs off before the work goes live.
Common use cases
- Suggest expense categories for the general ledger and reconciliation matches across systems
- Prepare vendor details and set up payment recipients
- Draft payment runs based on approved invoices, due dates, and terms
- Propose card limits, controls, and freezes for review
- Generate accounting entries and tax-preparation materials
Action for founders: Create a feedback loop for agent review
Use human reviews to build confidence in the agent over time. Capture corrections to improve future suggestions and create a track record of reliable output that can support greater responsibility. This verification loop can also reduce the human effort required for future reviews.
Use case: Reconciliation at scale
One respondent in the insurance tech industry uses a Claude Code agent to process more than 100 PDFs from the insurance carriers the company works with, including monthly commission and billing statements. The agent extracts invoice data, matches insurance carrier statements, and supports reconciliation across systems—tasks that previously required one to two full-time employees. The COO of a Series B startup said, “Straight money matching should never be touched by a human—reviewed, sure, but automated. We’re still doing manual reconciliations until we have confidence. But once we do, it’ll save us from having to hire.”
Stage 3: Agents execute financial actions within policy
This is the stage startups in this study are building toward: agents handling important but recurring finance work without someone approving every step of the way. Getting there requires trust built over potentially thousands of agentic tasks and tools that can properly control access to the financial data the agent needs.
It’s [the same way] you develop trust with another human: the more correct[ly] the agent does something, the more you trust it.”
Common use cases
- Move funds between approved accounts to maintain target balances
- Pay approved invoices within amount, vendor, and timing rules
- Convert currency when a balance or rate reaches a preapproved limit
- Issue and manage cards within predefined spending controls
- Initiate refunds that meet eligibility and amount rules
- Support machine-to-machine payments and delegated spending
Use case: Resolving eligible refund requests
Dust, an AI platform for building and sharing custom agents, illustrates what this stage can look like. Its internal refund agent combines customer context and the company’s refund policy with authorized Stripe data via the Stripe Model Context Protocol (MCP) server. The agent reviews the customer’s history, validates the payment, applies the policy, and triggers an eligible refund in Stripe before drafting a personalized confirmation email. Dust reduces the refund process from about an hour to seconds. “Instead of drowning under a long-tail of low value tasks, teams get to maximize time spent on critical work, and customers can feel it,” said Nicolas Chinot, US general manager at Dust.
Action for founders: Run a pilot for agentic money movement
Once an agent progresses through stages 1 and 2, run controlled money-movement pilots to validate its capabilities. Start with nominal, one-off transfers between internal financial accounts, and then test recurring or scheduled transactions. Record every action it takes and the policy that allowed it, and pressure-test escalation rules to ensure a human gets flagged to review exceptions and can pause or reverse the transaction.
Balancing access and control for agents
As an agent takes on more responsibility, it needs access to the financial data and tools that provide context for how the business operates. A single payment record, for example, does not explain whether a sudden spending spike is a new growth initiative or a service price increase. Broader data around revenue, expense, and even fundraising helps the agent accurately interpret what that spike means.
These [agents] don’t know about your architecture, they don’t know about what’s happening in Slack, they don’t know about this new product release that caused something to increase in spend.”
Unlocking that full view requires access to the systems where the data lives. But that doesn’t mean granting agents broad permissions across each one. Agentic finance infrastructure, then, needs to do two things: give agents the context they need and give startups precise control over what agents can access, prepare, and execute.
How Stripe powers agentic finance
Most startups are building agents around their own tools, data, and processes. Stripe makes financial data and money management capabilities available to the agents a business chooses, with clear controls over what those agents can access and do.
With Stripe MCP, businesses can connect AI tools to the Stripe API as well as Stripe documentation. Stripe Financial Connections securely adds external account balances and transaction data. Stripe is also developing support for agent workflows through Stripe CLI and Stripe Console, our built-in Dashboard assistant.
Stripe Treasury for agents extends these workflows to money management. Agents can analyze financial data, move funds between supported Stripe balances, create recipients, send payments, and convert currencies. Granular permissions control agent access, and actions that move live money require human confirmation by default.