Making an Accounting Practice AI-Compatible: Six Stages for Staff

Accounting firms are being told that their workflows have to become "AI-compatible". The models are good enough to do production work. Most practices are not set up so that the work can be checked. This article sets out what AI-compatible means for an accounting practice, why reviewability is the hard part, and six stages a member of staff moves through, with what the Document Substrate takes over at each one.


The argument

In a recent episode, Jason Staats, who coaches accounting firm owners, argues that current models from Anthropic and OpenAI change how a firm should run its production work: bookkeeping, tax preparation and review. Most firms run on processes designed decades ago for people working by hand. His point is that these processes need to be rebuilt as structured, reviewable systems that AI can work inside, not have AI added on top.

He makes two further points that matter here. A single accountant with AI can now do work that used to need a team, but growing past that still needs people for running the system and for client relationships. And answering a question about keeping books with Claude and a plain-text ledger, he says it works for small entities, but reviewability is still the problem on complex files.

That last point is the one this article is about.

Reviewability is the hard part

A model can read a bank statement and produce a list of transactions. The question a partner asks is not whether the list looks right. It is whether each figure is on the page, whether the totals add up, whether a month is missing, and who checked it. If the answer to all of that is "the model said so", the work has to be done again by hand before anyone signs it, and nothing has been saved.

AI-compatible therefore means three things for an accounting practice:

  • Structured inputs. Each job states which documents it needs. A missing document is found before any work is done, not at review.
  • Checks that do not depend on the model. The model reads. Code adds the totals and compares figures across documents, so the same files always get the same verdict.
  • A record a reviewer can check. Every figure is tied to its page, uncertain items come first with a reason, and a person signs.

A standardised Excel workbook is a good first step towards the first point. The second and third need software. That software is the Document Substrate.

What the Document Substrate provides

From doer to reviewer: validation checks what goes in, verification checks what comes out, and a person signs

  • The inbox is read in any state. PDF, Word, Excel, scans and photos, arriving by upload, email or API. Document type is detected from the contents. Confident detections are applied; the rest wait for a person.
  • Personal details never reach the model. Names, addresses, bank account, IRD and company numbers are replaced with placeholders before any model call and restored only for an authorised person. Every restore is logged.
  • Checks run in code. Twenty-two document types have field definitions and eighteen have arithmetic checks. Same files, same verdict.
  • The model never does arithmetic. Code adds the totals. The workpaper and draft letter are generated from checked records only.
  • Every figure is traced to its page. A figure that cannot be found on the page is flagged, never passed.
  • A person reviews and signs. Uncertain items come first, each with a reason.

A workflow defines each job: the slots (the documents the job needs), the rules that run in code over the figures read from them, and the outputs. For a rental property, the slots are the bank statement, rent summary, loan statement, rates bill and invoices. If the loan statement is missing, the workpaper stays DRAFT and the client gets a chase list.

The firm's own expertise sits on top of this as a substrate: the calculations and thresholds it applies, the judgement calls its seniors make, and the way it does the work. The practice manual records that expertise in the firm's words, and a test fails if the manual and the checks disagree.

Six stages for staff

Finance teams are being given roadmaps for AI adoption that run from doing the work by hand to managing an AI-supported system. The version below is adapted for staff in an accounting practice. For each stage it sets out where the person's time goes, what limits them, what the Document Substrate takes over, and what moves them to the next stage.

Six stages for accounting staff, from preparer to adviser, with what the Document Substrate adds at each: reading the inbox, redacting before the model, checking in code, testing rules on the firm's own jobs, drafting the manual from accepted findings, and keeping the same checks when the model changes

Not everyone needs to reach stage 6, and a firm does not move all its staff at once. Most practices will have people at three or four stages at the same time.

1. Preparer

Where the time goes. Reading client documents, keying figures, building the workpaper, chasing what is missing.

What limits it. The number of hours one person has. Most of those hours go on transcription, not judgement.

What the Document Substrate takes over. Reading the inbox. Documents are classified and the figures read, so the preparer checks figures instead of typing them.

To move on. Separate the work that needs your judgement from the work that is reading and copying. The second kind is what the machine does.

2. AI-assisted preparer

Where the time goes. Using a chat assistant task by task: summarising a document, drafting an email, explaining a variance. Each task starts again, and the context is pasted in every time.

What limits it. Nothing the assistant produces is checked against the source, so every figure has to be checked by hand. Client documents pasted into a general chat tool take personal details with them.

What the Document Substrate takes over. Personal details are replaced with placeholders before any model call. Figures are traced to their page, so checking one is a click, not a search.

To move on. Stop pasting client documents into general chat tools. Use a tool that removes personal details and traces what it reads.

3. Reviewer

Where the time goes. Working through one queue of findings. Each item is a check that fired or a figure the machine was unsure of, with a plain-language reason. You accept, correct or resolve each one and sign.

What limits it. You still hold the shape of each job in your head: which documents it needs, which differences are normal for this client, when to chase.

What the Document Substrate takes over. The checks: footing, continuity from one statement to the next, duplicates, figures agreeing across documents. Uncertain items come first. Confident ones have already passed every check.

To move on. Write down what a recurring job needs: the documents, the checks, what a finished workpaper looks like. That is the first draft of a workflow.

This is the stage where the work changes from doing to reviewing. The principle is set out in Doer to reviewer.

4. Workflow owner

Where the time goes. Defining the slots, rules and outputs for an engagement type the firm does many times a year, such as rental properties or annual accounts for small companies, and writing the practice manual chapter that explains each rule.

What limits it. A mistake in a rule affects every job that uses it. A rule that is too strict fills the queue with findings nobody needs to see; one that is too loose lets errors through.

What the Document Substrate takes over. Rules are declared as data, so the formula can be shown to the reviewer and a failure names the exact figures it compared. Corrected figures and rated answers become test cases that every change has to pass.

To move on. Add test cases from real jobs, including the awkward ones. Make the workflow reliable enough that nobody has to coordinate each step by hand.

5. System owner

Where the time goes. Looking after several workflows as one system. Deciding which results are applied automatically and which wait for a person. Reading which findings reviewers keep accepting, and why.

What limits it. More automation adds complexity. Every result applied without a person needs the evidence to justify it.

What the Document Substrate takes over. When reviewers repeatedly accept the same finding, a draft paragraph for the practice manual is proposed, with personal details removed, and a person decides whether to add it. Every restore of a personal detail and every decision is logged.

To move on. Decide on evidence, not on enthusiasm, where a person's time adds most value and where the checks are enough.

6. Adviser

Where the time goes. Judgement, client conversations, advice and signing. This is the work clients pay a practice for, and it is the work nobody had time for at stage 1.

What limits it. Nothing in the production process. The question is where AI and people each add most value, and that changes as the models improve.

What the Document Substrate takes over. Production, within the limits the firm has set. When a better model is released, it reads more accurately, and the checks in code stay the same. The verdict on a file does not depend on which model read it.

There is no stage 7. The work from here is improving how the firm operates as the tools change, and measuring that by the advice clients receive, not by how much AI the firm uses.

The solo practitioner

The stages apply to a practice of one as well. A sole practitioner with the Document Substrate covers stages 1 to 3 for every client, because the machine reads and checks and the practitioner reviews. Stages 4 and 5 are where a sole practitioner most often needs help: writing down the rules for each engagement type and testing them. That is what a Map captures, and it is done once for each engagement type, not for every job.

Signing does not change. However much the machine reads and checks, a person resolves the uncertain items and signs the work.

Where to start

Three questions show where a practice is on the roadmap and what to write down first.

  • Which job do you repeat most? Pick the engagement type with the most clients. List the documents it needs and the checks a senior runs before signing. That list is a first workflow.
  • What do new staff get wrong? It is usually something the senior knows and has not written down: a check, a threshold, a reason to chase. That is the most valuable thing to capture.
  • Where do client documents go today? If staff are pasting them into general chat tools, that is the first thing to change.

Six stages, explained simply · The Document Substrate in detail · Principles · Non-determinism and calculation validation, the short version · Worked examples across six professions