Accrual recently launched Arc, an AI agent platform built to let accounting firms delegate complex, multi-step work across their existing systems. The pitch is familiar: instead of a person moving a task from your tax software to your document management system to your client portal, an agent does it for you. That's a real capability, and it's worth taking seriously. But before any partner signs a contract, there's a more important question than "what can the agent do": where does the data actually live, and who has to approve what the agent does with it?
We build audit and tax engagement platforms for CPA firms for a living, including a system currently running for a multi-office firm handling 500-plus engagements a year, wired into DocuSign, SmartVault, a client portal, and a firm-wide deadline dashboard. Every one of those integrations forced us to answer the same question Arc and platforms like it now raise at scale: when a system acts on your behalf across multiple tools, who is actually in control of the record?
What "AI agent" actually means in practice
Strip away the marketing and an AI agent platform is an orchestration layer. It reads data out of one system, makes a decision or takes an action, and writes data into another. For a CPA firm, that means touching things like engagement status, client documents in SmartVault, signature requests in DocuSign, deadlines on your tracking dashboard, and possibly client communications through a portal.
That's not a small ask. Those are the systems that carry legal and compliance weight: a signed engagement letter, a filed return, a document a client uploaded under an expectation of confidentiality. An agent that can move work across those systems needs write access to all of them, or access to an integration layer that has write access. The question a partner needs answered isn't "is this AI powerful," it's "what does this agent have the authority to change, and where is the audit trail for that change kept?"
The gap between a demo and your firm's actual workflow
Vendor demos of agent platforms tend to show the happy path: a task gets created, the agent pulls the right document, sends it for signature, updates the status, done. What they don't show as clearly is what happens on the messy path, which is most of the actual work at a firm running hundreds of engagements a year:
- A partner needs to review and approve a document before it goes out for signature, not after.
- A client document in SmartVault has version history and permission rules tied to who on staff can see it.
- An engagement's deadline status needs to reflect actual manager sign-off, not just "the agent thinks this step is done."
- A staff member needs to be able to see exactly what the agent did and reverse it if it got something wrong.
When we built the deadline dashboard for our audit and tax client, the entire design point was that the dashboard reflects the firm's actual approval chain, not just document movement. A document landing in SmartVault doesn't mean an engagement step is complete. A partner sign-off does. That distinction sounds small until an agent platform automates past it and marks something "done" that a partner hasn't actually reviewed.
Six questions to ask before you buy an agent platform
1. Where is the system of record?
If the agent platform sits on top of your existing tools, is your engagement status still tracked in your own system, or does it now live inside the agent platform's database? If the vendor disappears, gets acquired, or changes pricing, do you still have your engagement history, or is it stuck behind their login?
2. What requires human approval, and is that configurable by you or fixed by the vendor?
A signature request going out under a partner's name is not something that should happen without an explicit human gate. Ask the vendor to show you, specifically, where that gate sits in their workflow, and whether your firm can move it, tighten it, or require dual approval for certain document types.
3. What write access does the agent have to DocuSign and SmartVault, and can you scope it down?
An agent that can read status is very different from one that can initiate a signature request or reorganize client folders. Firms should ask for the actual permission scope, not the marketing description, and should be able to restrict it per engagement type or per office if they run multiple locations.
4. Is there a full audit log of every action the agent takes, and who can see it?
If a document goes to the wrong recipient or a status gets marked complete incorrectly, you need to know exactly what the agent did, when, and based on what trigger. That log needs to be accessible to your firm, not just available on request from the vendor's support team.
5. What happens to your workflow if the agent platform changes its own rules?
Agent platforms improve their models and change their automation logic over time. If your engagement workflow depends on how the agent interprets a document or a status update today, a model update six months from now could quietly change how work moves through your firm. Ask how the vendor handles versioning and whether you'll be notified before behavior changes.
6. Does the platform actually understand your firm's engagement structure, or a generic one?
A multi-office firm running audit and tax engagements has a different approval chain than a single-office tax prep shop. Ask the vendor to show, concretely, how the agent's logic adapts to your specific engagement types, not a generic accounting workflow template.
Why this matters more for CPA firms than most industries
Firms are handling client documents under confidentiality expectations, signatures with legal weight, and deadlines with regulatory consequences. An operations mistake at a retail company is a customer service issue. An operations mistake at a CPA firm, caused by an agent that moved a document to the wrong client folder or triggered a signature request before partner review, is a professional liability issue.
That's not a reason to avoid AI-assisted workflow tools. It's a reason to be precise about where control actually sits before adopting one. The firms that get the most value out of automation are the ones that know exactly which decisions are automated and which ones still require a person to say yes.
Buy vs. build, honestly considered
An off-the-shelf agent platform like Arc has a real advantage: speed. It's built already, it plugs into common tools, and a firm can be testing it within weeks. For a firm that wants to experiment with automation on lower-stakes tasks, that speed is worth something.
But for the core engagement workflow, the one touching DocuSign, SmartVault, client portals, and partner sign-off, a generic platform is making assumptions about your approval chain that may not match how your firm actually operates. That's the exact problem we solved differently: instead of adopting a generic tool and adjusting our client's workflow to fit it, we built the deadline dashboard and integrations around the firm's actual approval structure, so the system reflects how the firm works rather than the other way around. The firm owns the platform, the data stays in systems they control, and every action is traceable to a specific approval step they defined.
That's not an argument against every agent platform. It's an argument for asking the six questions above before any of them touch a signature request or a client document. If a vendor can't answer where the data lives and who approves what clearly and specifically, that's the answer.
What to do next
Before evaluating any AI agent platform, sit down with your own team and write out your firm's actual approval chain for a single engagement, from document intake to final sign-off. Then ask any vendor to show you exactly where their agent sits inside that chain, not around it. If they can't map their product to your specific workflow, that mismatch will show up eventually, usually in a client's confidential document or a signature that went out before it should have.