The questions people actually ask before they buy.

You are going to build it yourself. That is the product. The question is only what you start from.

Point a coding agent at an empty folder, ask it for accounting software, and you will get something that demos beautifully and fails quietly. Not because the model is bad — because AI development follows inertia. At every fork it takes the path of least resistance, and it has no idea which forks matter in accounting.

In practice that means it will let the same identity prepare a journal entry and approve it. It will bind an approval to nothing, so a $400 reclass approved on Tuesday can execute as $40,000 on Wednesday. It will hardcode a client’s company ID — or your firm’s name — into a memo that posts to their ledger and stays there. It will proceed on partial data rather than stopping.

You will not catch these on day one. You will catch them the first month you run a real close, which is the most expensive time to find out.

What you are buying is a codebase where those shortcuts are already closed — hundreds of the 1,903 tests exist to assert a refusal — and guardrails that sit between your coding agent and your code: a build-intake gate it has to clear before it edits anything, a runtime where nothing reaches a ledger without a named human approving the exact figures, and five commit-time checks that refuse the commit — credentials, hardcoded client identifiers, mutation calls outside the approval path, undocumented stored fields, and, the one that came out of a real incident, a firm’s name hardcoded into a memo that would post to a client’s ledger and stay there.

You are still the one building. We are the reason your AI builds the right thing instead of the easy thing.

Those are the tools that do the building. This is what you point them at.

A coding agent is very good at writing code and has no opinion about whether a preparer should be able to approve their own work. It will happily build you a system where they can, because you did not say otherwise and nothing stopped it. The eleven questions in the builder are how you say otherwise, once, in business language — and AGENTS.md and the handoff validator are what hold your agent to it afterwards.

Keep your subscription. You will use it. This decides what it is allowed to do with your books.

You need to be able to open a file in your browser and answer questions about your own firm. That part takes about ten minutes and involves no code, no account and no install.

After that you are directing a coding agent in plain English — it proposes, explains, and waits for you to approve. Getting the project onto GitHub is done with GitHub Desktop buttons rather than a command line, and the guide walks you through it a click at a time.

What you do need is judgement about accounting and about your own firm’s controls. That is the part nobody can supply for you, and it is the part you already have.

Worth being precise about, because the answer is the difference between a good purchase and a disappointed one.

Working on day two: the platform source, the runtime controls, the approval path, the tenant isolation, the ASC 606 and ASC 842 engines, five commit-time checks, four worked examples and 1,903 passing tests you can run in front of a reviewer.

Still yours to build: the tool bodies. Ninety-eight tools are declared, classified and gated — the wiring, the permissions and the refusals are done — and the bodies are the build after that. They are also the part that makes it yours rather than everyone’s, and the part nobody could have written without knowing your firm.

Anything unfinished stays listed as work, in writing, rather than being quietly assumed. That is deliberate: a plan that overstates what is ready is how firms end up promising a client something that does not exist yet.

Straight answer: this tier is files, not support. That is what makes it $2,000 rather than a project fee.

Email Owen with the step number and what the terminal said and he will answer — he just cannot promise when. Corrections ship to every buyer at the same link, permanently, with no update to install.

If you would rather have someone in the room while you build it, that is the Bootcamp, and what you pay today credits against a seat in full.

Same code either way. The difference is whether anyone is in the room.

Take the Sprint if you are comfortable working through a well-documented package at your own pace and want to start this week. Take the Bootcamp if you would rather build it alongside someone, on a schedule, with the theory for how your firm keeps building these tools after we are gone.

You are not choosing between them. The Sprint credits in full against a Bootcamp seat and that credit does not expire, so starting with the files costs you nothing if you decide a year from now that you want the room.

Only if you decide it does, and the builder asks you that question directly rather than assuming an answer.

The recommended starting position is no — and if you choose it, the plan records that as a rule the build must enforce, and the validator will flag any configuration that contradicts it as a hard stop rather than letting it through.

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

The controls are built to make that specific thing hard, and they are the part of the package that is finished rather than declared.

Nothing executes without a named human approving it. The person who prepares work cannot approve it. An approval is bound to the exact figures approved, so approving a $400 reclass does not authorise a $40,000 one — that is a worked example in the box, and you can run it and watch it refuse. A service account cannot stand in for a person.

Hundreds of the tests exist to assert exactly these refusals, which means you can demonstrate them to a reviewer rather than describing them.

The terms of sale also put the obligation in writing in the other direction: a qualified person at your firm reviews and approves every output before it is posted, sent, filed or relied on.

Those obligations are yours and they do not move. We are not your counsel and this is not legal advice.

What the package changes is whether you can answer when someone asks. You end up with a written record of every decision you made about who may approve what, where client information is allowed to be processed, and what the software is permitted to do — plus controls that enforce those answers and tests that prove they hold.

If the question you actually have is “what should our firm’s written policy say before we turn any of this on,” that is what the AI Register is for — free, and it takes ninety seconds, which comes off anything you buy afterwards with a credit that does not expire, or is refunded on request within seven days of the session.

The platform is accounting-system agnostic by design — it sits on top of the ledger you already run rather than replacing it. The builder asks which one you use and writes the plan accordingly: QuickBooks Online, Xero, NetSuite, Sage Intacct, Sage Business Cloud, Sage Active, or another system you name.

A reference guide in the package covers getting credentials out of each of the major ones and holding them safely — scopes, sandbox first, and the expiry dates that arrive without warning — checked against each vendor’s own documentation.

The connector itself is one of the things your coding agent builds first, against the provider’s test company before anything real.

Precisely: you get a perpetual licence to source you can read, change and merge into your own systems, and you may operate it in your practice, including in work you do for your own clients. Nothing expires, there is nothing to renew, and nothing stops working if you stop paying us — because there is nothing to stop paying.

You own everything you write on top of it, all of your data and your clients’ data, and everything the software produces for you. We receive none of it; the software sends us nothing. You can pass the licence to a buyer of your firm or a successor in a merger.

What you do not get is the copyright in the original software — that stays with us, which is why you cannot resell or redistribute the package itself as a product or a competing offering. The full terms are on the terms of sale page and they are written to be read.

Yes. You may operate it in your practice, including in work you do for your own clients, and give those clients the results. You may install and run it on a client’s own infrastructure where you operate it on their behalf.

The licence covers one firm and everyone in it — no per-seat count. What it does not cover is handing the software itself onward: your client gets the results and the service, not a right to pass the package on.

The platform talks to AI providers through one common doorway rather than being written for a single company, so changing provider or model is a configuration decision instead of a rebuild. The builder’s recommended setup supports both OpenAI and Anthropic from the start, precisely so you are never negotiating with one vendor.

The parts most likely to age — provider setup, credentials, token policies — live in documentation that ships corrections to every buyer at the same link. There is no update to install and no migration to run.

Seven days, on request, no reason required. Email owen@byram-advisory.com and it is done. The licence ends and you delete your copies, and nobody is going to ask you to justify it.

We ask one thing in return, and the refund does not depend on it: tell us what you expected and did not get. At this stage that is worth more to us than the money.