Financial operations platform
Processing Content
“Accounting firms are under more pressure today than at any point in history,” said Geoff Charles, chief product officer at Ramp. “The firms we work with aren’t asking for another AI tool to prompt. They need something that actually does the work, with every decision reviewable and auditable. That’s what Stack is built to do, starting with the close, and expanding into every workflow that stands between their team and higher-value client work.”
Stack is said to allow accounting firms to deploy AI agents to execute any accounting task from start to finish. Among many other things, it can build and update recurring schedules (for fixed asset depreciation, prepaid amortization and deferred revenue) and post the resulting journal entries each period. It can audit a client’s existing books to surface duplicates, mispostings and unreconciled periods, then draft the correcting entries to fix them. The product will pull payroll reports, computing employer/employee cost splits and department allocations, and posting the resulting entries. It will be able to pull a summary of all changes made to a client’s books by user and time period. It can review and code bank feed transactions based on historical GL patterns; and compute and reconcile monthly or quarterly sales tax for clients, with agents picking up the prior session context.
The agents assigned to the task come loaded with the proper methodology, data sources and output format to complete the task, pulling data from connected systems to do the work and surface the result. This output might be an analysis, a formula-based workpaper, a journal entry ready to post, or all three. The agents can be told to run a task end-to-end or to pause at set points for human review and approval. Firms operate the solution from a single command center for their entire product portfolio, the Advisor Console, where partners and managers can see close status for every client, track open tasks and see who’s working on what.
Agents operate using Skills, an editable document that details the standard operating procedure for a specific workflow: how a task should be approached, what data sources to use, how the output should be structured, and how to handle the particulars of a given client. Stack comes pre-loaded with Skills covering the most common accounting workflows, and firms can edit any pre-loaded Skill, update existing ones as their processes evolve, or create entirely new ones for their own workflows and client needs. A workflow captured once runs consistently across the entire client book, regardless of who’s doing the work.
When an accountant assigns a task to a Stack agent, it first loads stored memories for the client and identifies which Skills apply to the task type. Then it queries connected systems directly, pulling relevant general ledger accounts, transaction history and reports from QuickBooks, reading uploaded files from Google Drive, and surfacing every data source accessed. If the agent identifies an ambiguity, it raises questions before proceeding. Then the agent presents a structured plan for human review before executing anything. Once the plan has been approved, the agent builds the workpaper and drafts the journal entry. The accountant reviews and approves, and the agent posts the journal entry to QuickBooks.
Afterward, Stack writes a memory describing how to approach that task for this client specifically: which GL accounts are relevant, which source files matter, what the calculation methodology is. This functions not as a log but a reference for next month. Finally, Stack may suggest updating the relevant Skill based on what it learned. The firm accepts or declines. It is through Skills that agents can learn and execute specific firm methodologies.
Underneath the hood, Stack agents rely on what Ramp calls a harness: the full system of tools, Skills, memory and connectors that surround the model and define how it receives tasks, accesses data, processes information and produces outputs. Ramp said that while the model is the reasoning engine itself, the harness is what makes that reasoning useful for accounting work specifically.
In the case of, for example, a payroll entry, the harness informs the agent of the firm’s documented process for structuring payroll entries, account mappings and department splits, gives the agent the connectors to pull the payroll data directly from the source, and the directions for structuring the workpaper. And once that’s approved, it can post the resulting journal entry directly to QuickBooks.
Stack connects directly to data sources like QuickBooks Online (for GL, transaction history, chart of accounts, read/write access), Google Drive, and external bank accounts and credit cards via Plaid, so all client transaction and spend data can be pulled and analyzed from one place. For data that lives outside a connected system (such as a payroll report or vendor contract), files can be uploaded on a one-off basis and the agent will incorporate them into its work. Stack itself is meant to be a central connectivity layer regardless of whether the firm’s clients are using Ramp or have a Ramp account. But for firms whose clients are also on Ramp, the card, bill pay and treasury data are natively connected as well.
Ramp noted that every action Stack takes is logged with the data sources used and the reasoning behind it. An accountant, or an auditor reviewing the client’s books, can trace from any journal entry back to the exact agent session, the workpaper and the source data that produced it. Meanwhile, the workpapers Stack produces are formula-based: Every number in the schedule is tied to its source, the same way a well-prepared manual workpaper would be. The entry and the supporting documentation are produced together. If the agent encounters a situation that requires human judgment, it surfaces the question before proceeding. Ramp noted that nothing posts to a client’s GL without explicit human authorization.
Meanwhile, client data is encrypted in transit and at rest, and segregated by firm and client. No firm has visibility into any other firm’s data; within a firm’s workspace, client data is scoped to that client. Further, Stack runs on foundation models operating under zero data retention agreements, meaning that prompts and data are not stored by the model providers after the session ends, and no customer or client data is used to train the models Stack runs on. The Skills, checklists and procedures a firm builds in Stack are that firm’s intellectual property. They are scoped to that firm and not accessible to any other firm on the platform. A firm’s proprietary approach to a particular accounting workflow remains exclusively its own.
While QuickBooks Online is the first accounting system Stack connects to, NetSuite and Sage Intacct are also slated for integration, followed by additional payroll, commerce and banking integrations prioritized by what partner firms need most. The long-term goal is for Stack to be connected to every system in a firm’s tech stack, so agents have everything they need to handle any client’s accounting work without manual data prep.
The company plans to further expand the solution’s capacities in the future. While right now it covers core CAS and bookkeeping workflows, Ramp plans to eventually expand it into tax preparation, audit support and advisory-level analysis as agent capability and data connectivity grow. As those capabilities come online, firms will be able to bring more of their service lines onto Stack without switching platforms.
Credit: Source link
