Straight answers. What you get, what you keep, and what it costs.

You will build it yourself. That is the product. The question is where you start.

Point a coding workspace at an empty folder. Ask it for accounting software. You may get a great demo that fails in quiet ways. The model takes the easy path. It does not know which choices matter in accounting.

It may let one person both prepare and approve an entry. It may not tie a yes to the exact figures. It may put a client ID in a ledger memo. You may only find these gaps in a real close.

The Sprint starts from tested controls. Hundreds of its 1,965 tests check that it says no when it should. Four handoff files record the result you want and the limits that matter. Five checks inspect the code at each commit.

The tests prove the parts that ship. They do not prove a live client build. Each real write must still pass through the controls. You must prove there is no second path.

Once you say yes to the result and the limits, your workspace builds the platform you chose. It does the routine technical steps on its own. It does not ask you to approve each one.

You are still building it. But you start from clear, tested rules, not an empty folder.

Claude Code and Codex can write code. The Sprint Workspace we support adds a firm plan and firm rules on top.

You use your own ChatGPT Work local project, with Work locally turned on. Or your own Claude Cowork task, with the permanent folder chosen. What works depends on the app, the plan, the settings, and your access. We do not bundle either service.

Quick Start asks you three business questions. A seven-stage path lets you add more detail. The four handoff files, the rules file, and the checker hold the workspace to what you chose.

Once you say yes to the result and the limits, it builds the platform you chose. It does the technical steps on its own. It does not ask you to approve each one.

Other coding workspaces need the same things. The same root. The same handoff. The same checks. The same limits on what they may do.

You need to open a file in your browser. Then you answer three questions. What is the work? Which accounting product should it prove first? What is the first useful result? That takes no code. No Byram account. Nothing to install. If you want to change the choices under that, there is an optional seven-stage Customize path.

You say yes to the business result and its limits. You do it in plain words. Then the Sprint Workspace we support does the routine work. The design, the code, the tests, the review, the fixes, and the commits. It comes back to you for three things only. A fact you have not given it. An action only the owner may take. Or a new limit that matters.

For that route you need your own ChatGPT Work or Claude Cowork access, of a kind that works with it. You may also need GitHub, hosting, and provider accounts later. The guide walks you through each one at the right step.

What you do need is judgement. About accounting, and about your own firm’s controls. No one can give you that. And you already have it.

Only if you decide it does. Quick Start starts from the safe default: client data is not sent to an AI provider. The optional Customize path lets you change that.

The handoff records your choice as a rule. Its checker refuses a plan that is missing, changed, or does not match. The finished build still has to enforce that rule. And its go-live evidence has to prove the real data path.

The planner itself sends nothing anywhere. It runs in your browser, from a file on your machine. No account. No network calls. Nothing you type in it reaches us, an AI company, or your accounting system.

The package ships tested controls for that exact risk. A person you name must say yes. The one who prepares is never the one who approves. And the yes is tied to the exact figures. The worked example proves that a yes to a $400 reclass is not a yes to a $40,000 one. Hundreds of tests check that it says no in those cases.

The package does not include a certified live connector. So those tests prove the controls that ship. They do not prove a full client build, end to end. Before any real write, the finished build must send every write to your accounting system through those controls. And it must prove there is no second path.

The terms of sale also put a duty on your side. It is in writing. A qualified person at your firm must review and approve every output. That happens before it is posted, sent, filed, or relied on.

Those duties are yours. They do not move. That includes keeping client data private. We are not your lawyer, and this is not legal advice.

The handoff gives you a written record of the rules you chose. Who may approve what. Where client data may be handled. What the finished software may do. The package includes tested control foundations. But the finished build still has to put those rules in place. And its go-live evidence has to show they hold in the real provider and the real deployment.

Is your real question which AI tools already touch client work, and what proof you hold for each one? That is what the AI Register is for. Free, private, and about ninety seconds.

QuickBooks Online has a small reference adapter. It is built. It is not a turnkey or certified live connection. If you pick QBO in Quick Start, that does not claim one. On Xero, NetSuite, or Sage Intacct? Then the adapter is built as part of your firm’s engagement. It sits on contracts that do not depend on the provider.

The package has a reference guide for this. It covers how to get the keys and logins out of each system. And how to hold them safely. The scopes. The test company first. The expiry dates that arrive without warning. It is checked against each vendor’s own documentation.

Each path to a provider is proved on its test company first. Nothing real is touched before that.