This is what you’d receive from an Orellis audit.
A short written assessment of one workflow in a 20-person bookkeeping practice. Three ranked AI opportunities, an honest view of where AI is the wrong answer, and the recommendation. Illustrative throughout — no real client, no measured results.
A 20-person administratiekantoor — monthly client reporting cycle.
Picture a 20-person administratiekantoor. Fifteen of those people are accountants or bookkeepers; the other five handle client contact, planning, and administration. Every month, the firm closes the books for approximately 80 SME clients and sends each one a management summary: a PDF with key figures, a short written commentary on anything unusual, and a list of open items requiring the client’s attention.
The cycle runs from the 1st to the 15th of each month. By the 12th, the whole firm is in heads-down mode. Commentary drafting — the part that feels like it should be quick — routinely runs to two to three hours per accountant per cycle. Nobody became an accountant to write the same twelve variants of “revenue is tracking broadly in line with the prior period.”
This audit focused on that one workflow: the monthly reporting cycle, from closed accounts to dispatched client package.
Ranked by impact per unit of implementation effort.
Each estimate is illustrative reasoning from the task structure — not a measured result from any real firm.
Commentary Drafting — First-Draft Generation
What it is
Once the numbers are finalized in the accounting system, a structured prompt pulls the key figures (revenue, cost movements, cash position, open debtors) and generates a first-draft commentary paragraph for the accountant to review, adjust, and send. The accountant corrects tone, adds context they hold in their head, and approves. They do not start from a blank page.
Human vs AI split
AI handles assembly: pulling the figures, applying the firm’s commentary template, flagging any line item that moved more than a defined threshold. The accountant handles judgment: knowing that the revenue dip is because the client’s biggest customer is on holiday in August, not a trading problem; knowing that the tone needs to soften because the client relationship is sensitive right now; making the call on what to flag versus what to leave unremarked.
The phrase that matters here: hire for the judgment, automate the assembly.
Illustrative time estimate
In a case like this — a practice of this size, 80 monthly clients, commentary averaging 20 minutes per client to draft from scratch — the total monthly drafting load across the team is in the range of 25–30 hours. First-draft generation of this type typically reduces the accountant’s time per report to 5–8 minutes of review and edit rather than 20 minutes of drafting. In a case like this, that is an estimated saving of roughly 15–20 hours of senior accountant time per month across the team.
These are reasoning-based estimates from the task structure, not measured results from this firm.
What stays human
Every word that goes to the client. The accountant reads, edits, and sends. The AI produces a draft, not a deliverable. Any commentary touching a sensitive client relationship, a going-concern judgment, or an unusual transaction stays fully in the accountant’s hands.
Open-Items Chasing — Automated Follow-Up Drafts
What it is
Every monthly close produces a list of open items: missing bank statements, unsigned documents, unreconciled transactions the client needs to clarify. Currently an accountant drafts a follow-up email to each client listing their open items. An AI agent pulls the open-items list per client from the accounting system, matches it to the client’s contact record, and drafts a follow-up email in the firm’s standard register. The accountant reviews and sends.
Human vs AI split
AI drafts the email, populates the client-specific item list, applies the right tone register, and flags any open item that looks like it needs a phone call rather than an email. The accountant decides whether to send, edit, or escalate to a call.
Illustrative time estimate
In a case like this, with 80 clients and an average of three open items per client per cycle, drafting follow-up emails is typically 30–45 minutes of accountant time per day for several days in the middle of the cycle. A drafting agent reduces that to a review task: 5–10 minutes to scan the batch and approve or edit. Across a month, that is an estimated saving of 2–4 hours of mid-cycle accountant time.
Modest individually, but meaningful in the context of a cycle where everyone is already stretched. Illustrative estimate only.
What stays human
The send decision. Any client where the open item is sensitive, long-overdue, or likely to produce a difficult conversation. The accountant knows which clients need a call, not an email. The AI does not.
Client Package Assembly — PDF Compilation and Dispatch
What it is
The final client package — management accounts PDF, commentary, open-items list, any supporting schedules — is currently assembled by hand: exporting from the accounting system, attaching the commentary, naming the file correctly, attaching to an email, sending. For 80 clients, this is pure assembly work. An agent pulls the relevant outputs per client, assembles them into a standardized PDF package, and either deposits it in the client portal or attaches it to a pre-drafted dispatch email. The accountant triggers the batch send after a final review.
Illustrative time estimate
In a case like this, package assembly for 80 clients typically runs 3–5 hours of administrative time per cycle. Automating this step is designed to recover that time. The implementation is the most mechanical of the three opportunities: the task is entirely assembly, the inputs are already digital, and the output format is fixed.
Illustrative estimate from task structure. Actual time depends on current tooling and export cleanliness.
What stays human
The dispatch decision. No package reaches a client without a human confirming the batch. Spot-checking for accuracy is non-negotiable — the accountant is professionally liable for the figures in that package.
This section is the point of the audit.
A vendor who only tells you what to automate is selling you something. Orellis tells you where to stop.
The Relationship Judgment Call
Several of the 80 clients in a practice of this size are going through something: a bad year, a divorce affecting the business, a dispute with a supplier, a conversation about whether to keep trading. The accountant knows this. That knowledge shapes the commentary, the tone of the follow-up, the decision to pick up the phone instead of sending the package.
No AI tool should be in the commentary chain for these clients. Not because the AI cannot produce fluent text — it can. Because the risk of a tone-deaf commentary landing at the wrong moment in a client relationship that the accountant has managed for ten years is not recoverable with an edit. The correct answer for sensitive-relationship clients: the accountant writes the commentary. Full stop.
The audit would flag these clients for exclusion from the drafting workflow — typically a small number, but the ones where it matters most.
The Accountant’s Professional Judgment on Unusual Transactions
The commentary drafting workflow described above works well for standard months: revenue in line with prior period, costs normal, nothing flagged. The moment an unusual transaction appears — a large one-off receipt, an unexplained cost movement, a balance that does not reconcile cleanly — the accountant’s judgment is the only appropriate tool.
An AI agent set up to flag threshold breaches will catch obvious movements. It will not catch the transaction that looks normal on the face of it but does not match what the accountant knows about the client’s business. That pattern recognition — “this doesn’t look right given what I know about this firm” — is the core of what a good bookkeeper provides. It cannot be automated, and any system that implies otherwise is misrepresenting what AI is currently capable of.
The correct design: AI handles the standard months, drafts the commentary, flags threshold breaches. The accountant always reviews the flag list before the draft goes out. Unusual transactions go back to the accountant’s judgment, not the AI’s phrasing.
Which opportunity to pilot first — and why.
Start with commentary drafting.
It has the highest time-saving potential of the three. It sits in the most senior-time part of the workflow (accountants, not administrators). And the implementation requires no system integration to start — a structured prompt against exported figures is sufficient to run a test in week one.
It also produces the most visible proof of value inside the firm. When an accountant sends the first batch of AI-drafted commentaries — reviewed and edited, but not written from scratch — and the clients respond normally, the internal skepticism about whether this is safe to use dissolves faster than any demo could achieve. That trust compounds into the second and third workflows.
The honest next step is a 30-minute scoping call to walk through three things: which accounting system the firm is running, what the current commentary template looks like, and whether the firm’s data export is clean enough to use as AI input without manual cleaning. Those three questions determine whether the build is one week or three.
On sequencing: if the firm wants to move faster, opportunities 1 and 2 can run in parallel — they use different inputs and have no dependencies between them. Opportunity 3 (package assembly) is worth building after the first two are stable.
This document is the written output of an Orellis AI Audit.
It is designed to arrive after a 30-minute call. The call covers three things: your goal for the year (not “what workflows do you have” — what you are actually trying to achieve), what is in the way of it, and where AI shortens the path fastest.
The written assessment ranks the opportunities Orellis identifies, gives you an honest view of what not to automate and why, and tells you which one to pilot first with the reasoning.
The assessment is yours whether or not you engage further. If the pilot makes sense, that is a separate fixed-price conversation. There is no open-ended retainer in the first meeting.
The audit is free while launch slots last. The point is to give you something useful before you commit to anything.
This is what you’d receive. Tell us where your time goes.
One line about the workflow your team repeats most. Rachel reads every submission and replies within 2 business days.
Tell us where your time goes →