AI Implementation — for accounting firms
Four live weeks. You leave owning the software.
Financial software specific to your firm — your ledger, your chart of accounts, your thresholds, your approval limits — built on real platform source, running on infrastructure you control, automating work your people do by hand today. Yours to keep and modify under a perpetual licence — not a subscription, not a seat, not an automation living inside a product somebody else can reprice.
And the harder half: a method for running the practice around it. What to build, what to buy, what to leave alone, what a control has to do before it goes near a client, and what you say when a peer reviewer, a client’s counsel or an insurer’s renewal questionnaire asks whether a human reviews the work.
Written by Owen Byram — a career in private equity finance running the close, then about 2,000 hours building automation software that drafted to client ledgers in production — the case study is that architecture in full. Roughly 73,000 words, six working chapters, and every price in it carries the date it was checked.
What you leave with
Each chapter ends in something that exists.
The six working chapters run across four live weeks. These are their actual promises—not learning outcomes, but things that exist when you finish.
AI and the way the 99% use it
By the end: a written model policy for your firm — which model does which job — backed by your own measured numbers on your own real work, not by anyone's benchmark.
Make it stop starting from zero
By the end: a context file per client, proven against a task you already graded — and ninety minutes of your most senior person's head on paper, in a form the rest of the firm can use.
This chapter alone pays for the programme. The chapter has you count the interruptions your firm currently absorbs because an answer lives in one person's head, price them properly, and then halve the number twice. It still pays back inside a month.
Let it read your real systems
By the end: a live, genuinely read-only connection to a system that matters — and one automation running off it that tells you something you would have gone and looked up.
Controls, in the order that makes the argument
By the end: one automation drafting posting-adjacent work with controls a reviewer would accept, and an always-on agent telling you what you are spending before you think to ask.
Vibe coding, and the gap you can't see from here
By the end: something you built yourself, running — and a costed answer to the only question that matters: what would have to be true before you would let it near client data.
Bringing it together, and watching it refuse
By the end: the refusal tests from chapter four driving your own code instead of skipping past it, one client's work running end to end without you carrying anything between the steps by hand, and a written statement of the one thing your system still cannot express.
What the six add up to
Six exercises is not the point. One system is.
Read as a list, chapter one to chapter six looks like six separate pieces of homework. They are not. Each one is a layer of the same thing, and chapter six is where they stop being separate — one client’s work running end to end, with nobody carrying anything between the steps by hand.
You leave owning software
- Bespoke to your firm, in your own repository, on your own hosting, on a model key that is yours
- Built on the complete platform source — not a teaching subset, and not a sandbox written to illustrate a point
- One firm, unlimited seats, permanently. Nothing to renew and nothing that stops working if you stop paying us
- Modifiable, because you have the source. A different model vendor or a different ledger is an edit, not a ticket
And a framework for running the practice around it
- Which work is worth automating, and — more often — which is not
- What a control has to do before you put its output in front of a client, and how you demonstrate that it refused
- What the whole thing costs a month, and how you would know whether it paid
- Who opens the approval queue on a Tuesday, and what happens the morning it says something nobody expected
That second column is AI-1, and it is six layers built in order: 1.1 Foundation — may we do this at all · 1.2 Frame — what the agent may ask for and what it may never do · 1.3 Systems — what it reads, what it writes, what it costs · 1.4 Inspection — how you prove it refused · 1.5 Occupancy — who uses it and who approves · 1.6 Upkeep — expiring credentials, changed models, drift. The Sprint is layers 1.2 to 1.4, the three that are code. The cohort is all six, against your own firm.
The layers that make this defensible are not the ones that make it work. Most programmes teach the second set.
Two ways to take it
Build it yourself. Or build it with me in the room.
In a world where everyone wants to sell your firm another software subscription, we would rather hand you the tools to do it yourself. Both routes end in the same place — your firm running AI it can actually defend — and both hand you the same code, because it is the same package. Not guardrails written to illustrate a point — we have built production-grade systems like this before, and this is that platform source, scrubbed of credentials and client identifiers and nothing else. The case study is what one of them did: the close, the flux review, the 13-week cash forecast, covenants, AR and AP, board reporting, FP&A, and the technical accounting under ASC 842, 606, 326 and 205-40. A seat includes the Platform in full. What the cohort adds is the six chapters of reasoning behind it, a weekly session in which every control is made to fire against your setup, a 1:1 on your own book, and a room of other firm owners doing it the same month. Buy the Sprint alone and it credits in full against a seat, so starting on your own costs nothing if you decide you want the room.
Build it yourself, on your own clock
The Sprint
- The complete platform source — the tool registry and the eight-invariant gate that refuses to boot on a half-registered tool, ninety-eight real tool schemas, the ASC 606 revenue engine, the ASC 842 lease engine, the standards corpus and the row-level tenant isolation. Not a description of how we build agents. The source
- The runtime controls — the approval queue, fingerprint binding, the idempotency ledger and the spend monitor — plus five commit-time checks and the four project files, installed as a pre-commit gate in shadow mode first, so you see what they catch before they can block a close
- 1,903 tests, hundreds of them asserting a refusal. Run them in front of a reviewer
- A guided builder that opens in your browser and turns your business decisions into build instructions — four files out, including the prompt you hand your coding agent. About two days to the foundation. No theory and no hand-holding — that is the trade
- A browser start guide — a plain-language path from download to a private repository and coding-agent handoff, using GitHub Desktop instead of Terminal Git setup
- A provenance file naming every source file, when it was written, and whether it is platform code, extracted declarations, or written for the package — so you can check what you bought, and a customization file naming every decision the code is making on your behalf
- Yours permanently. Corrections ship to every buyer
Pay, and Stripe returns you straight to a page carrying the zip, the current start guide and the build order — no email to wait for. See what is in the box first if you want the detail.
Four live weeks, capped seats
AI Implementation Bootcamp
- Everything in the Platform — the full package, not a teaching subset — plus the chapters behind it and the three things a document cannot do: the sessions, the 1:1, and the room
- Four weekly sessions where the controls you build are made to fire in your environment, because a control you have never watched refuse is one you are trusting on faith
- Four 1:1s, on your own repository. You add us as a collaborator under a written confidentiality undertaking we sign first; we pull your branch, reproduce the error and push a fix as a pull request you review and merge. We never receive your ledger credentials, and the access ends with the cohort
- Half of what you leave with involves no code at all — your firm’s written memory, your client context files, and a signed AI use policy where every clause points at the thing that proves it
- A small room of other firm owners solving the same problem in the same month
Work through the first two weeks. If you have nothing running that you didn’t have before, we refund all of it — ask any time up to two weeks after the first session.
Not sure it is for you? Book a close teardown → free, forty-five minutes, and no obligation either way.
The next cohort starts 15 September. Seats are capped, and we will tell you honestly whether the Sprint on its own would serve you better.
The short version of the difference: the Sprint hands you the source and the order to build in. The Bootcamp builds it with you, in your systems, against your own book. Either way the software is yours at the end — what four live weeks buys is that it is running.
Fit
Who this is for, and who it is not.
It is for you if
- You run a firm, lead a tax, audit, advisory or client-accounting practice inside one, or carry a book as a fractional CFO
- You are professionally liable for being wrong, and that shapes what you will accept
- You are numerate, sceptical, and have been sold to badly by an AI vendor already
- You want to know what to build and what not to
- You can find three or four hours a week for four weeks
It is not for you if
- You want somebody to do it for you — that is a custom build, and we will say so
- You have never used ChatGPT or Claude — read the close teardown first — it is free and takes forty-five minutes
- You want a certificate. There is no certificate
- You want to be told AI will replace your staff. We argue at length that it will not
On specialty: the worked examples run against a general ledger, because that is the system we build against. The method does not care. A context file that carries your firm’s positions, thresholds and the work that always needs a second pair of eyes is the same artifact whether the agent is reading a prior-year return, pulling figures out of a workpaper, or proposing a journal entry — and the control that makes it refuse is the same line of text. What you are buying is the discipline, not the dataset.
Why it reads differently
It repeatedly tells you not to spend money.
Running models on your own hardware will not save you money, and we do the arithmetic to show it. A platform spend limit is often an alert rather than a wall, so do not treat it as a control. QuickBooks Online has no read-only permission, which means the read-only you think you have is something you have to build. Most of what your firm should automate needs no approval machinery at all — and the chapter that could have sold you the heavy version spends a section arguing the opposite.
In a category where everyone oversells, candour is the entire differentiator — and you have already bought once on the strength of it.
Before you start
What you will need.
Questions
The ones worth asking.
Do I need to be technical?
No, and the material assumes you are not. It does assume you are numerate and that you understand internal controls, materiality and segregation of duties — because those turn out to be the ideas the whole thing is built on. Engineers learn this year what your profession has known for a century.
Will this be out of date in six months?
Parts of it, and we say so on the page. Every price and product capability carries the date it was checked, which is the main reason it is published as web pages rather than a PDF — updating is trivial and you keep access. What will not go stale is the method: measuring on your own work, writing decisions down, read-only first, and the control sequence in chapter four.
What is the difference between the Sprint and the Bootcamp, really?
Not the files. A seat includes the Platform in full — the complete source, not a teaching subset — so nobody is reading an abridgement. The cohort gets three things a package cannot carry: a weekly session in which each control is made to fire in their environment, a 1:1 on their own book, and a small room of other firm owners solving the same problem in the same month. And it goes two layers further: the Platform is AI-1 layers 1.2 to 1.4, the parts that are code, while the cohort takes you through 1.5 and 1.6 — who actually uses the thing, and what keeps it alive after the first credential expires. So: if you want somebody watching the controls fire and a room to do it in, take the cohort. If you would rather build alone, the Sprint is the honest choice, it credits in full against a seat, and we would rather tell you that than not.
Do I have to connect a client's live systems?
Not until chapter three, and not at all if you would rather not. Chapter three covers exactly what your engagement letter and professional standards have to say about connecting client data to a third-party vendor — including the parts most firms have never asked about — and it is legitimate to do that chapter against your own practice management data instead.
What if I get through it and decide to buy something instead?
That is a legitimate conclusion arrived at properly, and chapter five has a section on exactly that decision. A reader who finishes and concludes they should not build did not fail the exercise. They did it — and they can now interrogate whatever they are shown far better than they could at the start of the four live weeks.
Is there a refund?
Yes. Work through the first two weeks. If you have nothing running that you didn’t have before, we refund all of it. The window is two weeks from the start of the first session — after you have seen the first two live weeks of work, so you decide having seen them, not on faith. Email info@byram-advisory.com and it is done.
The only thing we ask in return is a phone call — and to be clear, the refund does not depend on it and nobody is going to try to talk you out of it. We want to know what you expected and did not get. At this stage that is worth considerably more to us than the price, and every one of those calls makes the next version better.
And if you would rather settle the fit question first, that is what the close teardown is for — forty-five minutes, free, and you leave with one written page either way.
Start where it costs you nothing.
The close teardown is free, takes forty-five minutes, and you leave with one written page about your own close. If that is not useful, none of the rest will be either.