Six stages, from preparer to adviser

AI can now read client documents well. The hard part is checking what it read. Here is how staff in an accounting practice move, one stage at a time, from typing figures to reviewing them and then to giving advice, and what the Document Substrate takes over at each stage.

The change: from doing to deciding

Today one person reads every page, types every figure, checks every total and signs. With the Document Substrate, the machine reads, code checks, and you decide. Your name still goes on the work. You just stop spending most of the day copying numbers.

Two rows. Today: you read, type, check and sign. With the Document Substrate: the machine reads the document type and figures, code checks totals, months and duplicates with the same answer every time, and you decide and sign, unsure items first.
Under the hood

The model does three jobs: it suggests a document type, reads the figures into a fixed set of fields for that type, and answers questions. It never adds anything up. Totals, footing and comparisons across documents run in code, so the same figures always give the same verdict. Twenty-two document types have field definitions and sixteen have arithmetic checks.

Six stages, one step at a time

Each stage moves some of your time from production to judgement. Not everyone needs to reach stage 6, and most practices have people at three or four stages at once. That is normal.

A staircase of six stages: preparer, AI-assisted, reviewer, workflow owner, system owner and adviser. Under each, what the Document Substrate takes over: reading the inbox, hiding personal details, checking with unsure items first, testing each rule, drafting the manual, and keeping the same checks when the model changes.
Under the hood

The stages are adapted from roadmaps written for finance teams adopting AI. Stage 3 is the turning point: it is where the work changes from doing to reviewing. Stages 4 and 5 are about writing down what the firm knows, so the system checks each job the way the firm’s seniors would.

Stages 1 and 2: stop pasting client documents into chat

At stage 1 you read and type everything. At stage 2 you ask a chat tool for help, one task at a time. That is faster, but the chat tool sees the client’s name and address, and nothing it gives back has been checked against the page. The Document Substrate swaps personal details for placeholders before the AI that answers sees the text, and every figure comes back with the page it was found on.

Two rows. Stage 2: a rates bill with a name and address is pasted into a chat tool, which sees the details as written and gives an answer nobody checked. With the Document Substrate: the name and address become placeholders, the AI that answers reads placeholders only, and the result is the figure with the page it came from.
Under the hood

Redaction runs in three stages on every page: pattern rules for values with a fixed shape or a checksum (IRD numbers, bank accounts, company numbers, emails, phone numbers), a model pass for names, organisations and addresses, and an independent check. Real values are encrypted into a vault and restored on the server only for people allowed to see that client’s documents. Reading the original document is the one exception: the service that turns the file into text, and for some document types the service that reads the figures, receive the original file, and the model pass sees the names it is there to replace. A figure that cannot be found on its page is marked and left for a person, not passed silently.

Stage 3: work through one queue

At stage 3 you stop preparing and start reviewing. Everything the system was unsure of goes into one queue, unsure items first, each with a reason in plain words and the figures it compared. You accept it, correct it, or ask the client. Everything that passed every check is already traced to its page, ready to look at if you want to.

A review queue. Three flagged items come first: a March closing balance that does not match the April opening balance, a rates bill amount not found on the page, and a missing loan statement. Below them, 41 figures passed every check. Beside the queue: accept with a note, correct the figure, ask the client, then sign the job when nothing is left open.
Under the hood

Rules come in a handful of declared kinds: two figures agree within a tolerance (reconcile), statements chain from one closing balance to the next opening balance (continuity), and nothing is counted twice (unique). A finding names the exact values it read from each document. Corrections are stored separately with the author, the time and the reason, so the AI’s first reading is never overwritten.

Stages 4 and 5: write down what the firm knows

At stage 4 you write down what a recurring job needs: the documents, the checks, and what a finished workpaper looks like. That is a workflow, and corrected figures can be drafted into tests that a person reviews and adds. Build the workpaper so that someone other than its author can find their way round it: that reader is now often an AI tool working inside Excel. At stage 5 you look after several workflows together, and learn from the queue: when reviewers keep accepting the same finding, the system drafts a paragraph for the practice manual and a person decides whether to add it.

Left: a rental property workflow with slots for the documents it needs (the loan statement missing, so the job stays draft), rules that run in code, outputs, and tests from corrected figures. Right: reviewers keep accepting one finding, a draft paragraph for the manual is proposed, and a person decides whether it goes into the practice manual.
Under the hood

A workflow definition names its slots, rules and output template, and states what a passing run must not be read as concluding. Rules are data, not code, so their formulas can be shown to a reviewer. The practice manual is versioned markdown with one set of chapters per engagement type. A build-time test checks that every rule a chapter names exists and that every workflow’s rules are covered. The manual explains the checks; it never changes them.

A workbook an AI tool can read needs what a new member of staff needs. Each figure has a descriptive name that stays the same from one job to the next, such as rental.rates.total, not a cell reference. A contents sheet lists each named figure, the document and page it came from, whether it passed its checks, and whether a person changed it. Figures are stored as values, and formulas are used only for arithmetic.

Stage 6: the work clients pay for

At stage 6 your time goes on judgement, advice and client conversations. When a better AI model comes out, the checks and the sign-off stay the same. The model and parser versions are pinned, and a change is tested against real files before it ships. There is no stage 7. From here, the firm keeps improving how it works as the tools change.

Client documents go to the AI model, which can be swapped for next year's model, then to code checks and a person who signs, both of which stay the same. Below: your time goes on judgement, advice and client conversations.
Under the hood

A change to the AI’s behaviour, including a new model, is tested against evaluation suites built from the firm’s own rated answers and corrected figures: field accuracy, no invented figures, classification, redaction and chat answers. For redaction, zero leaks is the pass mark. The rules engine makes no model call, so the same figures always get the same verdict. A new model can still read a figure differently, which is why versions are pinned.

The full article: Making an Accounting Practice AI-Compatible: Six Stages for Staff.